Private verified fixture stage b4cd2220-55a6-4224-9a90-aec53a188606

.gitattributes

@@ -1,60 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.avro filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.lz4 filter=lfs diff=lfs merge=lfs -text
-*.mds filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
-# Audio files - uncompressed
-*.pcm filter=lfs diff=lfs merge=lfs -text
-*.sam filter=lfs diff=lfs merge=lfs -text
-*.raw filter=lfs diff=lfs merge=lfs -text
-# Audio files - compressed
-*.aac filter=lfs diff=lfs merge=lfs -text
-*.flac filter=lfs diff=lfs merge=lfs -text
-*.mp3 filter=lfs diff=lfs merge=lfs -text
-*.ogg filter=lfs diff=lfs merge=lfs -text
-*.wav filter=lfs diff=lfs merge=lfs -text
-# Image files - uncompressed
-*.bmp filter=lfs diff=lfs merge=lfs -text
-*.gif filter=lfs diff=lfs merge=lfs -text
-*.png filter=lfs diff=lfs merge=lfs -text
-*.tiff filter=lfs diff=lfs merge=lfs -text
-# Image files - compressed
-*.jpg filter=lfs diff=lfs merge=lfs -text
-*.jpeg filter=lfs diff=lfs merge=lfs -text
-*.webp filter=lfs diff=lfs merge=lfs -text
-# Video files - compressed
-*.mp4 filter=lfs diff=lfs merge=lfs -text
-*.webm filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,620 @@
+---
+license: "mit"
+tags:
+- format-fixture
+- random-init
+- rtn
+- optimizer-not-run
+---
+
+# microfloat tiny CPU format fixture reproducibility
+
+Complete tiny random FORMAT fixture evidence. Round-to-nearest (RTN) storage/reader exercise only; **optimizer-not-run**. No GPTQ/AWQ/AutoRound optimization, calibrated ModelOpt/CT/QAT quality, trained-model quality ranking, GPU parity, or native serving-kernel correctness claim.
+
+Reconstructed weights are evaluated by the captured native forward. KL is own-head, full-vocabulary on the recorded panel, not a benchmark of training quality. Advisory reconstruction and activation omissions are retained verbatim in the raw evidence.
+
+This is a **dataset evidence bundle**, not a multi-model model repository. Each consumer model has its own repository:
+
+- https://huggingface.co/malaiwah/glm-moe-dsa-tiny-random-ct-mxfp4-rtn-format
+- https://huggingface.co/malaiwah/glm-moe-dsa-tiny-random-ct-nvfp4-rtn-format
+- https://huggingface.co/malaiwah/glm-moe-dsa-tiny-random-modelopt-nvfp4-rtn-format
+- https://huggingface.co/malaiwah/glm-moe-dsa-tiny-random-fp8-ue8m0-rtn-format
+- https://huggingface.co/malaiwah/glm-moe-dsa-tiny-random-modelopt-mixed-rtn-format
+
+Raw model/source trees, original cards/licenses, captures, reports and retained scripts are preserved under `raw/`. `qualification.json` records checked results, `source-inventory.json` records every original source byte, and `publication-plan.json` records the exact commands and provenance. Each upstream source retains its own license; the dataset license does not relicense bundled code or checkpoints.
+
+## Reproduction commands and source hashes
+
+Commands below are recorded argv, not executed by this publisher. Restore each recorded input from the dataset raw trees and map original absolute paths to your workspace. Retained tools/dependency files and byte hashes are in `source-inventory.json`; do not substitute moving upstream branches.
+
+```json
+[
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root",
+      "--dataset-id",
+      "fidelity--quant-integration.root",
+      "--dataset-name",
+      "Local quant integration root",
+      "--role",
+      "root"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate",
+      "--dataset-id",
+      "fidelity--quant-integration.candidate",
+      "--dataset-name",
+      "Local quant integration candidate",
+      "--role",
+      "quant",
+      "--codec",
+      "mxfp4",
+      "--declared-bits",
+      "4.0",
+      "--scope-file",
+      "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4/scope.json"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+      "compare",
+      "--reference",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root",
+      "--candidate",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/comparison",
+      "--force-compute",
+      "--device",
+      "cpu",
+      "--replay-device",
+      "numpy",
+      "--replay-dtype",
+      "float32",
+      "--verify-tensors",
+      "--own-heads"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/root",
+      "--dataset-id",
+      "fidelity--quant-integration.root",
+      "--dataset-name",
+      "Local quant integration root",
+      "--role",
+      "root"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-quant-microfloat-release/stage-main/ct-nvfp4",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/candidate",
+      "--dataset-id",
+      "fidelity--quant-integration.candidate",
+      "--dataset-name",
+      "Local quant integration candidate",
+      "--role",
+      "quant",
+      "--codec",
+      "nvfp4",
+      "--declared-bits",
+      "4.0",
+      "--scope-file",
+      "/tmp/qfs-quant-microfloat-release/stage-main/ct-nvfp4/scope.json"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+      "compare",
+      "--reference",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/root",
+      "--candidate",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/candidate",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/comparison",
+      "--force-compute",
+      "--device",
+      "cpu",
+      "--replay-device",
+      "numpy",
+      "--replay-dtype",
+      "float32",
+      "--verify-tensors",
+      "--own-heads"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/root",
+      "--dataset-id",
+      "fidelity--quant-integration.root",
+      "--dataset-name",
+      "Local quant integration root",
+      "--role",
+      "root"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-nvfp4",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/candidate",
+      "--dataset-id",
+      "fidelity--quant-integration.candidate",
+      "--dataset-name",
+      "Local quant integration candidate",
+      "--role",
+      "quant",
+      "--codec",
+      "nvfp4",
+      "--declared-bits",
+      "4.0",
+      "--scope-file",
+      "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-nvfp4/scope.json"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+      "compare",
+      "--reference",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/root",
+      "--candidate",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/candidate",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/comparison",
+      "--force-compute",
+      "--device",
+      "cpu",
+      "--replay-device",
+      "numpy",
+      "--replay-dtype",
+      "float32",
+      "--verify-tensors",
+      "--own-heads"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/root",
+      "--dataset-id",
+      "fidelity--quant-integration.root",
+      "--dataset-name",
+      "Local quant integration root",
+      "--role",
+      "root"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-quant-microfloat-release/stage-main/fp8-ue8m0",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/candidate",
+      "--dataset-id",
+      "fidelity--quant-integration.candidate",
+      "--dataset-name",
+      "Local quant integration candidate",
+      "--role",
+      "quant",
+      "--codec",
+      "fp8_e4m3",
+      "--declared-bits",
+      "8.0",
+      "--scope-file",
+      "/tmp/qfs-quant-microfloat-release/stage-main/fp8-ue8m0/scope.json"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+      "compare",
+      "--reference",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/root",
+      "--candidate",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/candidate",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/comparison",
+      "--force-compute",
+      "--device",
+      "cpu",
+      "--replay-device",
+      "numpy",
+      "--replay-dtype",
+      "float32",
+      "--verify-tensors",
+      "--own-heads"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/root",
+      "--dataset-id",
+      "fidelity--quant-integration.root",
+      "--dataset-name",
+      "Local quant integration root",
+      "--role",
+      "root"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-mixed",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/candidate",
+      "--dataset-id",
+      "fidelity--quant-integration.candidate",
+      "--dataset-name",
+      "Local quant integration candidate",
+      "--role",
+      "quant",
+      "--codec",
+      "mixed",
+      "--declared-bits",
+      "4.0",
+      "--scope-file",
+      "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-mixed/scope.json"
+    ]
+  },
+  {
+    "cwd": "/tmp/qfs-explorer-development",
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+      "compare",
+      "--reference",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/root",
+      "--candidate",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/candidate",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/comparison",
+      "--force-compute",
+      "--device",
+      "cpu",
+      "--replay-device",
+      "numpy",
+      "--replay-dtype",
+      "float32",
+      "--verify-tensors",
+      "--own-heads"
+    ]
+  }
+]
+```
+
+The publication receipt is written locally only after anonymous verification of complete remote bytes at exact commits. Publication is not additional model or serving qualification.

publication-plan.json

@@ -0,0 +1,839 @@
+{
+  "bundle": "microfloat-cpu-format-v1",
+  "candidates": [
+    {
+      "activation_omissions": [
+        "Reconstructed weights only; no activation quantization, QAT, calibration or GPU kernel validation"
+      ],
+      "architecture_restrictions": [
+        "GLM MoE DSA, dense and routed expert matrices; decoded BF16 execution, not full original quantized release"
+      ],
+      "classification": "lossy",
+      "control_max_kld": null,
+      "evidence": {
+        "candidate": "candidate",
+        "comparison": "comparison/comparison-receipt.json",
+        "dir": "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4",
+        "reference": "root"
+      },
+      "format": "ct-mxfp4; CPU RTN format fixture",
+      "head_policy": "native",
+      "model_dir": "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+      "name": "ct-mxfp4",
+      "repo_id": "malaiwah/glm-moe-dsa-tiny-random-ct-mxfp4-rtn-format",
+      "source": {
+        "license": "mit",
+        "license_file": "/tmp/qfs-complete-tiny-fixture/model/LICENSE",
+        "model_dir": "/tmp/qfs-complete-tiny-fixture/model",
+        "pin": {
+          "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "revision": "a1a973330e4e912678adbd8ea41e8b86b1c9dccb"
+        },
+        "pin_evidence": {
+          "files_pointer": "/source_sha256",
+          "path": "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4/provenance.json",
+          "repository_pointer": "/source/repository",
+          "revision_pointer": "/source/revision"
+        }
+      }
+    },
+    {
+      "activation_omissions": [
+        "Reconstructed weights only; no activation quantization, QAT, calibration or GPU kernel validation"
+      ],
+      "architecture_restrictions": [
+        "GLM MoE DSA, dense and routed expert matrices; decoded BF16 execution, not full original quantized release"
+      ],
+      "classification": "lossy",
+      "control_max_kld": null,
+      "evidence": {
+        "candidate": "candidate",
+        "comparison": "comparison/comparison-receipt.json",
+        "dir": "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4",
+        "reference": "root"
+      },
+      "format": "ct-nvfp4; CPU RTN format fixture",
+      "head_policy": "native",
+      "model_dir": "/tmp/qfs-quant-microfloat-release/stage-main/ct-nvfp4",
+      "name": "ct-nvfp4",
+      "repo_id": "malaiwah/glm-moe-dsa-tiny-random-ct-nvfp4-rtn-format",
+      "source": {
+        "license": "mit",
+        "license_file": "/tmp/qfs-complete-tiny-fixture/model/LICENSE",
+        "model_dir": "/tmp/qfs-complete-tiny-fixture/model",
+        "pin": {
+          "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "revision": "a1a973330e4e912678adbd8ea41e8b86b1c9dccb"
+        },
+        "pin_evidence": {
+          "files_pointer": "/source_sha256",
+          "path": "/tmp/qfs-quant-microfloat-release/stage-main/ct-nvfp4/provenance.json",
+          "repository_pointer": "/source/repository",
+          "revision_pointer": "/source/revision"
+        }
+      }
+    },
+    {
+      "activation_omissions": [
+        "Reconstructed weights only; no activation quantization, QAT, calibration or GPU kernel validation"
+      ],
+      "architecture_restrictions": [
+        "GLM MoE DSA, dense and routed expert matrices; decoded BF16 execution, not full original quantized release"
+      ],
+      "classification": "lossy",
+      "control_max_kld": null,
+      "evidence": {
+        "candidate": "candidate",
+        "comparison": "comparison/comparison-receipt.json",
+        "dir": "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4",
+        "reference": "root"
+      },
+      "format": "modelopt-nvfp4; CPU RTN format fixture",
+      "head_policy": "native",
+      "model_dir": "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-nvfp4",
+      "name": "modelopt-nvfp4",
+      "repo_id": "malaiwah/glm-moe-dsa-tiny-random-modelopt-nvfp4-rtn-format",
+      "source": {
+        "license": "mit",
+        "license_file": "/tmp/qfs-complete-tiny-fixture/model/LICENSE",
+        "model_dir": "/tmp/qfs-complete-tiny-fixture/model",
+        "pin": {
+          "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "revision": "a1a973330e4e912678adbd8ea41e8b86b1c9dccb"
+        },
+        "pin_evidence": {
+          "files_pointer": "/source_sha256",
+          "path": "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-nvfp4/provenance.json",
+          "repository_pointer": "/source/repository",
+          "revision_pointer": "/source/revision"
+        }
+      }
+    },
+    {
+      "activation_omissions": [
+        "Reconstructed weights only; no activation quantization, QAT, calibration or GPU kernel validation"
+      ],
+      "architecture_restrictions": [
+        "GLM MoE DSA, dense and routed expert matrices; decoded BF16 execution, not full original quantized release"
+      ],
+      "classification": "lossy",
+      "control_max_kld": null,
+      "evidence": {
+        "candidate": "candidate",
+        "comparison": "comparison/comparison-receipt.json",
+        "dir": "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0",
+        "reference": "root"
+      },
+      "format": "fp8-ue8m0; CPU RTN format fixture",
+      "head_policy": "native",
+      "model_dir": "/tmp/qfs-quant-microfloat-release/stage-main/fp8-ue8m0",
+      "name": "fp8-ue8m0",
+      "repo_id": "malaiwah/glm-moe-dsa-tiny-random-fp8-ue8m0-rtn-format",
+      "source": {
+        "license": "mit",
+        "license_file": "/tmp/qfs-complete-tiny-fixture/model/LICENSE",
+        "model_dir": "/tmp/qfs-complete-tiny-fixture/model",
+        "pin": {
+          "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "revision": "a1a973330e4e912678adbd8ea41e8b86b1c9dccb"
+        },
+        "pin_evidence": {
+          "files_pointer": "/source_sha256",
+          "path": "/tmp/qfs-quant-microfloat-release/stage-main/fp8-ue8m0/provenance.json",
+          "repository_pointer": "/source/repository",
+          "revision_pointer": "/source/revision"
+        }
+      }
+    },
+    {
+      "activation_omissions": [
+        "Reconstructed weights only; no activation quantization, QAT, calibration or GPU kernel validation"
+      ],
+      "architecture_restrictions": [
+        "GLM MoE DSA, dense and routed expert matrices; decoded BF16 execution, not full original quantized release"
+      ],
+      "classification": "lossy",
+      "control_max_kld": null,
+      "evidence": {
+        "candidate": "candidate",
+        "comparison": "comparison/comparison-receipt.json",
+        "dir": "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed",
+        "reference": "root"
+      },
+      "format": "modelopt-mixed; CPU RTN format fixture",
+      "head_policy": "native",
+      "model_dir": "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-mixed",
+      "name": "modelopt-mixed",
+      "repo_id": "malaiwah/glm-moe-dsa-tiny-random-modelopt-mixed-rtn-format",
+      "source": {
+        "license": "mit",
+        "license_file": "/tmp/qfs-complete-tiny-fixture/model/LICENSE",
+        "model_dir": "/tmp/qfs-complete-tiny-fixture/model",
+        "pin": {
+          "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "revision": "a1a973330e4e912678adbd8ea41e8b86b1c9dccb"
+        },
+        "pin_evidence": {
+          "files_pointer": "/source_sha256",
+          "path": "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-mixed/provenance.json",
+          "repository_pointer": "/source/repository",
+          "revision_pointer": "/source/revision"
+        }
+      }
+    }
+  ],
+  "commands": [
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root",
+        "--dataset-id",
+        "fidelity--quant-integration.root",
+        "--dataset-name",
+        "Local quant integration root",
+        "--role",
+        "root"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate",
+        "--dataset-id",
+        "fidelity--quant-integration.candidate",
+        "--dataset-name",
+        "Local quant integration candidate",
+        "--role",
+        "quant",
+        "--codec",
+        "mxfp4",
+        "--declared-bits",
+        "4.0",
+        "--scope-file",
+        "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4/scope.json"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+        "compare",
+        "--reference",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root",
+        "--candidate",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/comparison",
+        "--force-compute",
+        "--device",
+        "cpu",
+        "--replay-device",
+        "numpy",
+        "--replay-dtype",
+        "float32",
+        "--verify-tensors",
+        "--own-heads"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/root",
+        "--dataset-id",
+        "fidelity--quant-integration.root",
+        "--dataset-name",
+        "Local quant integration root",
+        "--role",
+        "root"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-quant-microfloat-release/stage-main/ct-nvfp4",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/candidate",
+        "--dataset-id",
+        "fidelity--quant-integration.candidate",
+        "--dataset-name",
+        "Local quant integration candidate",
+        "--role",
+        "quant",
+        "--codec",
+        "nvfp4",
+        "--declared-bits",
+        "4.0",
+        "--scope-file",
+        "/tmp/qfs-quant-microfloat-release/stage-main/ct-nvfp4/scope.json"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+        "compare",
+        "--reference",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/root",
+        "--candidate",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/candidate",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/ct-nvfp4/comparison",
+        "--force-compute",
+        "--device",
+        "cpu",
+        "--replay-device",
+        "numpy",
+        "--replay-dtype",
+        "float32",
+        "--verify-tensors",
+        "--own-heads"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/root",
+        "--dataset-id",
+        "fidelity--quant-integration.root",
+        "--dataset-name",
+        "Local quant integration root",
+        "--role",
+        "root"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-nvfp4",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/candidate",
+        "--dataset-id",
+        "fidelity--quant-integration.candidate",
+        "--dataset-name",
+        "Local quant integration candidate",
+        "--role",
+        "quant",
+        "--codec",
+        "nvfp4",
+        "--declared-bits",
+        "4.0",
+        "--scope-file",
+        "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-nvfp4/scope.json"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+        "compare",
+        "--reference",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/root",
+        "--candidate",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/candidate",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-nvfp4/comparison",
+        "--force-compute",
+        "--device",
+        "cpu",
+        "--replay-device",
+        "numpy",
+        "--replay-dtype",
+        "float32",
+        "--verify-tensors",
+        "--own-heads"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/root",
+        "--dataset-id",
+        "fidelity--quant-integration.root",
+        "--dataset-name",
+        "Local quant integration root",
+        "--role",
+        "root"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-quant-microfloat-release/stage-main/fp8-ue8m0",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/candidate",
+        "--dataset-id",
+        "fidelity--quant-integration.candidate",
+        "--dataset-name",
+        "Local quant integration candidate",
+        "--role",
+        "quant",
+        "--codec",
+        "fp8_e4m3",
+        "--declared-bits",
+        "8.0",
+        "--scope-file",
+        "/tmp/qfs-quant-microfloat-release/stage-main/fp8-ue8m0/scope.json"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+        "compare",
+        "--reference",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/root",
+        "--candidate",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/candidate",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/fp8-ue8m0/comparison",
+        "--force-compute",
+        "--device",
+        "cpu",
+        "--replay-device",
+        "numpy",
+        "--replay-dtype",
+        "float32",
+        "--verify-tensors",
+        "--own-heads"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/root",
+        "--dataset-id",
+        "fidelity--quant-integration.root",
+        "--dataset-name",
+        "Local quant integration root",
+        "--role",
+        "root"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+        "--panel",
+        "/tmp/qfs-complete-tiny-fixture/panel",
+        "--panel-role",
+        "final",
+        "--panel-binding",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/panel-binding.json",
+        "--panel-binding-sha256",
+        "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+        "--panel-tokenizer-root",
+        "/tmp/qfs-complete-tiny-fixture/model",
+        "--lane",
+        "other",
+        "--device",
+        "cpu",
+        "--dtype",
+        "bfloat16",
+        "--author",
+        "malaiwah",
+        "--schedule",
+        "layer-outer",
+        "--layer-residency",
+        "stream",
+        "--disclosures",
+        "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+        "--model",
+        "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-mixed",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/candidate",
+        "--dataset-id",
+        "fidelity--quant-integration.candidate",
+        "--dataset-name",
+        "Local quant integration candidate",
+        "--role",
+        "quant",
+        "--codec",
+        "mixed",
+        "--declared-bits",
+        "4.0",
+        "--scope-file",
+        "/tmp/qfs-quant-microfloat-release/stage-main/modelopt-mixed/scope.json"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    },
+    {
+      "argv": [
+        "/tmp/qfs-quant-venv/bin/python",
+        "-c",
+        "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+        "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+        "compare",
+        "--reference",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/root",
+        "--candidate",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/candidate",
+        "--out",
+        "/tmp/qfs-quant-microfloat-release/proof-public/modelopt-mixed/comparison",
+        "--force-compute",
+        "--device",
+        "cpu",
+        "--replay-device",
+        "numpy",
+        "--replay-dtype",
+        "float32",
+        "--verify-tensors",
+        "--own-heads"
+      ],
+      "cwd": "/tmp/qfs-explorer-development"
+    }
+  ],
+  "evidence_license": "mit",
+  "evidence_repo_id": "malaiwah/qfs-microfloat-tiny-cpu-format-v1",
+  "owner": "malaiwah",
+  "retain": [
+    {
+      "kind": "dependency",
+      "name": "requirements-verified.txt",
+      "path": "/tmp/qfs-quant-integration/requirements-verified.txt"
+    },
+    {
+      "kind": "script",
+      "name": "smoke_capture.py",
+      "path": "/tmp/qfs-quant-integration/smoke_capture.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--hf_capture.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/hf_capture.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--layer_outer.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/layer_outer.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--quant_stream.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/quant_stream.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--affine_surface.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/affine_surface.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--microscale_surface.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/microscale_surface.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--gguf_qwen35.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/gguf_qwen35.py"
+    },
+    {
+      "kind": "source",
+      "name": "engines--tools--gguf_surface.py",
+      "path": "/tmp/qfs-explorer-development/engines/tools/gguf_surface.py"
+    },
+    {
+      "kind": "source",
+      "name": "bin--fidelity--dscompare.py",
+      "path": "/tmp/qfs-explorer-development/bin/fidelity/dscompare.py"
+    },
+    {
+      "kind": "script",
+      "name": "build_microfloat.py",
+      "path": "/tmp/qfs-quant-microfloat-release/build_microfloat.py"
+    },
+    {
+      "kind": "oracle",
+      "name": "oracle-report.json",
+      "path": "/tmp/qfs-quant-microfloat-release/probe-main.json"
+    }
+  ],
+  "schema": "qfs.fixture-publication-plan.v1",
+  "title": "microfloat tiny CPU format fixture reproducibility"
+}

qualification.json

@@ -0,0 +1,1550 @@
+[
+  {
+    "checkpoint_identity_sha256": "aa51ae526ba79294a678ada6189d0418fa78f1d5e51091e5d10696c56d3470ed",
+    "classification": "lossy",
+    "comparison_receipt_sha256": "c8a4ea7eb65d3ad812514a3faadd0913f27d2cf03e45f11d170682ded1b6726d",
+    "decode": {
+      "activation_quantization": null,
+      "checkpoint_bytes_read": 125392,
+      "comparison_class": "advisory",
+      "components_consumed": 212,
+      "components_not_consumed": [],
+      "components_planned": 212,
+      "formats_decoded": {
+        "mxfp4-pack-quantized": 106
+      },
+      "method": "microfloat-weight-reconstruction",
+      "modules_decoded": 106,
+      "modules_not_decoded": [],
+      "modules_planned": 106,
+      "output_dtype": "bfloat16",
+      "quantization_config": {
+        "config_groups": {
+          "group_0": {
+            "format": "mxfp4-pack-quantized",
+            "input_activations": null,
+            "output_activations": null,
+            "targets": [
+              "model.layers.0.mlp.down_proj",
+              "model.layers.0.mlp.gate_proj",
+              "model.layers.0.mlp.up_proj",
+              "model.layers.0.self_attn.indexer.weights_proj",
+              "model.layers.0.self_attn.indexer.wk",
+              "model.layers.0.self_attn.indexer.wq_b",
+              "model.layers.0.self_attn.kv_a_proj_with_mqa",
+              "model.layers.0.self_attn.o_proj",
+              "model.layers.0.self_attn.q_a_proj",
+              "model.layers.0.self_attn.q_b_proj",
+              "model.layers.1.mlp.experts.0.down_proj",
+              "model.layers.1.mlp.experts.0.gate_proj",
+              "model.layers.1.mlp.experts.0.up_proj",
+              "model.layers.1.mlp.experts.1.down_proj",
+              "model.layers.1.mlp.experts.1.gate_proj",
+              "model.layers.1.mlp.experts.1.up_proj",
+              "model.layers.1.mlp.experts.2.down_proj",
+              "model.layers.1.mlp.experts.2.gate_proj",
+              "model.layers.1.mlp.experts.2.up_proj",
+              "model.layers.1.mlp.experts.3.down_proj",
+              "model.layers.1.mlp.experts.3.gate_proj",
+              "model.layers.1.mlp.experts.3.up_proj",
+              "model.layers.1.mlp.experts.4.down_proj",
+              "model.layers.1.mlp.experts.4.gate_proj",
+              "model.layers.1.mlp.experts.4.up_proj",
+              "model.layers.1.mlp.experts.5.down_proj",
+              "model.layers.1.mlp.experts.5.gate_proj",
+              "model.layers.1.mlp.experts.5.up_proj",
+              "model.layers.1.mlp.experts.6.down_proj",
+              "model.layers.1.mlp.experts.6.gate_proj",
+              "model.layers.1.mlp.experts.6.up_proj",
+              "model.layers.1.mlp.experts.7.down_proj",
+              "model.layers.1.mlp.experts.7.gate_proj",
+              "model.layers.1.mlp.experts.7.up_proj",
+              "model.layers.1.mlp.shared_experts.down_proj",
+              "model.layers.1.mlp.shared_experts.gate_proj",
+              "model.layers.1.mlp.shared_experts.up_proj",
+              "model.layers.1.self_attn.kv_a_proj_with_mqa",
+              "model.layers.1.self_attn.o_proj",
+              "model.layers.1.self_attn.q_a_proj",
+              "model.layers.1.self_attn.q_b_proj",
+              "model.layers.2.mlp.experts.0.down_proj",
+              "model.layers.2.mlp.experts.0.gate_proj",
+              "model.layers.2.mlp.experts.0.up_proj",
+              "model.layers.2.mlp.experts.1.down_proj",
+              "model.layers.2.mlp.experts.1.gate_proj",
+              "model.layers.2.mlp.experts.1.up_proj",
+              "model.layers.2.mlp.experts.2.down_proj",
+              "model.layers.2.mlp.experts.2.gate_proj",
+              "model.layers.2.mlp.experts.2.up_proj",
+              "model.layers.2.mlp.experts.3.down_proj",
+              "model.layers.2.mlp.experts.3.gate_proj",
+              "model.layers.2.mlp.experts.3.up_proj",
+              "model.layers.2.mlp.experts.4.down_proj",
+              "model.layers.2.mlp.experts.4.gate_proj",
+              "model.layers.2.mlp.experts.4.up_proj",
+              "model.layers.2.mlp.experts.5.down_proj",
+              "model.layers.2.mlp.experts.5.gate_proj",
+              "model.layers.2.mlp.experts.5.up_proj",
+              "model.layers.2.mlp.experts.6.down_proj",
+              "model.layers.2.mlp.experts.6.gate_proj",
+              "model.layers.2.mlp.experts.6.up_proj",
+              "model.layers.2.mlp.experts.7.down_proj",
+              "model.layers.2.mlp.experts.7.gate_proj",
+              "model.layers.2.mlp.experts.7.up_proj",
+              "model.layers.2.mlp.shared_experts.down_proj",
+              "model.layers.2.mlp.shared_experts.gate_proj",
+              "model.layers.2.mlp.shared_experts.up_proj",
+              "model.layers.2.self_attn.indexer.weights_proj",
+              "model.layers.2.self_attn.indexer.wk",
+              "model.layers.2.self_attn.indexer.wq_b",
+              "model.layers.2.self_attn.kv_a_proj_with_mqa",
+              "model.layers.2.self_attn.o_proj",
+              "model.layers.2.self_attn.q_a_proj",
+              "model.layers.2.self_attn.q_b_proj",
+              "model.layers.3.mlp.experts.0.down_proj",
+              "model.layers.3.mlp.experts.0.gate_proj",
+              "model.layers.3.mlp.experts.0.up_proj",
+              "model.layers.3.mlp.experts.1.down_proj",
+              "model.layers.3.mlp.experts.1.gate_proj",
+              "model.layers.3.mlp.experts.1.up_proj",
+              "model.layers.3.mlp.experts.2.down_proj",
+              "model.layers.3.mlp.experts.2.gate_proj",
+              "model.layers.3.mlp.experts.2.up_proj",
+              "model.layers.3.mlp.experts.3.down_proj",
+              "model.layers.3.mlp.experts.3.gate_proj",
+              "model.layers.3.mlp.experts.3.up_proj",
+              "model.layers.3.mlp.experts.4.down_proj",
+              "model.layers.3.mlp.experts.4.gate_proj",
+              "model.layers.3.mlp.experts.4.up_proj",
+              "model.layers.3.mlp.experts.5.down_proj",
+              "model.layers.3.mlp.experts.5.gate_proj",
+              "model.layers.3.mlp.experts.5.up_proj",
+              "model.layers.3.mlp.experts.6.down_proj",
+              "model.layers.3.mlp.experts.6.gate_proj",
+              "model.layers.3.mlp.experts.6.up_proj",
+              "model.layers.3.mlp.experts.7.down_proj",
+              "model.layers.3.mlp.experts.7.gate_proj",
+              "model.layers.3.mlp.experts.7.up_proj",
+              "model.layers.3.mlp.shared_experts.down_proj",
+              "model.layers.3.mlp.shared_experts.gate_proj",
+              "model.layers.3.mlp.shared_experts.up_proj",
+              "model.layers.3.self_attn.kv_a_proj_with_mqa",
+              "model.layers.3.self_attn.o_proj",
+              "model.layers.3.self_attn.q_a_proj",
+              "model.layers.3.self_attn.q_b_proj"
+            ],
+            "weights": {
+              "dynamic": false,
+              "group_size": 32,
+              "num_bits": 4,
+              "scale_dtype": "torch.uint8",
+              "strategy": "group",
+              "symmetric": true,
+              "type": "float"
+            }
+          }
+        },
+        "format": "mxfp4-pack-quantized",
+        "ignore": [],
+        "quant_method": "compressed-tensors",
+        "quantization_status": "compressed"
+      },
+      "reader_evidence": {
+        "decode_plan_sha256": "2bb2eef8f5dea9e0662e3fce6fbb6ecf22842b18c7192484e78ea5a8850ce9fa",
+        "gate_header_barrier": false,
+        "source_files": {
+          "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+          "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+          "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+          "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+        }
+      },
+      "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers",
+      "scientific_disclosure": "Stored-weight reconstruction only; no native quantized GEMM, activation arithmetic, or quantizer optimization quality claim.",
+      "scope": "weights_reconstructed"
+    },
+    "disclosures": [
+      {
+        "affects_comparability": false,
+        "code": "native_head_replay",
+        "detail": "HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.",
+        "severity": "info"
+      },
+      {
+        "affects_comparability": true,
+        "code": "weights_reconstructed",
+        "detail": "candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (microfloat-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.",
+        "severity": "caveat"
+      }
+    ],
+    "head_application": "native_head",
+    "kld_nats": 0.000866341033772075,
+    "measurement_scope": {
+      "contexts": 4,
+      "covers_full_panel": true,
+      "position_filter": "all",
+      "positions_per_context": 63,
+      "scored_positions": 252,
+      "subset_detail": null
+    },
+    "name": "ct-mxfp4",
+    "scope": {
+      "activation_quantization": null,
+      "assignments": [
+        {
+          "bits_per_weight": 4,
+          "format": "ct-mxfp4",
+          "layer_range": "all",
+          "note": "4 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 12 x native:bf16@16 (model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.kv_a_layernorm.weight); 6 x quantized:ct-mxfp4@4 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.other",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 4 x native:bf16@16 (model.layers.N.self_attn.kv_b_proj.weight); 12 x quantized:ct-mxfp4@4 (model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.q_a_proj.weight, model.layers.N.self_attn.q_b_proj.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.qkv",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: model.embed_tokens.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "embed_tokens",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: lm_head.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "lm_head",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-mxfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.down_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.down",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-mxfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.gate_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.gate",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-mxfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.up",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-mxfp4",
+          "layer_range": "all",
+          "note": "72 tensors: model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.experts",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x native:bf16@16 (model.layers.N.mlp.gate.weight); 3 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.router",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-mxfp4",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.shared_expert",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "norm",
+          "treatment": "native"
+        }
+      ],
+      "head_policy": "native",
+      "kv_cache_dtype": "bf16",
+      "mtp_included": false,
+      "policy": "mixed"
+    },
+    "source_checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e"
+  },
+  {
+    "checkpoint_identity_sha256": "8c943a9a4f3620b5a9014721f9f6565d25e868dd1e68407f28b618216a248317",
+    "classification": "lossy",
+    "comparison_receipt_sha256": "53824e510d77db71eefa6beb9403fbe7db4a4290cef2a119dc02b2ec9ffb3d1a",
+    "decode": {
+      "activation_quantization": null,
+      "checkpoint_bytes_read": 136664,
+      "comparison_class": "advisory",
+      "components_consumed": 330,
+      "components_not_consumed": [],
+      "components_planned": 330,
+      "formats_decoded": {
+        "nvfp4-pack-quantized": 110
+      },
+      "method": "microfloat-weight-reconstruction",
+      "modules_decoded": 110,
+      "modules_not_decoded": [],
+      "modules_planned": 110,
+      "output_dtype": "bfloat16",
+      "quantization_config": {
+        "config_groups": {
+          "group_0": {
+            "format": "nvfp4-pack-quantized",
+            "input_activations": null,
+            "output_activations": null,
+            "targets": [
+              "model.layers.0.mlp.down_proj",
+              "model.layers.0.mlp.gate_proj",
+              "model.layers.0.mlp.up_proj",
+              "model.layers.0.self_attn.indexer.weights_proj",
+              "model.layers.0.self_attn.indexer.wk",
+              "model.layers.0.self_attn.indexer.wq_b",
+              "model.layers.0.self_attn.kv_a_proj_with_mqa",
+              "model.layers.0.self_attn.kv_b_proj",
+              "model.layers.0.self_attn.o_proj",
+              "model.layers.0.self_attn.q_a_proj",
+              "model.layers.0.self_attn.q_b_proj",
+              "model.layers.1.mlp.experts.0.down_proj",
+              "model.layers.1.mlp.experts.0.gate_proj",
+              "model.layers.1.mlp.experts.0.up_proj",
+              "model.layers.1.mlp.experts.1.down_proj",
+              "model.layers.1.mlp.experts.1.gate_proj",
+              "model.layers.1.mlp.experts.1.up_proj",
+              "model.layers.1.mlp.experts.2.down_proj",
+              "model.layers.1.mlp.experts.2.gate_proj",
+              "model.layers.1.mlp.experts.2.up_proj",
+              "model.layers.1.mlp.experts.3.down_proj",
+              "model.layers.1.mlp.experts.3.gate_proj",
+              "model.layers.1.mlp.experts.3.up_proj",
+              "model.layers.1.mlp.experts.4.down_proj",
+              "model.layers.1.mlp.experts.4.gate_proj",
+              "model.layers.1.mlp.experts.4.up_proj",
+              "model.layers.1.mlp.experts.5.down_proj",
+              "model.layers.1.mlp.experts.5.gate_proj",
+              "model.layers.1.mlp.experts.5.up_proj",
+              "model.layers.1.mlp.experts.6.down_proj",
+              "model.layers.1.mlp.experts.6.gate_proj",
+              "model.layers.1.mlp.experts.6.up_proj",
+              "model.layers.1.mlp.experts.7.down_proj",
+              "model.layers.1.mlp.experts.7.gate_proj",
+              "model.layers.1.mlp.experts.7.up_proj",
+              "model.layers.1.mlp.shared_experts.down_proj",
+              "model.layers.1.mlp.shared_experts.gate_proj",
+              "model.layers.1.mlp.shared_experts.up_proj",
+              "model.layers.1.self_attn.kv_a_proj_with_mqa",
+              "model.layers.1.self_attn.kv_b_proj",
+              "model.layers.1.self_attn.o_proj",
+              "model.layers.1.self_attn.q_a_proj",
+              "model.layers.1.self_attn.q_b_proj",
+              "model.layers.2.mlp.experts.0.down_proj",
+              "model.layers.2.mlp.experts.0.gate_proj",
+              "model.layers.2.mlp.experts.0.up_proj",
+              "model.layers.2.mlp.experts.1.down_proj",
+              "model.layers.2.mlp.experts.1.gate_proj",
+              "model.layers.2.mlp.experts.1.up_proj",
+              "model.layers.2.mlp.experts.2.down_proj",
+              "model.layers.2.mlp.experts.2.gate_proj",
+              "model.layers.2.mlp.experts.2.up_proj",
+              "model.layers.2.mlp.experts.3.down_proj",
+              "model.layers.2.mlp.experts.3.gate_proj",
+              "model.layers.2.mlp.experts.3.up_proj",
+              "model.layers.2.mlp.experts.4.down_proj",
+              "model.layers.2.mlp.experts.4.gate_proj",
+              "model.layers.2.mlp.experts.4.up_proj",
+              "model.layers.2.mlp.experts.5.down_proj",
+              "model.layers.2.mlp.experts.5.gate_proj",
+              "model.layers.2.mlp.experts.5.up_proj",
+              "model.layers.2.mlp.experts.6.down_proj",
+              "model.layers.2.mlp.experts.6.gate_proj",
+              "model.layers.2.mlp.experts.6.up_proj",
+              "model.layers.2.mlp.experts.7.down_proj",
+              "model.layers.2.mlp.experts.7.gate_proj",
+              "model.layers.2.mlp.experts.7.up_proj",
+              "model.layers.2.mlp.shared_experts.down_proj",
+              "model.layers.2.mlp.shared_experts.gate_proj",
+              "model.layers.2.mlp.shared_experts.up_proj",
+              "model.layers.2.self_attn.indexer.weights_proj",
+              "model.layers.2.self_attn.indexer.wk",
+              "model.layers.2.self_attn.indexer.wq_b",
+              "model.layers.2.self_attn.kv_a_proj_with_mqa",
+              "model.layers.2.self_attn.kv_b_proj",
+              "model.layers.2.self_attn.o_proj",
+              "model.layers.2.self_attn.q_a_proj",
+              "model.layers.2.self_attn.q_b_proj",
+              "model.layers.3.mlp.experts.0.down_proj",
+              "model.layers.3.mlp.experts.0.gate_proj",
+              "model.layers.3.mlp.experts.0.up_proj",
+              "model.layers.3.mlp.experts.1.down_proj",
+              "model.layers.3.mlp.experts.1.gate_proj",
+              "model.layers.3.mlp.experts.1.up_proj",
+              "model.layers.3.mlp.experts.2.down_proj",
+              "model.layers.3.mlp.experts.2.gate_proj",
+              "model.layers.3.mlp.experts.2.up_proj",
+              "model.layers.3.mlp.experts.3.down_proj",
+              "model.layers.3.mlp.experts.3.gate_proj",
+              "model.layers.3.mlp.experts.3.up_proj",
+              "model.layers.3.mlp.experts.4.down_proj",
+              "model.layers.3.mlp.experts.4.gate_proj",
+              "model.layers.3.mlp.experts.4.up_proj",
+              "model.layers.3.mlp.experts.5.down_proj",
+              "model.layers.3.mlp.experts.5.gate_proj",
+              "model.layers.3.mlp.experts.5.up_proj",
+              "model.layers.3.mlp.experts.6.down_proj",
+              "model.layers.3.mlp.experts.6.gate_proj",
+              "model.layers.3.mlp.experts.6.up_proj",
+              "model.layers.3.mlp.experts.7.down_proj",
+              "model.layers.3.mlp.experts.7.gate_proj",
+              "model.layers.3.mlp.experts.7.up_proj",
+              "model.layers.3.mlp.shared_experts.down_proj",
+              "model.layers.3.mlp.shared_experts.gate_proj",
+              "model.layers.3.mlp.shared_experts.up_proj",
+              "model.layers.3.self_attn.kv_a_proj_with_mqa",
+              "model.layers.3.self_attn.kv_b_proj",
+              "model.layers.3.self_attn.o_proj",
+              "model.layers.3.self_attn.q_a_proj",
+              "model.layers.3.self_attn.q_b_proj"
+            ],
+            "weights": {
+              "dynamic": false,
+              "group_size": 16,
+              "num_bits": 4,
+              "scale_dtype": "torch.float8_e4m3fn",
+              "strategy": "tensor_group",
+              "symmetric": true,
+              "type": "float"
+            }
+          }
+        },
+        "format": "nvfp4-pack-quantized",
+        "ignore": [],
+        "quant_method": "compressed-tensors",
+        "quantization_status": "compressed"
+      },
+      "reader_evidence": {
+        "decode_plan_sha256": "b460ed95a3e245541b4d510acbaa5efd96d1d87e0b04d2da3d97e814ee542427",
+        "gate_header_barrier": false,
+        "source_files": {
+          "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+          "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+          "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+          "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+        }
+      },
+      "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers",
+      "scientific_disclosure": "Stored-weight reconstruction only; no native quantized GEMM, activation arithmetic, or quantizer optimization quality claim.",
+      "scope": "weights_reconstructed"
+    },
+    "disclosures": [
+      {
+        "affects_comparability": false,
+        "code": "native_head_replay",
+        "detail": "HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.",
+        "severity": "info"
+      },
+      {
+        "affects_comparability": true,
+        "code": "weights_reconstructed",
+        "detail": "candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (microfloat-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.",
+        "severity": "caveat"
+      }
+    ],
+    "head_application": "native_head",
+    "kld_nats": 0.0007752665823432193,
+    "measurement_scope": {
+      "contexts": 4,
+      "covers_full_panel": true,
+      "position_filter": "all",
+      "positions_per_context": 63,
+      "scored_positions": 252,
+      "subset_detail": null
+    },
+    "name": "ct-nvfp4",
+    "scope": {
+      "activation_quantization": null,
+      "assignments": [
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "4 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 12 x native:bf16@16 (model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.kv_a_layernorm.weight); 6 x quantized:ct-nvfp4@4 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.other",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "16 tensors: model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight, model.layers.N.self_attn.q_b_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.qkv",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: model.embed_tokens.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "embed_tokens",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: lm_head.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "lm_head",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.down_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.down",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.gate_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.gate",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.up",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "72 tensors: model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.experts",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x native:bf16@16 (model.layers.N.mlp.gate.weight); 3 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.router",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "ct-nvfp4",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.shared_expert",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "norm",
+          "treatment": "native"
+        }
+      ],
+      "head_policy": "native",
+      "kv_cache_dtype": "bf16",
+      "mtp_included": false,
+      "policy": "mixed"
+    },
+    "source_checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e"
+  },
+  {
+    "checkpoint_identity_sha256": "2fb12f3cf616cdb60b7003e9900d0cbc659c01df01c34cd0ae67f7ad5ad4bb6b",
+    "classification": "lossy",
+    "comparison_receipt_sha256": "5dcaf9dda1bf5084fa38578a18531f6aed8c99ee64aedffee8f70b1c4063bb4b",
+    "decode": {
+      "activation_quantization": null,
+      "checkpoint_bytes_read": 136664,
+      "comparison_class": "advisory",
+      "components_consumed": 330,
+      "components_not_consumed": [],
+      "components_planned": 330,
+      "formats_decoded": {
+        "modelopt-nvfp4": 110
+      },
+      "method": "microfloat-weight-reconstruction",
+      "modules_decoded": 110,
+      "modules_not_decoded": [],
+      "modules_planned": 110,
+      "output_dtype": "bfloat16",
+      "quantization_config": {
+        "group_size": 16,
+        "num_bits": 4,
+        "quant_algo": "NVFP4",
+        "quant_method": "modelopt",
+        "with_input_scale": false
+      },
+      "reader_evidence": {
+        "decode_plan_sha256": "bbdc8be9985af40d3a7adbcf3c895e105558d28af93c1b018b08d9e500a5c15e",
+        "gate_header_barrier": false,
+        "source_files": {
+          "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+          "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+          "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+          "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+        }
+      },
+      "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers",
+      "scientific_disclosure": "Stored-weight reconstruction only; no native quantized GEMM, activation arithmetic, or quantizer optimization quality claim.",
+      "scope": "weights_reconstructed"
+    },
+    "disclosures": [
+      {
+        "affects_comparability": false,
+        "code": "native_head_replay",
+        "detail": "HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.",
+        "severity": "info"
+      },
+      {
+        "affects_comparability": true,
+        "code": "weights_reconstructed",
+        "detail": "candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (microfloat-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.",
+        "severity": "caveat"
+      }
+    ],
+    "head_application": "native_head",
+    "kld_nats": 0.0007775931894536412,
+    "measurement_scope": {
+      "contexts": 4,
+      "covers_full_panel": true,
+      "position_filter": "all",
+      "positions_per_context": 63,
+      "scored_positions": 252,
+      "subset_detail": null
+    },
+    "name": "modelopt-nvfp4",
+    "scope": {
+      "activation_quantization": null,
+      "assignments": [
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "4 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 12 x native:bf16@16 (model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.kv_a_layernorm.weight); 6 x quantized:modelopt-nvfp4@4 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.other",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "16 tensors: model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight, model.layers.N.self_attn.q_b_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.qkv",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: model.embed_tokens.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "embed_tokens",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: lm_head.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "lm_head",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.down_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.down",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.gate_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.gate",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.up",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "72 tensors: model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.experts",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x native:bf16@16 (model.layers.N.mlp.gate.weight); 3 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.router",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.shared_expert",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "norm",
+          "treatment": "native"
+        }
+      ],
+      "head_policy": "native",
+      "kv_cache_dtype": "bf16",
+      "mtp_included": false,
+      "policy": "mixed"
+    },
+    "source_checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e"
+  },
+  {
+    "checkpoint_identity_sha256": "7239a3bd9a4ca38a32c32d110586c688197ee04f6662bc900a175d3107ad9a18",
+    "classification": "lossy",
+    "comparison_receipt_sha256": "0db49970160ff0d412b90abce7062ef372cd1ad62c016cf9614cb6a834f3470e",
+    "decode": {
+      "activation_quantization": {
+        "applied_to_weights": false,
+        "declared": {
+          "activation_scheme": "dynamic",
+          "scale_fmt": "ue8m0"
+        },
+        "kv_cache_scheme": null,
+        "scope": "weights-reconstructed/advisory; activation-not-captured",
+        "stored_components": []
+      },
+      "checkpoint_bytes_read": 242286,
+      "comparison_class": "advisory",
+      "components_consumed": 220,
+      "components_not_consumed": [],
+      "components_planned": 220,
+      "formats_decoded": {
+        "fp8-ue8m0": 110
+      },
+      "method": "microfloat-weight-reconstruction",
+      "modules_decoded": 110,
+      "modules_not_decoded": [],
+      "modules_planned": 110,
+      "output_dtype": "bfloat16",
+      "quantization_config": {
+        "activation_scheme": "dynamic",
+        "fmt": "e4m3",
+        "quant_method": "fp8",
+        "scale_fmt": "ue8m0",
+        "weight_block_size": [
+          128,
+          128
+        ]
+      },
+      "reader_evidence": {
+        "decode_plan_sha256": "5ba92234f4b0487318ab6ee1ebd8c7bdc20a826d3376a4ad5933500488a6ca6f",
+        "gate_header_barrier": false,
+        "source_files": {
+          "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+          "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+          "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+          "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+        }
+      },
+      "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers",
+      "scientific_disclosure": "Stored-weight reconstruction only; no native quantized GEMM, activation arithmetic, or quantizer optimization quality claim.",
+      "scope": "weights_reconstructed"
+    },
+    "disclosures": [
+      {
+        "affects_comparability": false,
+        "code": "native_head_replay",
+        "detail": "HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.",
+        "severity": "info"
+      },
+      {
+        "affects_comparability": true,
+        "code": "weights_reconstructed",
+        "detail": "candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (microfloat-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.",
+        "severity": "caveat"
+      },
+      {
+        "affects_comparability": true,
+        "code": "activation_quantization_not_captured",
+        "detail": "candidate was captured from a bf16 materialisation of its weights (microfloat-weight-reconstruction); the checkpoint declares activation quantization (dynamic, reader-declared activation quantization; see sealed decode evidence) that a weights-only capture does not apply, so a served deployment also quantizes activations at runtime. That term is not in this number, which is expected to understate the served divergence; it is not a mathematical bound. The comparison is advisory.",
+        "severity": "caveat"
+      }
+    ],
+    "head_application": "native_head",
+    "kld_nats": 0.00020246177849756768,
+    "measurement_scope": {
+      "contexts": 4,
+      "covers_full_panel": true,
+      "position_filter": "all",
+      "positions_per_context": 63,
+      "scored_positions": 252,
+      "subset_detail": null
+    },
+    "name": "fp8-ue8m0",
+    "scope": {
+      "activation_quantization": null,
+      "assignments": [
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "4 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 12 x native:bf16@16 (model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.kv_a_layernorm.weight); 6 x quantized:fp8-ue8m0@8 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.other",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "16 tensors: model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight, model.layers.N.self_attn.q_b_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.qkv",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: model.embed_tokens.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "embed_tokens",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: lm_head.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "lm_head",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.down_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.down",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.gate_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.gate",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.up",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "72 tensors: model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.experts",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x native:bf16@16 (model.layers.N.mlp.gate.weight); 3 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.router",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "fp8-ue8m0",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.shared_expert",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "norm",
+          "treatment": "native"
+        }
+      ],
+      "head_policy": "native",
+      "kv_cache_dtype": "bf16",
+      "mtp_included": false,
+      "policy": "mixed"
+    },
+    "source_checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e"
+  },
+  {
+    "checkpoint_identity_sha256": "822ca7cc58e9909f2832e980051622ccb4a4a4ce1472c618a45e3f4863beb013",
+    "classification": "lossy",
+    "comparison_receipt_sha256": "3a8413a6ebb35c46f92ba0ef5513d8fa9ce93cb26024f2d00bf5a4b1bf176136",
+    "decode": {
+      "activation_quantization": null,
+      "checkpoint_bytes_read": 207190,
+      "comparison_class": "advisory",
+      "components_consumed": 257,
+      "components_not_consumed": [],
+      "components_planned": 257,
+      "formats_decoded": {
+        "modelopt-fp8-block": 36,
+        "modelopt-fp8-tensor": 37,
+        "modelopt-nvfp4": 37
+      },
+      "method": "microfloat-weight-reconstruction",
+      "modules_decoded": 110,
+      "modules_not_decoded": [],
+      "modules_planned": 110,
+      "output_dtype": "bfloat16",
+      "quantization_config": {
+        "quant_algo": "MIXED_PRECISION",
+        "quant_method": "modelopt",
+        "quantized_layers": {
+          "model.layers.0.mlp.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.0.mlp.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.0.mlp.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.0.self_attn.indexer.weights_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.0.self_attn.indexer.wk": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.0.self_attn.indexer.wq_b": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.0.self_attn.kv_a_proj_with_mqa": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.0.self_attn.kv_b_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.0.self_attn.o_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.0.self_attn.q_a_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.0.self_attn.q_b_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.0.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.0.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.0.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.1.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.1.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.1.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.2.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.2.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.2.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.3.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.3.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.3.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.4.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.4.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.4.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.5.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.5.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.5.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.6.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.6.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.6.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.experts.7.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.experts.7.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.experts.7.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.mlp.shared_experts.down_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.mlp.shared_experts.gate_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.mlp.shared_experts.up_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.self_attn.kv_a_proj_with_mqa": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.self_attn.kv_b_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.1.self_attn.o_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.1.self_attn.q_a_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.1.self_attn.q_b_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.0.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.0.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.0.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.1.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.1.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.1.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.2.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.2.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.2.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.3.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.3.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.3.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.4.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.4.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.4.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.5.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.5.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.5.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.6.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.6.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.6.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.experts.7.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.experts.7.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.experts.7.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.mlp.shared_experts.down_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.mlp.shared_experts.gate_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.mlp.shared_experts.up_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.self_attn.indexer.weights_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.self_attn.indexer.wk": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.self_attn.indexer.wq_b": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.self_attn.kv_a_proj_with_mqa": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.self_attn.kv_b_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.2.self_attn.o_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.2.self_attn.q_a_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.2.self_attn.q_b_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.0.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.0.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.0.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.1.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.1.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.1.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.2.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.2.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.2.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.3.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.3.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.3.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.4.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.4.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.4.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.5.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.5.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.5.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.6.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.6.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.6.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.experts.7.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.experts.7.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.experts.7.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.mlp.shared_experts.down_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.mlp.shared_experts.gate_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.mlp.shared_experts.up_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.self_attn.kv_a_proj_with_mqa": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.self_attn.kv_b_proj": {
+            "quant_algo": "FP8"
+          },
+          "model.layers.3.self_attn.o_proj": {
+            "group_size": 128,
+            "quant_algo": "FP8_PB_WO"
+          },
+          "model.layers.3.self_attn.q_a_proj": {
+            "group_size": 16,
+            "quant_algo": "NVFP4"
+          },
+          "model.layers.3.self_attn.q_b_proj": {
+            "quant_algo": "FP8"
+          }
+        }
+      },
+      "reader_evidence": {
+        "decode_plan_sha256": "de7e9aa33267002a2ab5284d6438d4e3dff7617b953904356ff486590972f79f",
+        "gate_header_barrier": false,
+        "source_files": {
+          "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+          "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+          "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+          "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+        }
+      },
+      "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers",
+      "scientific_disclosure": "Stored-weight reconstruction only; no native quantized GEMM, activation arithmetic, or quantizer optimization quality claim.",
+      "scope": "weights_reconstructed"
+    },
+    "disclosures": [
+      {
+        "affects_comparability": false,
+        "code": "native_head_replay",
+        "detail": "HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.",
+        "severity": "info"
+      },
+      {
+        "affects_comparability": true,
+        "code": "weights_reconstructed",
+        "detail": "candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (microfloat-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.",
+        "severity": "caveat"
+      }
+    ],
+    "head_application": "native_head",
+    "kld_nats": 0.0005261778079916031,
+    "measurement_scope": {
+      "contexts": 4,
+      "covers_full_panel": true,
+      "position_filter": "all",
+      "positions_per_context": 63,
+      "scored_positions": 252,
+      "subset_detail": null
+    },
+    "name": "modelopt-mixed",
+    "scope": {
+      "activation_quantization": null,
+      "assignments": [
+        {
+          "bits_per_weight": 8,
+          "format": "modelopt-fp8-block",
+          "layer_range": "0,3",
+          "note": "2 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "modelopt-fp8-tensor",
+          "layer_range": "1",
+          "note": "1 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "2",
+          "note": "1 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.o",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 12 x native:bf16@16 (model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.kv_a_layernorm.weight); 2 x quantized:modelopt-fp8-block@8 (model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight); 2 x quantized:modelopt-fp8-tensor@8 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight); 2 x quantized:modelopt-nvfp4@4 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wq_b.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.other",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "0-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 4 x quantized:modelopt-fp8-block@8 (model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight); 6 x quantized:modelopt-fp8-tensor@8 (model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight); 6 x quantized:modelopt-nvfp4@4 (model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.kv_b_proj.weight, model.layers.N.self_attn.q_a_proj.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "attn.qkv",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: model.embed_tokens.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "embed_tokens",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "1 tensors: lm_head.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "lm_head",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": 4,
+          "format": "modelopt-nvfp4",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.down_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.down",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "modelopt-fp8-tensor",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.gate_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.gate",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 8,
+          "format": "modelopt-fp8-block",
+          "layer_range": "all",
+          "note": "1 tensors: model.layers.N.mlp.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "mlp.up",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 24 x quantized:modelopt-fp8-block@8 (model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight); 24 x quantized:modelopt-fp8-tensor@8 (model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight); 24 x quantized:modelopt-nvfp4@4 (model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.experts",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x native:bf16@16 (model.layers.N.mlp.gate.weight); 3 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.router",
+          "treatment": "native"
+        },
+        {
+          "bits_per_weight": null,
+          "format": "mixed",
+          "layer_range": "1-3",
+          "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x quantized:modelopt-fp8-block@8 (model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight); 3 x quantized:modelopt-fp8-tensor@8 (model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight); 3 x quantized:modelopt-nvfp4@4 (model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "moe.shared_expert",
+          "treatment": "quantized"
+        },
+        {
+          "bits_per_weight": 16,
+          "format": "bf16",
+          "layer_range": "all",
+          "note": "9 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+          "tensor_class": "norm",
+          "treatment": "native"
+        }
+      ],
+      "head_policy": "native",
+      "kv_cache_dtype": "bf16",
+      "mtp_included": false,
+      "policy": "mixed"
+    },
+    "source_checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e"
+  }
+]

raw/candidates/ct-mxfp4/evidence/candidate/README.md

@@ -0,0 +1,95 @@
+---
+license: mit
+tags:
+- fidelity
+- kl-divergence
+- fidelity-provenance
+- fidelity-dataset
+- fidelity-quant
+x_fidelity:
+  spec: https://github.com/malaiwah/quant-fidelity-suite/blob/main/docs/FIDELITY-DATASET-SPEC.md
+  spec_version: fidelity-provenance/v1
+  role: fidelity-dataset
+  captured_model:
+    repository: /tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4
+    revision: null
+    role: quant
+  form: hidden
+  lane: other
+  panel:
+    panel_id: panel--panel
+    suite_token_hash_sha256: 525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36
+    contexts: 4
+    scored_positions: 252
+    repository: null
+    revision: null
+  head:
+    policy: native
+    quantized: false
+    bits: 16
+    lm_head_tensor_content_sha256: 2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510
+    lm_head_file_sha256: 07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913
+    final_norm_tensor_content_sha256: null
+    final_norm_file_sha256: null
+    equality_receipt: null
+    replay_permitted: true
+  scope_digest: attn.o=quantized:ct-mxfp4@4|attn.other=quantized:mixed|attn.qkv=quantized:mixed|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=quantized:ct-mxfp4@4|mlp.gate=quantized:ct-mxfp4@4|mlp.up=quantized:ct-mxfp4@4|moe.experts=quantized:ct-mxfp4@4|moe.router=native:mixed|moe.shared_expert=quantized:ct-mxfp4@4|norm=native:bf16@16|head=native|kv=bf16
+  seal:
+    manifest: fidelity-dataset.json
+    dataset_sha256: null
+    checksums: checksums.txt
+    checksums_sha256: '0000000000000000000000000000000000000000000000000000000000000000'
+    note: 'written before the seal: dataset_sha256 and checksums_sha256 are digests of the tree this file
+      is inside, so they cannot appear in it. Read them from fidelity-dataset.json.'
+  interop:
+    compatible_with:
+    - kimi-k3-distribution-fidelity/1
+    k3_compat_emitted: false
+  registry:
+    dataset: malaiwah/quant-fidelity-registry
+    schema_version: quant-fidelity-registry/v1
+    artifact_id: null
+    measurement_ids: []
+    snapshot: null
+  fidelity_dataset:
+    repository: null
+    revision: null
+    dataset_sha256: null
+    capture_content_digest: 95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876
+    form: hidden
+    role: quant
+  measurements: []
+---
+# Local quant integration candidate
+
+A **quant** fidelity dataset in **hidden form**, produced by `engines/tools/hf_capture.py` from `/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4`.
+
+## The cut
+
+the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits after it). Same cut as engines/tools/hidden_replay.py and as Festr's kimi-k3 hidden-replay qualification.
+
+## What is here
+
+| | |
+|---|---|
+| records | 4 |
+| scored positions | 252 |
+| context length | 64 |
+| hidden width | 64 |
+| vocab size | 260 |
+| capture bytes (hidden form) | 33348 |
+| the same capture in logit form would be | 262080 bytes (8x) |
+| scope policy | mixed |
+| lane | other |
+
+## Does the model still generate sensibly
+
+`The capital of France is` -> `�` (p=0.0068, entropy 5.548 of a possible 5.561 nats). Recorded, not enforced. One extra window through the same schedule; its hidden state was discarded and is not part of this dataset.
+
+## How to check it
+
+```
+bin/fidelity-dataset verify <this directory>
+```
+
+Full specification: <https://github.com/malaiwah/quant-fidelity-suite/blob/main/docs/FIDELITY-DATASET-SPEC.md>

raw/candidates/ct-mxfp4/evidence/candidate/capture/hidden_0000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d1f5e1f4c7b7306fe5734fc8327b15888b0783aba848e04ae760404c5037960
+size 8337

raw/candidates/ct-mxfp4/evidence/candidate/capture/hidden_0001.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5749a836a7ed4349e649e7c9a0a641f990f1aee64b762540dd3af6977aa56ee
+size 8337

raw/candidates/ct-mxfp4/evidence/candidate/capture/hidden_0002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0a719e3c2aa47a15bea2c00572b37fe09a9877384a4b388c1ace9666496e4db
+size 8337

raw/candidates/ct-mxfp4/evidence/candidate/capture/hidden_0003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb02c95788a64ef2cc00523c399efb7b62586bd286e74af0df9b9c8426536149
+size 8337

raw/candidates/ct-mxfp4/evidence/candidate/capture/manifest.json

@@ -0,0 +1,159 @@
+{
+  "capture_content_digest": "95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876",
+  "context_length": 64,
+  "coverage": {
+    "complete": true,
+    "declared_records": 4,
+    "index_range": [
+      0,
+      3
+    ],
+    "missing_indices_sample": [],
+    "present_records": 4,
+    "shard_of": null,
+    "subset_detail": null
+  },
+  "created_utc": "2026-09-07T16:52:52Z",
+  "dtype": "BF16",
+  "dtype_lossless": true,
+  "form": "hidden",
+  "format_version": 1,
+  "hidden_width": 64,
+  "receipt_sha256": "d7ade9a3dec0af4823385c1146a6196d4fa4fc53496780601c74407ff8ca3c5a",
+  "records": [
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 0,
+      "document_id": "embedded-synthetic/garden.txt",
+      "domain": "garden",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.038,
+      "file": "hidden_0000.safetensors",
+      "index": 0,
+      "key": "hidden_states",
+      "payload_sha256": "a8eb741d7fd134901c2dc959779109fc4c8abbf00bfba61fcb0ca1d66ae93dee",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "0d1f5e1f4c7b7306fe5734fc8327b15888b0783aba848e04ae760404c5037960",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "a8eb741d7fd134901c2dc959779109fc4c8abbf00bfba61fcb0ca1d66ae93dee",
+      "token_ids_json_sha256": "08f8d8d449f7f679b24ada75dc4db5b0108508e16ccefb0f61cad6144593e315",
+      "token_ids_sha256_legacy": "9d5378f0316be57e37f80d2bae7895ba4a1140c302c87ff8cce2fbd9afc83b77",
+      "window_id": "final-0000",
+      "window_index": 0
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 1,
+      "document_id": "embedded-synthetic/arithmetic.txt",
+      "domain": "arithmetic",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.031,
+      "file": "hidden_0001.safetensors",
+      "index": 1,
+      "key": "hidden_states",
+      "payload_sha256": "8e3dfc446f7a189a020abc054ac5f400281a6995ee4020eacbc735ca9960faff",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "a5749a836a7ed4349e649e7c9a0a641f990f1aee64b762540dd3af6977aa56ee",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "8e3dfc446f7a189a020abc054ac5f400281a6995ee4020eacbc735ca9960faff",
+      "token_ids_json_sha256": "5bc88e4e47faf077b2fc85ae0758eee5308003ff0fefa5d6691a7f24f22ecc46",
+      "token_ids_sha256_legacy": "e0f95114b71a46cb489539dc38909fb67fb378f25fafbb513e32690cc7f11b37",
+      "window_id": "final-0001",
+      "window_index": 1
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 2,
+      "document_id": "embedded-synthetic/code.txt",
+      "domain": "code",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.03,
+      "file": "hidden_0002.safetensors",
+      "index": 2,
+      "key": "hidden_states",
+      "payload_sha256": "039ee5458e561911aa87d5f9848afbf0cbe81f7a30055e63c4dbc9bd4bb93508",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "f0a719e3c2aa47a15bea2c00572b37fe09a9877384a4b388c1ace9666496e4db",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "039ee5458e561911aa87d5f9848afbf0cbe81f7a30055e63c4dbc9bd4bb93508",
+      "token_ids_json_sha256": "51b956fbf40161ce10fe3da6c1e186368a567676467030535468064aea0d7d07",
+      "token_ids_sha256_legacy": "482b890c7b0e2f7400257c43680093b7f2b90d6df5361c6de3ae97ddfdd2504e",
+      "window_id": "final-0002",
+      "window_index": 2
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 3,
+      "document_id": "embedded-synthetic/instructions.txt",
+      "domain": "instructions",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.031,
+      "file": "hidden_0003.safetensors",
+      "index": 3,
+      "key": "hidden_states",
+      "payload_sha256": "bc3708dc8fcdd573a450ff6deac4a699b7c03c18a3c5f807a10da328c0704450",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "fb02c95788a64ef2cc00523c399efb7b62586bd286e74af0df9b9c8426536149",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "bc3708dc8fcdd573a450ff6deac4a699b7c03c18a3c5f807a10da328c0704450",
+      "token_ids_json_sha256": "8653267d2a6dc67550759c29ccb80a7a5529d927e6c367c681f5018d55497a2d",
+      "token_ids_sha256_legacy": "12e01edd3900ff2ca82621feb5504f24b003a53c0afe98a9ee0b490eb93141d4",
+      "window_id": "final-0003",
+      "window_index": 3
+    }
+  ],
+  "run_name": "hf-capture",
+  "runtime_manifest": "../runtime/capture-runtime.json",
+  "runtime_manifest_sha256": "6838263dda3c253365372dece3f23a449687b4cf1a2175b54d9f264da3e16748",
+  "schema": "malaiwah.fidelity-capture-manifest.v1",
+  "scored_rows_per_context": 63,
+  "semantic_point": "after_final_rmsnorm_before_lm_head",
+  "tensor_key": "hidden_states",
+  "total_scored_rows": 252,
+  "total_size_bytes": 33348,
+  "vocab_size": 260
+}

raw/candidates/ct-mxfp4/evidence/candidate/checksums.txt

@@ -0,0 +1,20 @@
+9456750e9e6193900df9bd4361ce92e6f7539b62255a120bbe9e3afdbc4880e9  README.md
+0d1f5e1f4c7b7306fe5734fc8327b15888b0783aba848e04ae760404c5037960  capture/hidden_0000.safetensors
+a5749a836a7ed4349e649e7c9a0a641f990f1aee64b762540dd3af6977aa56ee  capture/hidden_0001.safetensors
+f0a719e3c2aa47a15bea2c00572b37fe09a9877384a4b388c1ace9666496e4db  capture/hidden_0002.safetensors
+fb02c95788a64ef2cc00523c399efb7b62586bd286e74af0df9b9c8426536149  capture/hidden_0003.safetensors
+78ea9d1cadee7df930e55fa9dbac6a6cd8fa457cffcaa7d1ac7aab166d72ff1f  capture/manifest.json
+777b038f747bed25d1f5db8adbfd2a78c7762fb5c870ff8c3c785a984d48a168  head/head.json
+07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913  head/weight.safetensors
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0000.npy
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0001.npy
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0002.npy
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0003.npy
+b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263  panel/panel-receipt.json
+68efa258ea3c5d4ba759c3c15aadd6e8c984711831bea3f335903109345cdcef  panel/panel.json
+08f8d8d449f7f679b24ada75dc4db5b0108508e16ccefb0f61cad6144593e315  panel/tokens/context-0000.json
+5bc88e4e47faf077b2fc85ae0758eee5308003ff0fefa5d6691a7f24f22ecc46  panel/tokens/context-0001.json
+51b956fbf40161ce10fe3da6c1e186368a567676467030535468064aea0d7d07  panel/tokens/context-0002.json
+8653267d2a6dc67550759c29ccb80a7a5529d927e6c367c681f5018d55497a2d  panel/tokens/context-0003.json
+6838263dda3c253365372dece3f23a449687b4cf1a2175b54d9f264da3e16748  runtime/capture-runtime.json
+d3024ee8845d123018358b9c98a844bcdc3c985fbbc92d9e1b8565b121fe8a45  validation/structural-validation.json

raw/candidates/ct-mxfp4/evidence/candidate/fidelity-dataset.json

@@ -0,0 +1,422 @@
+{
+  "capture": {
+    "capture_content_digest": "95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876",
+    "dtype": "BF16",
+    "dtype_lossless": true,
+    "form": "hidden",
+    "head_not_separable_reason": null,
+    "head_separable": true,
+    "hidden_width": 64,
+    "lossy_codec": null,
+    "manifest_file": "capture/manifest.json",
+    "manifest_file_sha256": "78ea9d1cadee7df930e55fa9dbac6a6cd8fa457cffcaa7d1ac7aab166d72ff1f",
+    "records_count": 4,
+    "scored_rows_total": 252,
+    "semantic_point": "after_final_rmsnorm_before_lm_head",
+    "tensor_key": "hidden_states",
+    "total_size_bytes": 33348,
+    "vocab_size": 260
+  },
+  "coverage": {
+    "complete": true,
+    "declared_records": 4,
+    "index_range": [
+      0,
+      3
+    ],
+    "missing_indices_sample": [],
+    "present_records": 4,
+    "shard_of": null,
+    "subset_detail": null
+  },
+  "created_utc": "2026-09-07T16:52:52Z",
+  "dataset": {
+    "author": {
+      "handle": null,
+      "is_registry_maintainer": false,
+      "name": "malaiwah",
+      "role": "capture-author",
+      "url": null
+    },
+    "base_capture": null,
+    "id": "fidelity--quant-integration.candidate",
+    "license": "mit",
+    "name": "Local quant integration candidate",
+    "qualification": null,
+    "repository": null,
+    "revision": null,
+    "role": "quant",
+    "structural_status": "sealed"
+  },
+  "dataset_sha256": "7bbd1d3e4f4d78f1a154187e48b138acfbb557c5e9d740e686036101501dc761",
+  "determinism": {
+    "cold_start_per_run": true,
+    "distinct_evidence_hash_count": 1,
+    "evidence_hashes": [
+      "95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876"
+    ],
+    "evidence_kind": "hidden_state_tensor_sha256",
+    "identical_across_runs": null,
+    "note": "a single cold capture; cross-run determinism is NOT asserted here. Capture a second cold run and run `fidelity-dataset compare --self-compare` -- an exactly 0.0 result is the SC-1 reproduction confirmation.",
+    "repeat_noise": null,
+    "repeats": [],
+    "run_count": 1
+  },
+  "disclosures": [
+    {
+      "affects_comparability": false,
+      "code": "rtn_format_fixture",
+      "detail": "Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.",
+      "severity": "info"
+    },
+    {
+      "affects_comparability": false,
+      "code": "reduced_run_count",
+      "detail": "this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus `fidelity-dataset compare --self-compare`, whose exactly-0.0 result is the SC-1 reproduction confirmation.",
+      "severity": "caveat"
+    }
+  ],
+  "format_version": 1,
+  "generation_sanity_probe": {
+    "degenerate": false,
+    "enforced": false,
+    "entropy_nats": 5.548143460498403,
+    "expect": null,
+    "logit_spread": 0.984375,
+    "note": "one extra window through the same schedule; its hidden state is discarded and never enters the dataset.",
+    "prompt": "The capital of France is",
+    "schema": "malaiwah.generation-sanity-probe.v1",
+    "status": "recorded",
+    "top1_id": 113,
+    "top1_probability": 0.00677271665330186,
+    "top1_text": "�",
+    "top_k": [
+      {
+        "id": 113,
+        "probability": 0.00677271665330186,
+        "text": "�"
+      },
+      {
+        "id": 237,
+        "probability": 0.006263738959055852,
+        "text": "�"
+      },
+      {
+        "id": 193,
+        "probability": 0.005895742083529586,
+        "text": "\u0001"
+      },
+      {
+        "id": 236,
+        "probability": 0.0057143491692576796,
+        "text": "�"
+      },
+      {
+        "id": 240,
+        "probability": 0.005399680497105314,
+        "text": "�"
+      }
+    ],
+    "uniform_entropy_nats": 5.560681631015528,
+    "vocab_size": 260
+  },
+  "head": {
+    "applied_in_capture": false,
+    "bias": null,
+    "bits": 16,
+    "compat_tensor_key": "weight",
+    "dtype": "BF16",
+    "equality_receipt": null,
+    "file": "head/weight.safetensors",
+    "file_sha256": "07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913",
+    "final_norm": {
+      "applied_at_replay": false,
+      "applied_in_capture": true,
+      "dtype": null,
+      "file": null,
+      "file_sha256": null,
+      "shape": null,
+      "tensor_content_sha256": null,
+      "tensor_key": null
+    },
+    "head_json": "head/head.json",
+    "present": true,
+    "quantized": false,
+    "raw_tensor_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+    "shape": [
+      260,
+      64
+    ],
+    "source": "native",
+    "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+    "tensor_key": "lm_head.weight"
+  },
+  "interop": {
+    "adapted_from": null,
+    "compatible_with": [
+      "kimi-k3-distribution-fidelity/1"
+    ],
+    "divergences": [
+      {
+        "field": "head",
+        "id": "D-1",
+        "reason": "every capture declares its OWN head identity by tensor content digest; the comparator refuses a hidden-form comparison across differing heads."
+      },
+      {
+        "field": "runtime.lane",
+        "id": "D-2",
+        "reason": "lane is required and gated at compare time so registry BIAS-006 is checked before a number exists, not at submission."
+      },
+      {
+        "field": "panel.scoring_window",
+        "id": "D-3",
+        "reason": "scoring_window is part of panel identity, not a comparator flag: a score_from=1024 number is a different panel, not a variant."
+      },
+      {
+        "field": "comparison_kind",
+        "id": "D-4",
+        "reason": "self-compare is a first-class mode asserting exactly 0.0."
+      },
+      {
+        "field": "estimator.accumulation_dtype",
+        "id": "D-5",
+        "reason": "required, fp64 default; an fp32 estimator lands in a different comparability class on the same data and that must be visible."
+      },
+      {
+        "field": "coverage",
+        "id": "D-6",
+        "reason": "declared vs present record counts with shard_of; fixes our own manifest that claimed 5,120 captures in a 512-file repository."
+      },
+      {
+        "field": "panel",
+        "id": "D-7",
+        "reason": "the panel is a separately referenceable object, which is what makes a candidate capture publishable standalone."
+      },
+      {
+        "field": "capture.lossy_codec",
+        "id": "D-8",
+        "reason": "lets a llama.cpp .kld be ingested without laundering a lossy capture as exact."
+      },
+      {
+        "field": "capture.records[].attention_mask_sha256",
+        "id": "D-9",
+        "reason": "our packed and streaming lanes vary mask construction; a single-request capture path makes it invariant and has no equivalent field."
+      },
+      {
+        "field": "capture.head_separable",
+        "id": "D-10",
+        "reason": "hidden form is the default; logit form must declare head_separable:false with a reason."
+      }
+    ],
+    "k3_compat_emitted": false,
+    "note": null
+  },
+  "panel": {
+    "contamination": {
+      "checked": false,
+      "hits": null,
+      "method": "not established",
+      "receipt": null
+    },
+    "context_length": 64,
+    "contexts": 4,
+    "panel_file": "panel/panel.json",
+    "panel_file_sha256": "68efa258ea3c5d4ba759c3c15aadd6e8c984711831bea3f335903109345cdcef",
+    "panel_id": "panel--panel",
+    "panel_receipt_file": "panel/panel-receipt.json",
+    "panel_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5",
+    "panel_token_sha256_legacy": "3c9467c85be43d238926f6d18bf825c2e4748814c95b8b0e8b8f1420546c3089",
+    "remap_file": null,
+    "repository": null,
+    "revision": null,
+    "scored_positions_total": 252,
+    "scoring_window": {
+      "dropped_positions_total": 0,
+      "min_left_context_tokens": 1,
+      "policy": "every causal prediction position of every window is scored; nothing is dropped",
+      "score_from": 0,
+      "windowed": false
+    },
+    "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36",
+    "tokenizer": {
+      "files": [
+        {
+          "bytes": 1479,
+          "name": "config.json",
+          "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+        },
+        {
+          "bytes": 216,
+          "name": "generation_config.json",
+          "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+        },
+        {
+          "bytes": 6763,
+          "name": "tokenizer.json",
+          "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+        },
+        {
+          "bytes": 236,
+          "name": "tokenizer_config.json",
+          "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+        }
+      ],
+      "files_verified": true,
+      "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+      "maximum_token_id_exclusive": 260,
+      "receipt": null,
+      "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+      "vocab_size": 260
+    }
+  },
+  "runtime": {
+    "backend_identity_sha256": null,
+    "file": "runtime/capture-runtime.json",
+    "file_sha256": "6838263dda3c253365372dece3f23a449687b4cf1a2175b54d9f264da3e16748",
+    "lane": "other",
+    "lane_identity_sha256": "34f640ff350d24093734d97a81c1048ce0443d8b3c7bda0953f8283b8ea9269f",
+    "lane_inferred": false,
+    "runtime_reader_sha256": null,
+    "source": "native",
+    "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd"
+  },
+  "schema": "malaiwah.fidelity-dataset.v1",
+  "scope": {
+    "assignments": [
+      {
+        "bits_per_weight": 4,
+        "format": "ct-mxfp4",
+        "layer_range": "all",
+        "note": "4 tensors: model.layers.N.self_attn.o_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "attn.o",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": null,
+        "format": "mixed",
+        "layer_range": "0-3",
+        "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 12 x native:bf16@16 (model.layers.N.self_attn.indexer.k_norm.bias, model.layers.N.self_attn.indexer.k_norm.weight, model.layers.N.self_attn.kv_a_layernorm.weight); 6 x quantized:ct-mxfp4@4 (model.layers.N.self_attn.indexer.weights_proj.weight, model.layers.N.self_attn.indexer.wk.weight, model.layers.N.self_attn.indexer.wq_b.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "attn.other",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": null,
+        "format": "mixed",
+        "layer_range": "0-3",
+        "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 4 x native:bf16@16 (model.layers.N.self_attn.kv_b_proj.weight); 12 x quantized:ct-mxfp4@4 (model.layers.N.self_attn.kv_a_proj_with_mqa.weight, model.layers.N.self_attn.q_a_proj.weight, model.layers.N.self_attn.q_b_proj.weight). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "attn.qkv",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": "all",
+        "note": "1 tensors: model.embed_tokens.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "embed_tokens",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": "all",
+        "note": "1 tensors: lm_head.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "lm_head",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 4,
+        "format": "ct-mxfp4",
+        "layer_range": "all",
+        "note": "1 tensors: model.layers.N.mlp.down_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "mlp.down",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": 4,
+        "format": "ct-mxfp4",
+        "layer_range": "all",
+        "note": "1 tensors: model.layers.N.mlp.gate_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "mlp.gate",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": 4,
+        "format": "ct-mxfp4",
+        "layer_range": "all",
+        "note": "1 tensors: model.layers.N.mlp.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "mlp.up",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": 4,
+        "format": "ct-mxfp4",
+        "layer_range": "all",
+        "note": "72 tensors: model.layers.N.mlp.experts.E.down_proj.weight, model.layers.N.mlp.experts.E.gate_proj.weight, model.layers.N.mlp.experts.E.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "moe.experts",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": null,
+        "format": "mixed",
+        "layer_range": "1-3",
+        "note": "class mixes formats on the same layers (SCOPE-004 admits one row per class and layer_range): 3 x native:bf16@16 (model.layers.N.mlp.gate.weight); 3 x native:fp32@32 (model.layers.N.mlp.gate.e_score_correction_bias). FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "moe.router",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 4,
+        "format": "ct-mxfp4",
+        "layer_range": "all",
+        "note": "9 tensors: model.layers.N.mlp.shared_experts.down_proj.weight, model.layers.N.mlp.shared_experts.gate_proj.weight, model.layers.N.mlp.shared_experts.up_proj.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "moe.shared_expert",
+        "treatment": "quantized"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": "all",
+        "note": "9 tensors: model.layers.N.input_layernorm.weight, model.layers.N.post_attention_layernorm.weight, model.norm.weight. FORMAT fixture CPU RTN; reconstructed weights only; activation-not-captured, no optimizer/native-kernel quality claim",
+        "tensor_class": "norm",
+        "treatment": "native"
+      }
+    ],
+    "head_policy": "native",
+    "kv_cache_dtype": "bf16",
+    "policy": "mixed",
+    "scope_digest": "attn.o=quantized:ct-mxfp4@4|attn.other=quantized:mixed|attn.qkv=quantized:mixed|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=quantized:ct-mxfp4@4|mlp.gate=quantized:ct-mxfp4@4|mlp.up=quantized:ct-mxfp4@4|moe.experts=quantized:ct-mxfp4@4|moe.router=native:mixed|moe.shared_expert=quantized:ct-mxfp4@4|norm=native:bf16@16|head=native|kv=bf16"
+  },
+  "seal": {
+    "checksums_file": "checksums.txt",
+    "checksums_format": "sha256sum",
+    "checksums_sha256": "838cf94a88ea50cab58a834076243558521828afa3df5f46fc20ef18c9d3c5e9",
+    "covers": [
+      "every published file except checksums.txt and fidelity-dataset.json, via checksums.txt",
+      "fidelity-dataset.json itself, via the self-blanked seal",
+      "every capture tensor a second time by CONTENT, via capture.capture_content_digest"
+    ],
+    "excludes": [
+      "checksums.txt",
+      "fidelity-dataset.json"
+    ],
+    "external_anchor": null,
+    "method": "self_blanked_canonical_json_sha256",
+    "seal_field": "dataset_sha256"
+  },
+  "upstream_receipts": [],
+  "weights": {
+    "artifact_ref": null,
+    "checkpoint_identity_sha256": "aa51ae526ba79294a678ada6189d0418fa78f1d5e51091e5d10696c56d3470ed",
+    "codec": "mxfp4",
+    "config_sha256": "c6cc76ca973555cae99d0c42da0a9385bf1ac67c67f4748cac0b4c0778ff5d90",
+    "declared_bits": 4.0,
+    "declared_head_bits": 16,
+    "index_sha256": null,
+    "model_ref": null,
+    "model_revision": null,
+    "quantized": true,
+    "repository": "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+    "revision": null
+  }
+}

raw/candidates/ct-mxfp4/evidence/candidate/head/head.json

@@ -0,0 +1,34 @@
+{
+  "applied_in_capture": false,
+  "bias": null,
+  "bits": 16,
+  "compat_tensor_key": "weight",
+  "dtype": "BF16",
+  "equality_receipt": null,
+  "file": "head/weight.safetensors",
+  "file_sha256": "07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913",
+  "final_norm": {
+    "applied_at_replay": false,
+    "applied_in_capture": true,
+    "dtype": null,
+    "file": null,
+    "file_sha256": null,
+    "shape": null,
+    "tensor_content_sha256": null,
+    "tensor_key": null
+  },
+  "format_version": 1,
+  "note": "the head is shipped verbatim from the checkpoint so a third party can replay logits = hidden @ head^T without the weights",
+  "present": true,
+  "quantized": false,
+  "raw_tensor_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+  "receipt_sha256": "2da3cfbfc607296b3bb32eb6fec1f122e4ce30fe9ea96653e5bb1b4ebab6514f",
+  "schema": "malaiwah.fidelity-head-identity.v1",
+  "shape": [
+    260,
+    64
+  ],
+  "source": "native",
+  "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+  "tensor_key": "lm_head.weight"
+}

raw/candidates/ct-mxfp4/evidence/candidate/head/weight.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913
+size 33365

raw/candidates/ct-mxfp4/evidence/candidate/panel/masks/context-0000.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/candidate/panel/masks/context-0001.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/candidate/panel/masks/context-0002.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/candidate/panel/masks/context-0003.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/candidate/panel/panel-receipt.json

@@ -0,0 +1,376 @@
+{
+  "corpus": {
+    "documents_considered": 4,
+    "documents_selected": 4,
+    "path_prefix": "embedded-synthetic/",
+    "repository": null,
+    "revision": null,
+    "source": "Synthetic texts embedded verbatim in this sealed receipt and builder."
+  },
+  "created_utc": "2026-09-07T00:00:00Z",
+  "created_utc_policy": "Fixed artifact epoch for deterministic bytes; not an execution timestamp.",
+  "documents": [
+    {
+      "bytes": 135,
+      "document": "garden.txt",
+      "document_sha256": "2aca14290cd4753c3b15d8a20aaa7a6963d2cb48d78e8538a4ce34d4665dcb94",
+      "num_tokens": 64,
+      "selected": true,
+      "special_tokens_mask": [
+        1,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        1
+      ],
+      "stratum": "garden",
+      "text_utf8": "The small garden has red flowers, green leaves, and a stone path. A gentle breeze moves the branches while a robin sings near the gate.",
+      "tokens_total": 137,
+      "window_id": "final-0000"
+    },
+    {
+      "bytes": 144,
+      "document": "arithmetic.txt",
+      "document_sha256": "b867a2e8cbe81cc8d5781bfe6718a43e171fb5c7c138b9ed6ac86bebc24bca83",
+      "num_tokens": 64,
+      "selected": true,
+      "special_tokens_mask": [
+        1,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        1
+      ],
+      "stratum": "arithmetic",
+      "text_utf8": "A shelf holds three blue boxes and five red boxes. There are eight boxes in total. Move two boxes to the desk and count the six boxes remaining.",
+      "tokens_total": 146,
+      "window_id": "final-0001"
+    },
+    {
+      "bytes": 162,
+      "document": "code.txt",
+      "document_sha256": "08c8f75f45899cbc9656a7a3fcdbf3092abfa7bc97fe353957b08d7e9660240b",
+      "num_tokens": 64,
+      "selected": true,
+      "special_tokens_mask": [
+        1,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        1
+      ],
+      "stratum": "code",
+      "text_utf8": "def add_values(left, right):\n    return left + right\n\nresult = add_values(7, 9)\nprint(result)\n# This readable snippet is synthetic input, not a coding benchmark.\n",
+      "tokens_total": 164,
+      "window_id": "final-0002"
+    },
+    {
+      "bytes": 140,
+      "document": "instructions.txt",
+      "document_sha256": "1dba5052b28929f94f322ad9e8c0b238bcbd653d6fae54ebaea8e5e3f6073378",
+      "num_tokens": 64,
+      "selected": true,
+      "special_tokens_mask": [
+        1,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        0,
+        1
+      ],
+      "stratum": "instructions",
+      "text_utf8": "First place a clean cup on the table. Next pour cool water into the cup. Finally put the bottle away and write a short note about the steps.",
+      "tokens_total": 142,
+      "window_id": "final-0003"
+    }
+  ],
+  "environment": {
+    "numpy": "2.5.3",
+    "tokenizers": "0.23.2",
+    "transformers": "5.16.1"
+  },
+  "format_builder": "engines/tools/build_token_panel.py",
+  "format_builder_sha256": "30eb4d4d7e465aaa8d9ef2d947a0ac7fd42ec45388f20c8b6db242c88dc2d851",
+  "format_version": 1,
+  "panel_id": "panel--glm-moe-dsa-tiny.malaiwah.cpu-smoke-v1",
+  "panel_name": "Tiny random-init GLM-MoE-DSA synthetic CPU smoke (not a benchmark)",
+  "parameters": {
+    "context_length": 64,
+    "generation_seed": null,
+    "generation_seed_policy": "No RNG is used.",
+    "prediction_positions_per_window": 63,
+    "scored_positions_total": 252,
+    "strata": [
+      "garden",
+      "arithmetic",
+      "code",
+      "instructions"
+    ],
+    "windows_per_stratum": 1,
+    "windows_total": 4
+  },
+  "receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5",
+  "resolver_sha256": "0bf78fca76289920e0dc10d58082f42f24f4f2db6e0ecde61daefa4c22de286d",
+  "schema": "malaiwah.token-panel-build-receipt.v1",
+  "selection_rule": "Four embedded readable synthetic documents in declared order; native tokenizer with add_special_tokens=True, truncation=True, max_length=64, padding=False; each encoded window must be exactly 64 tokens; no chat template; all 63 shifted next-token positions scored; no RNG.",
+  "separation": {
+    "checked": false,
+    "method": "No contamination scan or benchmark selection.",
+    "note": "Authored synthetic CPU smoke inputs for a random-init model; not a benchmark, held-out quality claim, or registry-row submission."
+  },
+  "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36",
+  "tokenizer": {
+    "add_special_tokens": true,
+    "chat_template_applied": false,
+    "class": "TokenizersBackend",
+    "files_sha256": {
+      "config.json": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7",
+      "generation_config.json": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34",
+      "tokenizer.json": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da",
+      "tokenizer_config.json": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+    },
+    "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+    "local_files_only": true,
+    "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+    "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+    "trust_remote_code": false,
+    "vocab_size": 260
+  },
+  "tool": "build_panel.py",
+  "tool_sha256": "f17b17262816676d6ededcc969756ac07ff991eab73117040ec41005f6ff63f3"
+}

raw/candidates/ct-mxfp4/evidence/candidate/panel/panel.json

@@ -0,0 +1,298 @@
+{
+  "contamination": {
+    "checked": false,
+    "hits": null,
+    "method": "not established",
+    "receipt": null
+  },
+  "context_length": 64,
+  "contexts": 4,
+  "format_version": 1,
+  "name": "token panel panel--panel",
+  "panel_id": "panel--panel",
+  "panel_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5",
+  "panel_token_sha256_legacy": "3c9467c85be43d238926f6d18bf825c2e4748814c95b8b0e8b8f1420546c3089",
+  "positions_per_context": 63,
+  "receipt_sha256": "8a4e27a74ca6ce39569b78c3d972077ab7966f6da118110b3b08adca1e284c6f",
+  "records": [
+    {
+      "allocation_stratum": null,
+      "attention_mask_file": "panel/masks/context-0000.npy",
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 0,
+      "document_id": "embedded-synthetic/garden.txt",
+      "domain": "garden",
+      "index": 0,
+      "num_tokens": 64,
+      "partition": null,
+      "prediction_positions": 63,
+      "role": "final",
+      "semantic_class": null,
+      "sentinel": false,
+      "source_cluster_id": null,
+      "token_file": "panel/tokens/context-0000.json",
+      "token_ids_first16": [
+        1,
+        55,
+        75,
+        72,
+        224,
+        86,
+        80,
+        68,
+        79,
+        79,
+        224,
+        74,
+        68,
+        85,
+        71,
+        72
+      ],
+      "token_ids_json_sha256": "08f8d8d449f7f679b24ada75dc4db5b0108508e16ccefb0f61cad6144593e315",
+      "token_ids_last16": [
+        224,
+        68,
+        81,
+        71,
+        224,
+        68,
+        224,
+        86,
+        87,
+        82,
+        81,
+        72,
+        224,
+        83,
+        68,
+        2
+      ],
+      "token_ids_sha256_legacy": "9d5378f0316be57e37f80d2bae7895ba4a1140c302c87ff8cce2fbd9afc83b77",
+      "window_id": "final-0000",
+      "window_index": 0
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_file": "panel/masks/context-0001.npy",
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 1,
+      "document_id": "embedded-synthetic/arithmetic.txt",
+      "domain": "arithmetic",
+      "index": 1,
+      "num_tokens": 64,
+      "partition": null,
+      "prediction_positions": 63,
+      "role": "final",
+      "semantic_class": null,
+      "sentinel": false,
+      "source_cluster_id": null,
+      "token_file": "panel/tokens/context-0001.json",
+      "token_ids_first16": [
+        1,
+        36,
+        224,
+        86,
+        75,
+        72,
+        79,
+        73,
+        224,
+        75,
+        82,
+        79,
+        71,
+        86,
+        224,
+        87
+      ],
+      "token_ids_json_sha256": "5bc88e4e47faf077b2fc85ae0758eee5308003ff0fefa5d6691a7f24f22ecc46",
+      "token_ids_last16": [
+        72,
+        86,
+        17,
+        224,
+        55,
+        75,
+        72,
+        85,
+        72,
+        224,
+        68,
+        85,
+        72,
+        224,
+        72,
+        2
+      ],
+      "token_ids_sha256_legacy": "e0f95114b71a46cb489539dc38909fb67fb378f25fafbb513e32690cc7f11b37",
+      "window_id": "final-0001",
+      "window_index": 1
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_file": "panel/masks/context-0002.npy",
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 2,
+      "document_id": "embedded-synthetic/code.txt",
+      "domain": "code",
+      "index": 2,
+      "num_tokens": 64,
+      "partition": null,
+      "prediction_positions": 63,
+      "role": "final",
+      "semantic_class": null,
+      "sentinel": false,
+      "source_cluster_id": null,
+      "token_file": "panel/tokens/context-0002.json",
+      "token_ids_first16": [
+        1,
+        71,
+        72,
+        73,
+        224,
+        68,
+        71,
+        71,
+        66,
+        89,
+        68,
+        79,
+        88,
+        72,
+        86,
+        11
+      ],
+      "token_ids_json_sha256": "51b956fbf40161ce10fe3da6c1e186368a567676467030535468064aea0d7d07",
+      "token_ids_last16": [
+        85,
+        76,
+        74,
+        75,
+        87,
+        202,
+        202,
+        85,
+        72,
+        86,
+        88,
+        79,
+        87,
+        224,
+        32,
+        2
+      ],
+      "token_ids_sha256_legacy": "482b890c7b0e2f7400257c43680093b7f2b90d6df5361c6de3ae97ddfdd2504e",
+      "window_id": "final-0002",
+      "window_index": 2
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_file": "panel/masks/context-0003.npy",
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 3,
+      "document_id": "embedded-synthetic/instructions.txt",
+      "domain": "instructions",
+      "index": 3,
+      "num_tokens": 64,
+      "partition": null,
+      "prediction_positions": 63,
+      "role": "final",
+      "semantic_class": null,
+      "sentinel": false,
+      "source_cluster_id": null,
+      "token_file": "panel/tokens/context-0003.json",
+      "token_ids_first16": [
+        1,
+        41,
+        76,
+        85,
+        86,
+        87,
+        224,
+        83,
+        79,
+        68,
+        70,
+        72,
+        224,
+        68,
+        224,
+        70
+      ],
+      "token_ids_json_sha256": "8653267d2a6dc67550759c29ccb80a7a5529d927e6c367c681f5018d55497a2d",
+      "token_ids_last16": [
+        224,
+        70,
+        82,
+        82,
+        79,
+        224,
+        90,
+        68,
+        87,
+        72,
+        85,
+        224,
+        76,
+        81,
+        87,
+        2
+      ],
+      "token_ids_sha256_legacy": "12e01edd3900ff2ca82621feb5504f24b003a53c0afe98a9ee0b490eb93141d4",
+      "window_id": "final-0003",
+      "window_index": 3
+    }
+  ],
+  "repository": null,
+  "revision": null,
+  "schema": "malaiwah.fidelity-panel-binding.v1",
+  "scored_positions_total": 252,
+  "scoring_window": {
+    "dropped_positions_total": 0,
+    "min_left_context_tokens": 1,
+    "policy": "every causal prediction position of every window is scored; nothing is dropped",
+    "score_from": 0,
+    "windowed": false
+  },
+  "strata": {},
+  "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36",
+  "token_digest_algorithm": {
+    "aggregate": "sha256('\\n'.join(per_record_hex).encode('ascii'))",
+    "legacy_aggregate": "sha256(''.join(per_record_hex).encode('utf-8'))",
+    "legacy_per_record": "sha256(json.dumps(ids).encode('utf-8'))",
+    "per_record": "sha256(json.dumps(ids, separators=(',',':')).encode('utf-8'))"
+  },
+  "tokenizer": {
+    "files": [
+      {
+        "bytes": 1479,
+        "name": "config.json",
+        "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+      },
+      {
+        "bytes": 216,
+        "name": "generation_config.json",
+        "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+      },
+      {
+        "bytes": 6763,
+        "name": "tokenizer.json",
+        "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+      },
+      {
+        "bytes": 236,
+        "name": "tokenizer_config.json",
+        "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+      }
+    ],
+    "files_verified": true,
+    "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+    "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+    "maximum_token_id_exclusive": 260,
+    "receipt": null,
+    "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+    "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+    "vocab_size": 260
+  }
+}

raw/candidates/ct-mxfp4/evidence/candidate/panel/tokens/context-0000.json

@@ -0,0 +1 @@
+[1,55,75,72,224,86,80,68,79,79,224,74,68,85,71,72,81,224,75,68,86,224,85,72,71,224,73,79,82,90,72,85,86,15,224,74,85,72,72,81,224,79,72,68,89,72,86,15,224,68,81,71,224,68,224,86,87,82,81,72,224,83,68,2]

raw/candidates/ct-mxfp4/evidence/candidate/panel/tokens/context-0001.json

@@ -0,0 +1 @@
+[1,36,224,86,75,72,79,73,224,75,82,79,71,86,224,87,75,85,72,72,224,69,79,88,72,224,69,82,91,72,86,224,68,81,71,224,73,76,89,72,224,85,72,71,224,69,82,91,72,86,17,224,55,75,72,85,72,224,68,85,72,224,72,2]

raw/candidates/ct-mxfp4/evidence/candidate/panel/tokens/context-0002.json

@@ -0,0 +1 @@
+[1,71,72,73,224,68,71,71,66,89,68,79,88,72,86,11,79,72,73,87,15,224,85,76,74,75,87,12,29,202,224,224,224,224,85,72,87,88,85,81,224,79,72,73,87,224,14,224,85,76,74,75,87,202,202,85,72,86,88,79,87,224,32,2]

raw/candidates/ct-mxfp4/evidence/candidate/panel/tokens/context-0003.json

@@ -0,0 +1 @@
+[1,41,76,85,86,87,224,83,79,68,70,72,224,68,224,70,79,72,68,81,224,70,88,83,224,82,81,224,87,75,72,224,87,68,69,79,72,17,224,49,72,91,87,224,83,82,88,85,224,70,82,82,79,224,90,68,87,72,85,224,76,81,87,2]

raw/candidates/ct-mxfp4/evidence/candidate/runtime/capture-runtime.json

@@ -0,0 +1,407 @@
+{
+  "capture_tool": {
+    "expert_fill": {
+      "bytes_filled": 294912,
+      "declined": {},
+      "decoded_slices": 72,
+      "identity": "a byte copy into the fused parameter's slice; bin/selftest_layer_outer.py L4/L6 assert equality with from_pretrained and with the converter path",
+      "layers_filled": 3,
+      "mode": "direct",
+      "reference": "transformers.core_model_loading.MergeModulelist(dim=0) + Concatenate(dim=1)",
+      "slices_filled": 72,
+      "staged_slices": 0,
+      "staging": "pageable",
+      "targets_filled": 6
+    },
+    "file": "engines/tools/hf_capture.py",
+    "head_module_path": "lm_head",
+    "head_resolution": "lm_head",
+    "layer_residency": "stream",
+    "mechanism": "transformers forward pass, layer-outer/window-inner schedule: for each decoder layer the model's own forward is run once per window with the layers below replaying their memoised output and the layers above suspending, so each layer's weights are materialised once for the whole panel. Windows are pushed through sequentially, never batched. Forward pre-hook on model.get_output_embeddings(), fired by the model's own epilogue after the last layer",
+    "model_view": null,
+    "parallel_plan_dropped": false,
+    "resolved_classes": {
+      "config": "transformers.models.glm_moe_dsa.configuration_glm_moe_dsa.GlmMoeDsaConfig",
+      "model": "transformers.models.glm_moe_dsa.modeling_glm_moe_dsa.GlmMoeDsaForCausalLM"
+    },
+    "resolved_panel_binding": {
+      "binding": {
+        "content": {
+          "archive": {
+            "algorithm": "sha256(ustar: sorted regular files; mode=0644; uid=gid=mtime=0)",
+            "bytes": 20480,
+            "compression": "none",
+            "format": "ustar",
+            "sha256": "389706dfa3335a75ec70a40d079995af1a60c05512334125dbd336790f12702f"
+          },
+          "manifest": [
+            {
+              "bytes": 192,
+              "path": "arrays/causal-mask-64.npy",
+              "sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02"
+            },
+            {
+              "bytes": 384,
+              "path": "arrays/final-0000.tokens.npy",
+              "sha256": "d60234f4b3b6044fb21eca18f7df82de9822f388e4d93d19d8a4eb69d6d490af"
+            },
+            {
+              "bytes": 384,
+              "path": "arrays/final-0001.tokens.npy",
+              "sha256": "20188213e4dfaaed04b1396a4ec46a5a2a6236c87d870c64d7102568d585ceea"
+            },
+            {
+              "bytes": 384,
+              "path": "arrays/final-0002.tokens.npy",
+              "sha256": "00341b323c286034a1c3a80ae66318cb6ce67153c08b1eedf6990b7a0310662c"
+            },
+            {
+              "bytes": 384,
+              "path": "arrays/final-0003.tokens.npy",
+              "sha256": "a13ec9b40a39027396ec752ee8595e0613569b78e27ef25d254f4f9a82d8775a"
+            },
+            {
+              "bytes": 4195,
+              "path": "panel.json",
+              "sha256": "fe67acc87ce549ce174fa521a775e9e41c8e1ccbc66155ee673a080fea5e24a6"
+            },
+            {
+              "bytes": 7808,
+              "path": "panel.receipt.json",
+              "sha256": "b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263"
+            }
+          ],
+          "manifest_sha256": "e5f2d6410a6290da1197016c5b57ae522a3570665215ddc23e2a0bd45f912181"
+        },
+        "panel": {
+          "bytes": 4195,
+          "context_length": 64,
+          "contexts": 4,
+          "file": "panel.json",
+          "id": "panel--glm-moe-dsa-tiny.malaiwah.cpu-smoke-v1",
+          "name": "Tiny random-init GLM-MoE-DSA synthetic CPU smoke (not a benchmark)",
+          "positions_per_context": 63,
+          "role": "final",
+          "scored_positions_total": 252,
+          "sha256": "fe67acc87ce549ce174fa521a775e9e41c8e1ccbc66155ee673a080fea5e24a6",
+          "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36"
+        },
+        "receipt": {
+          "bytes": 7808,
+          "declared_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5",
+          "file": "panel.receipt.json",
+          "receipt_file_sha256": "b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263",
+          "receipt_seal_mode": "self-blank"
+        },
+        "schema": "malaiwah.resolved-panel.v1",
+        "tokenizer": {
+          "files": [
+            {
+              "bytes": 1479,
+              "name": "config.json",
+              "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+            },
+            {
+              "bytes": 216,
+              "name": "generation_config.json",
+              "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+            },
+            {
+              "bytes": 6763,
+              "name": "tokenizer.json",
+              "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+            },
+            {
+              "bytes": 236,
+              "name": "tokenizer_config.json",
+              "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+            }
+          ],
+          "files_verified": true,
+          "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+          "maximum_token_id_exclusive": 260,
+          "receipt": null,
+          "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+          "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+          "vocab_size": 260
+        }
+      },
+      "binding_file": "panel-binding.json",
+      "binding_file_sha256": "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "tokenizer_equivalences": []
+    },
+    "schedule": "layer-outer",
+    "sha256": "44afcdac9f7c9cb26a35a2b7d9c1aed562344b76c8045a753651d2f2dbae99d0",
+    "unexpected_tensor_allowlist": null,
+    "verified_code": null,
+    "version": "hf_capture/1",
+    "weights_decode": {
+      "activation_quantization": null,
+      "checkpoint_bytes_read": 125392,
+      "comparison_class": "advisory",
+      "components_consumed": 212,
+      "components_not_consumed": [],
+      "components_planned": 212,
+      "formats_decoded": {
+        "mxfp4-pack-quantized": 106
+      },
+      "method": "microfloat-weight-reconstruction",
+      "modules_decoded": 106,
+      "modules_not_decoded": [],
+      "modules_planned": 106,
+      "output_dtype": "bfloat16",
+      "quantization_config": {
+        "config_groups": {
+          "group_0": {
+            "format": "mxfp4-pack-quantized",
+            "input_activations": null,
+            "output_activations": null,
+            "targets": [
+              "model.layers.0.mlp.down_proj",
+              "model.layers.0.mlp.gate_proj",
+              "model.layers.0.mlp.up_proj",
+              "model.layers.0.self_attn.indexer.weights_proj",
+              "model.layers.0.self_attn.indexer.wk",
+              "model.layers.0.self_attn.indexer.wq_b",
+              "model.layers.0.self_attn.kv_a_proj_with_mqa",
+              "model.layers.0.self_attn.o_proj",
+              "model.layers.0.self_attn.q_a_proj",
+              "model.layers.0.self_attn.q_b_proj",
+              "model.layers.1.mlp.experts.0.down_proj",
+              "model.layers.1.mlp.experts.0.gate_proj",
+              "model.layers.1.mlp.experts.0.up_proj",
+              "model.layers.1.mlp.experts.1.down_proj",
+              "model.layers.1.mlp.experts.1.gate_proj",
+              "model.layers.1.mlp.experts.1.up_proj",
+              "model.layers.1.mlp.experts.2.down_proj",
+              "model.layers.1.mlp.experts.2.gate_proj",
+              "model.layers.1.mlp.experts.2.up_proj",
+              "model.layers.1.mlp.experts.3.down_proj",
+              "model.layers.1.mlp.experts.3.gate_proj",
+              "model.layers.1.mlp.experts.3.up_proj",
+              "model.layers.1.mlp.experts.4.down_proj",
+              "model.layers.1.mlp.experts.4.gate_proj",
+              "model.layers.1.mlp.experts.4.up_proj",
+              "model.layers.1.mlp.experts.5.down_proj",
+              "model.layers.1.mlp.experts.5.gate_proj",
+              "model.layers.1.mlp.experts.5.up_proj",
+              "model.layers.1.mlp.experts.6.down_proj",
+              "model.layers.1.mlp.experts.6.gate_proj",
+              "model.layers.1.mlp.experts.6.up_proj",
+              "model.layers.1.mlp.experts.7.down_proj",
+              "model.layers.1.mlp.experts.7.gate_proj",
+              "model.layers.1.mlp.experts.7.up_proj",
+              "model.layers.1.mlp.shared_experts.down_proj",
+              "model.layers.1.mlp.shared_experts.gate_proj",
+              "model.layers.1.mlp.shared_experts.up_proj",
+              "model.layers.1.self_attn.kv_a_proj_with_mqa",
+              "model.layers.1.self_attn.o_proj",
+              "model.layers.1.self_attn.q_a_proj",
+              "model.layers.1.self_attn.q_b_proj",
+              "model.layers.2.mlp.experts.0.down_proj",
+              "model.layers.2.mlp.experts.0.gate_proj",
+              "model.layers.2.mlp.experts.0.up_proj",
+              "model.layers.2.mlp.experts.1.down_proj",
+              "model.layers.2.mlp.experts.1.gate_proj",
+              "model.layers.2.mlp.experts.1.up_proj",
+              "model.layers.2.mlp.experts.2.down_proj",
+              "model.layers.2.mlp.experts.2.gate_proj",
+              "model.layers.2.mlp.experts.2.up_proj",
+              "model.layers.2.mlp.experts.3.down_proj",
+              "model.layers.2.mlp.experts.3.gate_proj",
+              "model.layers.2.mlp.experts.3.up_proj",
+              "model.layers.2.mlp.experts.4.down_proj",
+              "model.layers.2.mlp.experts.4.gate_proj",
+              "model.layers.2.mlp.experts.4.up_proj",
+              "model.layers.2.mlp.experts.5.down_proj",
+              "model.layers.2.mlp.experts.5.gate_proj",
+              "model.layers.2.mlp.experts.5.up_proj",
+              "model.layers.2.mlp.experts.6.down_proj",
+              "model.layers.2.mlp.experts.6.gate_proj",
+              "model.layers.2.mlp.experts.6.up_proj",
+              "model.layers.2.mlp.experts.7.down_proj",
+              "model.layers.2.mlp.experts.7.gate_proj",
+              "model.layers.2.mlp.experts.7.up_proj",
+              "model.layers.2.mlp.shared_experts.down_proj",
+              "model.layers.2.mlp.shared_experts.gate_proj",
+              "model.layers.2.mlp.shared_experts.up_proj",
+              "model.layers.2.self_attn.indexer.weights_proj",
+              "model.layers.2.self_attn.indexer.wk",
+              "model.layers.2.self_attn.indexer.wq_b",
+              "model.layers.2.self_attn.kv_a_proj_with_mqa",
+              "model.layers.2.self_attn.o_proj",
+              "model.layers.2.self_attn.q_a_proj",
+              "model.layers.2.self_attn.q_b_proj",
+              "model.layers.3.mlp.experts.0.down_proj",
+              "model.layers.3.mlp.experts.0.gate_proj",
+              "model.layers.3.mlp.experts.0.up_proj",
+              "model.layers.3.mlp.experts.1.down_proj",
+              "model.layers.3.mlp.experts.1.gate_proj",
+              "model.layers.3.mlp.experts.1.up_proj",
+              "model.layers.3.mlp.experts.2.down_proj",
+              "model.layers.3.mlp.experts.2.gate_proj",
+              "model.layers.3.mlp.experts.2.up_proj",
+              "model.layers.3.mlp.experts.3.down_proj",
+              "model.layers.3.mlp.experts.3.gate_proj",
+              "model.layers.3.mlp.experts.3.up_proj",
+              "model.layers.3.mlp.experts.4.down_proj",
+              "model.layers.3.mlp.experts.4.gate_proj",
+              "model.layers.3.mlp.experts.4.up_proj",
+              "model.layers.3.mlp.experts.5.down_proj",
+              "model.layers.3.mlp.experts.5.gate_proj",
+              "model.layers.3.mlp.experts.5.up_proj",
+              "model.layers.3.mlp.experts.6.down_proj",
+              "model.layers.3.mlp.experts.6.gate_proj",
+              "model.layers.3.mlp.experts.6.up_proj",
+              "model.layers.3.mlp.experts.7.down_proj",
+              "model.layers.3.mlp.experts.7.gate_proj",
+              "model.layers.3.mlp.experts.7.up_proj",
+              "model.layers.3.mlp.shared_experts.down_proj",
+              "model.layers.3.mlp.shared_experts.gate_proj",
+              "model.layers.3.mlp.shared_experts.up_proj",
+              "model.layers.3.self_attn.kv_a_proj_with_mqa",
+              "model.layers.3.self_attn.o_proj",
+              "model.layers.3.self_attn.q_a_proj",
+              "model.layers.3.self_attn.q_b_proj"
+            ],
+            "weights": {
+              "dynamic": false,
+              "group_size": 32,
+              "num_bits": 4,
+              "scale_dtype": "torch.uint8",
+              "strategy": "group",
+              "symmetric": true,
+              "type": "float"
+            }
+          }
+        },
+        "format": "mxfp4-pack-quantized",
+        "ignore": [],
+        "quant_method": "compressed-tensors",
+        "quantization_status": "compressed"
+      },
+      "reader_evidence": {
+        "decode_plan_sha256": "2bb2eef8f5dea9e0662e3fce6fbb6ecf22842b18c7192484e78ea5a8850ce9fa",
+        "gate_header_barrier": false,
+        "source_files": {
+          "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+          "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+          "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+          "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+        }
+      },
+      "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers",
+      "scientific_disclosure": "Stored-weight reconstruction only; no native quantized GEMM, activation arithmetic, or quantizer optimization quality claim.",
+      "scope": "weights_reconstructed"
+    },
+    "weights_license": null,
+    "wraps": []
+  },
+  "container": {
+    "image_digest": null,
+    "image_reference": null,
+    "image_repository_digest": null
+  },
+  "format_version": 1,
+  "lane": "other",
+  "lane_identity_inputs": [
+    "torch_version",
+    "cuda_runtime_version",
+    "device_name",
+    "grouped_mm_kernel",
+    "numeric_policy",
+    "attention_backend",
+    "experts_implementation",
+    "parallelism",
+    "ep_emulate",
+    "reduce_order"
+  ],
+  "lane_identity_sha256": "34f640ff350d24093734d97a81c1048ce0443d8b3c7bda0953f8283b8ea9269f",
+  "lane_inferred": false,
+  "receipt_sha256": "58787c21c5e9a502b5a667beb8860eab3fdc68a3cc8c1325298a432359940c96",
+  "resources": {
+    "bytes": {
+      "checkpoint_read": 209072,
+      "hidden_d2h": 36616,
+      "weights_h2d": null
+    },
+    "checkpoint_bytes": 244397,
+    "checkpoint_files": 2,
+    "device_name": null,
+    "forward_timing": "wall-clock",
+    "note": "provenance, not identity: peaks and stage seconds this run measured; forward_sum is device time under cuda-events, wall time otherwise; seal covers head serialisation up to this block's assembly",
+    "peak_cuda_allocated_bytes": null,
+    "peak_cuda_reserved_bytes": null,
+    "peak_resident_weight_bytes": 208032,
+    "peak_rss_bytes": 368132096,
+    "rss_units_source": "Linux ru_maxrss",
+    "seconds": {
+      "decode_sum": 0.071,
+      "elapsed": 4.636,
+      "fill_sum": 0.0,
+      "forward_sum": 0.152,
+      "identity": 0.003,
+      "layer_load_max": 0.031,
+      "layer_load_sum": 0.114,
+      "layer_loads": 4,
+      "resident_load": 3.702,
+      "seal": 0.01
+    }
+  },
+  "runtime_environment": {
+    "cold_run": "20260907T165247Z-2063858",
+    "numeric_policy_observed": {
+      "NVIDIA_TF32_OVERRIDE": null,
+      "allow_tf32_cudnn": true,
+      "allow_tf32_matmul": false,
+      "deviates_from_default": false,
+      "float32_matmul_precision": "highest",
+      "observed_at": "seal time, after the forward"
+    },
+    "python": "3.12.14"
+  },
+  "schema": "malaiwah.fidelity-capture-runtime.v1",
+  "source_files": {
+    "bin/fidelity/codepin.py": "ce2a2c04d64adf7a268978a7ffe163c7d4cb5b675d5dc8e156f8a26a3b7bb5fd",
+    "bin/fidelity/panel.py": "0bf78fca76289920e0dc10d58082f42f24f4f2db6e0ecde61daefa4c22de286d",
+    "engines/tools/affine_surface.py": "37e7074797d53ae836e0af12fdbd2dc36c88778098aa44c72c1cd936d2b4c053",
+    "engines/tools/hf_capture.py": "44afcdac9f7c9cb26a35a2b7d9c1aed562344b76c8045a753651d2f2dbae99d0",
+    "engines/tools/layer_outer.py": "e3ea5715112663b1b60cf596eeffc2140e49b399809ece89a174bce41a969274",
+    "engines/tools/microscale_surface.py": "588923f0a4700436221923c794a8c112971d9f08e75c746742b0f0ede7921682",
+    "engines/tools/nvfp4_surface.py": "044b6fd2ee22d58512476e96060f7fa09030212a33736da1393c0103eb73e0c8",
+    "engines/tools/quant_stream.py": "7c6cc4cbe399d066eaff80ccf1e3d3190185f194644f738a404d3c537ad8e1ab"
+  },
+  "stack_fingerprint": {
+    "attention_backend": "model default",
+    "cuda_runtime_version": null,
+    "device": "cpu",
+    "device_name": null,
+    "engine": "transformers-eager",
+    "numeric_policy": "default torch matmul precision; no TF32 override applied",
+    "schema": "malaiwah.stack-fingerprint.v1",
+    "torch_version": "2.11.0+cpu",
+    "transformers_version": "5.16.1"
+  },
+  "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd",
+  "upstream_receipts": [],
+  "weights": {
+    "checkpoint_files": [
+      {
+        "name": "config.json",
+        "sha256": "c6cc76ca973555cae99d0c42da0a9385bf1ac67c67f4748cac0b4c0778ff5d90",
+        "size": 7517
+      },
+      {
+        "name": "model.safetensors",
+        "sha256": "9222707b17727c48479eb37416d96289344b48d70eee2d919b282063ccee3b3d",
+        "size": 236880
+      }
+    ],
+    "checkpoint_identity_algorithm": "sha256 over the canonical JSON {\"algorithm\": <this string>, \"files\": [{\"name\", \"size\", \"sha256\"}, ...]} of every *.safetensors shard plus config.json in the checkpoint directory, sorted by name; for a llama.cpp GGUF tree the shards are every *.gguf of the one build directory (name = <build>/<file>) and config.json is the official release's copy beside it",
+    "checkpoint_identity_sha256": "aa51ae526ba79294a678ada6189d0418fa78f1d5e51091e5d10696c56d3470ed",
+    "model_revision": null,
+    "repository": "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+    "revision": null
+  }
+}

raw/candidates/ct-mxfp4/evidence/candidate/validation/structural-validation.json

@@ -0,0 +1,14 @@
+{
+  "checks_run": [
+    "pre-seal"
+  ],
+  "error_count": 0,
+  "errors": [],
+  "format_version": 1,
+  "receipt_sha256": "df49786fa6ab7b99b4ccb964492beb1dfda015c837b9483c9f730bf69e82513c",
+  "schema": "malaiwah.fidelity-structural-validation.v1",
+  "structural_status": "sealed",
+  "subject": "dataset:fidelity--quant-integration.candidate",
+  "warning_count": 0,
+  "warnings": []
+}

raw/candidates/ct-mxfp4/evidence/capture-candidate.stderr.log

@@ -0,0 +1,15 @@
+
+Loading weights:   0%|          | 0/6 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 6/6 [00:00<00:00, 7469.82it/s]
+
+Loading weights:   0%|          | 0/17 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 17/17 [00:00<00:00, 5676.55it/s]
+
+Loading weights:   0%|          | 0/13 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 13/13 [00:00<00:00, 6210.95it/s]
+
+Loading weights:   0%|          | 0/18 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 18/18 [00:00<00:00, 5874.38it/s]
+
+Loading weights:   0%|          | 0/13 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 13/13 [00:00<00:00, 4924.23it/s]

raw/candidates/ct-mxfp4/evidence/capture-candidate.stdout.log

@@ -0,0 +1,37 @@
+{"files": 2, "seconds": 0.003, "sha256": "aa51ae526ba79294a678ada6189d0418fa78f1d5e51091e5d10696c56d3470ed", "stage": "checkpoint_identity"}
+{"bytes": 236880, "index_present": true, "shards": 1, "stage": "checkpoint_audit", "tensors": 245}
+{"components": 212, "method": "microfloat-weight-reconstruction", "modules": 106, "reference": "compressed-tensors@0.18.0; DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1/inference/kernel.py; nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47/config.json+headers", "stage": "packed_decode_plan"}
+{"layers_prefix": "model.layers", "renamed_checkpoint_keys": 72, "stage": "stream_routing", "total_checkpoint_keys": 139}
+{"deferred_buffers": 0, "layers_prefix": "model.layers", "resident_shards": null, "resident_tensors": 6, "stage": "stream_base", "streamed_layers": 4, "streamed_params": 67}
+{"conversion_errors": 0, "conversion_errors_visible": true, "error_msgs": 0, "mismatched": 0, "missing": 0, "observed": true, "stage": "load_report", "unexpected": 0, "unexpected_sample": []}
+{"architectures": ["GlmMoeDsaForCausalLM"], "device": "cpu", "dtype": "torch.bfloat16", "hidden_size": 64, "stage": "model", "vocab_size": 260}
+{"panel_json": "/tmp/qfs-complete-tiny-fixture/panel/panel.json", "receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5", "stage": "panel", "windows": 4}
+{"expect": null, "prompt": "The capital of France is", "prompt_tokens": 26, "stage": "sanity_probe_planned"}
+{"forward_windows": 5, "layers": 4, "residency": "stream", "schedule": "layer-outer", "stage": "schedule", "windows": 4}
+{"decode_seconds": 0.013, "decoded_modules": 10, "dequantized": 0, "fp8_elements": 0, "index": 0, "stage": "packed_decode_layer"}
+{"checkpoint_tensors": 17, "index": 0, "resident_weight_bytes": 145568, "seconds": 0.024, "stage": "layer_load"}
+{"index": 0, "stage": "layer", "windows": 5}
+{"decode_seconds": 0.018, "decoded_modules": 31, "dequantized": 0, "fp8_elements": 0, "index": 1, "stage": "packed_decode_layer"}
+{"checkpoint_tensors": 37, "index": 1, "resident_weight_bytes": 201312, "seconds": 0.028, "stage": "layer_load"}
+{"index": 1, "stage": "layer", "windows": 5}
+{"decode_seconds": 0.02, "decoded_modules": 34, "dequantized": 0, "fp8_elements": 0, "index": 2, "stage": "packed_decode_layer"}
+{"checkpoint_tensors": 42, "index": 2, "resident_weight_bytes": 208032, "seconds": 0.031, "stage": "layer_load"}
+{"index": 2, "stage": "layer", "windows": 5}
+{"decode_seconds": 0.019, "decoded_modules": 31, "dequantized": 0, "fp8_elements": 0, "index": 3, "stage": "packed_decode_layer"}
+{"checkpoint_tensors": 37, "index": 3, "resident_weight_bytes": 201312, "seconds": 0.03, "stage": "layer_load"}
+{"index": 3, "stage": "layer", "windows": 5}
+{"enforced": false, "entropy_nats": 5.5481, "probability": 0.006773, "stage": "sanity_probe", "status": "recorded", "top1": "\ufffd", "uniform_entropy_nats": 5.5607}
+{"stage": "load_report_final", "unexpected": 0, "unexpected_sample": []}
+{"bytes": 8337, "index": 0, "rows": 63, "seconds": 0.038, "stage": "window", "window_id": "final-0000"}
+progress: windows 1/4 25% [00:00<00:00, 563.74 it/s]
+{"bytes": 8337, "index": 1, "rows": 63, "seconds": 0.031, "stage": "window", "window_id": "final-0001"}
+progress: windows 2/4 50% [00:00<00:00, 650.29 it/s]
+{"bytes": 8337, "index": 2, "rows": 63, "seconds": 0.03, "stage": "window", "window_id": "final-0002"}
+progress: windows 3/4 75% [00:00<00:00, 664.57 it/s]
+{"bytes": 8337, "index": 3, "rows": 63, "seconds": 0.031, "stage": "window", "window_id": "final-0003"}
+progress: windows 4/4 100% [00:00<00:00, 740.24 it/s]
+progress: windows 4/4 100% [00:00<00:00, 737.14 it/s] capture complete
+{"bytes": 33365, "file": "head/weight.safetensors", "stage": "head", "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510"}
+{"context_length": 64, "decoder_layers": 4, "device": "cpu", "hidden_size": 64, "layer_residency": "stream", "model": "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4", "peak_resident_weight_bytes": 208032, "peak_resident_weight_detail": {"buffers": 128, "parameters": 207904, "total": 208032}, "peak_resident_weight_gb": 0.0, "peak_rss_bytes": 368132096, "peak_rss_gb": 0.368, "resident_weight_note": "the maximum, over the run, of the materialised parameter+buffer bytes. On the CPU path peak_rss_bytes ALSO counts the safetensors mmap page cache, which is evictable and is not held by the schedule; this figure is not confounded by it. On CUDA, peak_cuda_allocated_bytes is the authoritative number because it includes activations and workspace and has no page cache.", "rss_units_source": "Linux ru_maxrss", "schedule": "layer-outer", "stage": "peak_memory", "vocab_size": 260, "windows": 4}
+{"bytes": 7808, "declared_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5", "file": "panel/panel-receipt.json", "receipt_file_sha256": "b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263", "stage": "panel_receipt"}
+{"capture_content_digest": "95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876", "dataset_sha256": "7bbd1d3e4f4d78f1a154187e48b138acfbb557c5e9d740e686036101501dc761", "elapsed_seconds": 4.7, "hidden_form_bytes": 33348, "logit_form_bytes_equivalent": 262080, "out": "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate", "records": 4, "scored_rows": 252, "stage": "sealed", "storage_ratio": 7.9}

raw/candidates/ct-mxfp4/evidence/capture-root.stderr.log

@@ -0,0 +1,15 @@
+
+Loading weights:   0%|          | 0/6 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 6/6 [00:00<00:00, 10042.23it/s]
+
+Loading weights:   0%|          | 0/17 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 17/17 [00:00<00:00, 4844.95it/s]
+
+Loading weights:   0%|          | 0/13 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 13/13 [00:00<00:00, 4630.65it/s]
+
+Loading weights:   0%|          | 0/18 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 18/18 [00:00<00:00, 4035.36it/s]
+
+Loading weights:   0%|          | 0/13 [00:00<?, ?it/s]
+Loading weights: 100%|██████████| 13/13 [00:00<00:00, 4974.09it/s]

raw/candidates/ct-mxfp4/evidence/capture-root.stdout.log

@@ -0,0 +1,32 @@
+{"files": 2, "seconds": 0.005, "sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e", "stage": "checkpoint_identity"}
+{"bytes": 571000, "index_present": false, "shards": 1, "stage": "checkpoint_audit", "tensors": 139}
+{"layers_prefix": "model.layers", "renamed_checkpoint_keys": 72, "stage": "stream_routing", "total_checkpoint_keys": 139}
+{"deferred_buffers": 0, "layers_prefix": "model.layers", "resident_shards": null, "resident_tensors": 6, "stage": "stream_base", "streamed_layers": 4, "streamed_params": 67}
+{"conversion_errors": 0, "conversion_errors_visible": true, "error_msgs": 0, "mismatched": 0, "missing": 0, "observed": true, "stage": "load_report", "unexpected": 0, "unexpected_sample": []}
+{"architectures": ["GlmMoeDsaForCausalLM"], "device": "cpu", "dtype": "torch.bfloat16", "hidden_size": 64, "stage": "model", "vocab_size": 260}
+{"panel_json": "/tmp/qfs-complete-tiny-fixture/panel/panel.json", "receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5", "stage": "panel", "windows": 4}
+{"expect": null, "prompt": "The capital of France is", "prompt_tokens": 26, "stage": "sanity_probe_planned"}
+{"forward_windows": 5, "layers": 4, "residency": "stream", "schedule": "layer-outer", "stage": "schedule", "windows": 4}
+{"checkpoint_tensors": 17, "index": 0, "resident_weight_bytes": 145568, "seconds": 0.013, "stage": "layer_load"}
+{"index": 0, "stage": "layer", "windows": 5}
+{"checkpoint_tensors": 37, "index": 1, "resident_weight_bytes": 201312, "seconds": 0.017, "stage": "layer_load"}
+{"index": 1, "stage": "layer", "windows": 5}
+{"checkpoint_tensors": 42, "index": 2, "resident_weight_bytes": 208032, "seconds": 0.016, "stage": "layer_load"}
+{"index": 2, "stage": "layer", "windows": 5}
+{"checkpoint_tensors": 37, "index": 3, "resident_weight_bytes": 201312, "seconds": 0.012, "stage": "layer_load"}
+{"index": 3, "stage": "layer", "windows": 5}
+{"enforced": false, "entropy_nats": 5.548, "probability": 0.00672, "stage": "sanity_probe", "status": "recorded", "top1": "\ufffd", "uniform_entropy_nats": 5.5607}
+{"stage": "load_report_final", "unexpected": 0, "unexpected_sample": []}
+{"bytes": 8337, "index": 0, "rows": 63, "seconds": 0.034, "stage": "window", "window_id": "final-0000"}
+progress: windows 1/4 25% [00:00<00:00, 786.48 it/s]
+{"bytes": 8337, "index": 1, "rows": 63, "seconds": 0.027, "stage": "window", "window_id": "final-0001"}
+progress: windows 2/4 50% [00:00<00:00, 971.12 it/s]
+{"bytes": 8337, "index": 2, "rows": 63, "seconds": 0.028, "stage": "window", "window_id": "final-0002"}
+progress: windows 3/4 75% [00:00<00:00, 1080.40 it/s]
+{"bytes": 8337, "index": 3, "rows": 63, "seconds": 0.027, "stage": "window", "window_id": "final-0003"}
+progress: windows 4/4 100% [00:00<00:00, 1146.22 it/s]
+progress: windows 4/4 100% [00:00<00:00, 1140.48 it/s] capture complete
+{"bytes": 33365, "file": "head/weight.safetensors", "stage": "head", "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510"}
+{"context_length": 64, "decoder_layers": 4, "device": "cpu", "hidden_size": 64, "layer_residency": "stream", "model": "/tmp/qfs-complete-tiny-fixture/model", "peak_resident_weight_bytes": 208032, "peak_resident_weight_detail": {"buffers": 128, "parameters": 207904, "total": 208032}, "peak_resident_weight_gb": 0.0, "peak_rss_bytes": 367259648, "peak_rss_gb": 0.367, "resident_weight_note": "the maximum, over the run, of the materialised parameter+buffer bytes. On the CPU path peak_rss_bytes ALSO counts the safetensors mmap page cache, which is evictable and is not held by the schedule; this figure is not confounded by it. On CUDA, peak_cuda_allocated_bytes is the authoritative number because it includes activations and workspace and has no page cache.", "rss_units_source": "Linux ru_maxrss", "schedule": "layer-outer", "stage": "peak_memory", "vocab_size": 260, "windows": 4}
+{"bytes": 7808, "declared_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5", "file": "panel/panel-receipt.json", "receipt_file_sha256": "b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263", "stage": "panel_receipt"}
+{"capture_content_digest": "d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39", "dataset_sha256": "4cdf53f91210943b5c87bcddcfbee61034ed056af87d1b174cec28d7a2405ae0", "elapsed_seconds": 4.4, "hidden_form_bytes": 33348, "logit_form_bytes_equivalent": 262080, "out": "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root", "records": 4, "scored_rows": 252, "stage": "sealed", "storage_ratio": 7.9}

raw/candidates/ct-mxfp4/evidence/commands.json

@@ -0,0 +1,119 @@
+[
+  {
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root",
+      "--dataset-id",
+      "fidelity--quant-integration.root",
+      "--dataset-name",
+      "Local quant integration root",
+      "--role",
+      "root"
+    ],
+    "returncode": 0
+  },
+  {
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/engines/tools/hf_capture.py",
+      "--panel",
+      "/tmp/qfs-complete-tiny-fixture/panel",
+      "--panel-role",
+      "final",
+      "--panel-binding",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/panel-binding.json",
+      "--panel-binding-sha256",
+      "333b063b8f4cac60c9ee62a8a4cc5913dbdb3c7e3055680dcc3fa7bc5c208d69",
+      "--panel-tokenizer-root",
+      "/tmp/qfs-complete-tiny-fixture/model",
+      "--lane",
+      "other",
+      "--device",
+      "cpu",
+      "--dtype",
+      "bfloat16",
+      "--author",
+      "malaiwah",
+      "--schedule",
+      "layer-outer",
+      "--layer-residency",
+      "stream",
+      "--disclosures",
+      "[{\"code\": \"rtn_format_fixture\", \"severity\": \"info\", \"affects_comparability\": false, \"detail\": \"Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.\"}]",
+      "--model",
+      "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate",
+      "--dataset-id",
+      "fidelity--quant-integration.candidate",
+      "--dataset-name",
+      "Local quant integration candidate",
+      "--role",
+      "quant",
+      "--codec",
+      "mxfp4",
+      "--declared-bits",
+      "4.0",
+      "--scope-file",
+      "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4/scope.json"
+    ],
+    "returncode": 0
+  },
+  {
+    "argv": [
+      "/tmp/qfs-quant-venv/bin/python",
+      "-c",
+      "import runpy,sys,torch; torch.set_num_threads(2); torch.set_num_interop_threads(2); p=sys.argv.pop(1); sys.path.insert(0,str(__import__('pathlib').Path(p).parent)); sys.argv[0]=p; runpy.run_path(p,run_name='__main__')",
+      "/tmp/qfs-explorer-development/bin/fidelity_dataset.py",
+      "compare",
+      "--reference",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/root",
+      "--candidate",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/candidate",
+      "--out",
+      "/tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/comparison",
+      "--force-compute",
+      "--device",
+      "cpu",
+      "--replay-device",
+      "numpy",
+      "--replay-dtype",
+      "float32",
+      "--verify-tensors",
+      "--own-heads"
+    ],
+    "returncode": 2
+  }
+]

raw/candidates/ct-mxfp4/evidence/compare.stdout.log

@@ -0,0 +1,13 @@
+MEASUREMENT
+  metric              mean_tokenwise_kld = 0.000866341033772075 nats
+  direction           KL(reference || candidate)
+  top-1 agreement     0.5873015873015873
+  kl                  {"mean": 0.000866341033772075, "median": 0.0006639956888964968, "p95": 0.0021510926922378047, "p99": 0.0038031541906571086, "p99_9": 0.005432225354176228, "max": 0.005562118890622922}
+  scored positions    252 over 4 contexts
+  estimator           full vocabulary, float64, head_policy=native_head, stack=same_stack
+  backend             torch:k6_kld_report._token_kld (replay numpy:cpu:float32)
+  comparability       class=advisory same_lane=True usable_as_floor=True
+  disclosure          native_head_replay (info)
+  disclosure          weights_reconstructed (caveat)
+  receipt             /tmp/qfs-quant-microfloat-release/proof-public/ct-mxfp4/comparison/comparison-receipt.json
+  tokenwise           tokenwise-kld.npy (2144 bytes, 9970e64072dd76ae)

raw/candidates/ct-mxfp4/evidence/comparison/comparison-receipt.json

@@ -0,0 +1,473 @@
+{
+  "candidate": {
+    "capture_content_digest": "95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876",
+    "dataset_id": "fidelity--quant-integration.candidate",
+    "dataset_sha256": "7bbd1d3e4f4d78f1a154187e48b138acfbb557c5e9d740e686036101501dc761",
+    "form": "hidden",
+    "head": {
+      "bits": 16,
+      "quantized": false,
+      "raw_tensor_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+      "source": "native",
+      "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510"
+    },
+    "label": "candidate",
+    "lane": "other",
+    "lane_identity_sha256": "34f640ff350d24093734d97a81c1048ce0443d8b3c7bda0953f8283b8ea9269f",
+    "repository": null,
+    "revision": null,
+    "role": "quant",
+    "scope_digest": "attn.o=quantized:ct-mxfp4@4|attn.other=quantized:mixed|attn.qkv=quantized:mixed|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=quantized:ct-mxfp4@4|mlp.gate=quantized:ct-mxfp4@4|mlp.up=quantized:ct-mxfp4@4|moe.experts=quantized:ct-mxfp4@4|moe.router=native:mixed|moe.shared_expert=quantized:ct-mxfp4@4|norm=native:bf16@16|head=native|kv=bf16",
+    "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd",
+    "weights": {
+      "artifact_ref": null,
+      "checkpoint_identity_sha256": "aa51ae526ba79294a678ada6189d0418fa78f1d5e51091e5d10696c56d3470ed",
+      "codec": "mxfp4",
+      "config_sha256": "c6cc76ca973555cae99d0c42da0a9385bf1ac67c67f4748cac0b4c0778ff5d90",
+      "declared_bits": 4.0,
+      "declared_head_bits": 16,
+      "index_sha256": null,
+      "model_ref": null,
+      "model_revision": null,
+      "quantized": true,
+      "repository": "/tmp/qfs-quant-microfloat-release/stage-main/ct-mxfp4",
+      "revision": null
+    }
+  },
+  "comparability": {
+    "bias": null,
+    "class": "advisory",
+    "key": "cmp--38569dc9bf229373",
+    "key_inputs": {
+      "accumulation_dtype": "float64",
+      "direction": "reference_to_candidate",
+      "head_policy": "native_head",
+      "metric_name": "mean_tokenwise_kld",
+      "note": "provisional; the registry recomputes with its reference record id (registry CMP-001)",
+      "panel_id": "panel--panel",
+      "reference_id": "fidelity--quant-integration.root",
+      "stack_relation": "same_stack"
+    },
+    "same_lane": true,
+    "usable_as_floor": true
+  },
+  "comparator": {
+    "accumulation_dtype": "float64",
+    "bf16_reduced_precision_reduction": false,
+    "cublas_workspace_config": null,
+    "deterministic_algorithms": true,
+    "device": "cpu",
+    "estimator_backend": "torch:k6_kld_report._token_kld",
+    "head_applied_candidate_tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+    "head_applied_reference_tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+    "head_applied_tensor_content_sha256": null,
+    "logprob_dtype": "float64",
+    "position_block": 128,
+    "replay_backend": "numpy:cpu:float32",
+    "replay_env": {
+      "blas_configuration": "OpenBLAS 0.3.34.106.0  USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell MAX_THREADS=64",
+      "blas_name": "scipy-openblas",
+      "blas_threads": 16,
+      "blas_threads_source": "ctypes:scipy_openblas_get_num_threads64_",
+      "blas_version": "0.3.34.106.0",
+      "cpu_model": "Intel(R) Xeon(R) CPU           X5570  @ 2.93GHz",
+      "library": "numpy",
+      "numpy_version": "2.5.3",
+      "replay_dtype": "float32"
+    },
+    "replay_peak_device_bytes": null,
+    "short_circuited": false,
+    "source_file_hashes_verified": true,
+    "tensor_content_digests_verified": true,
+    "tf32": false,
+    "tool": {
+      "entrypoint": "bin/fidelity-dataset compare",
+      "note": "the fp64 estimator is k6_kld_report._token_kld, imported not copied, whenever torch is importable; estimator_backend records which path ran."
+    },
+    "two_pass": true,
+    "vocab_chunk": null
+  },
+  "comparison_kind": "measurement",
+  "context_depth_buckets": {
+    "0000-0255": 0.0008663410337720752
+  },
+  "created_utc": "2026-09-07T16:52:55Z",
+  "determinism": {
+    "cold_start_per_run": null,
+    "distinct_evidence_hash_count": 2,
+    "evidence_hashes": [
+      "95413ca78defc45614f623378c6cbebd919f1c429546e0c92997030a812f2876",
+      "d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39"
+    ],
+    "evidence_kind": "hidden_state_tensor_sha256",
+    "identical_across_runs": null,
+    "max_run_mean": 0.000866341033772075,
+    "min_run_mean": 0.000866341033772075,
+    "note": "one comparison of two sealed captures. identical_across_runs is null, not true: that claim needs two independently produced captures of the SAME weights (SC-2), which is a different receipt.",
+    "population_stddev_of_run_means": 0.0,
+    "run_count": 1,
+    "run_means": [
+      0.000866341033772075
+    ]
+  },
+  "disclosures": [
+    {
+      "affects_comparability": false,
+      "code": "native_head_replay",
+      "detail": "HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.",
+      "severity": "info"
+    },
+    {
+      "affects_comparability": true,
+      "code": "weights_reconstructed",
+      "detail": "candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (microfloat-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.",
+      "severity": "caveat"
+    }
+  ],
+  "estimator": {
+    "accumulation_dtype": "float64",
+    "head_policy": "native_head",
+    "hidden_dtype": "bf16",
+    "logits_dtype": "float32",
+    "softmax_note": "full vocabulary, torch.log_softmax in float64; no truncation, no top-k, no clamp. A non-finite intermediate is a hard refusal.",
+    "stack_relation": "same_stack",
+    "two_pass": true,
+    "vocab_chunk": null,
+    "zero_handling": null
+  },
+  "format_version": 1,
+  "gates": {
+    "coverage": {
+      "detail": "4 of 4/4 declared records shared",
+      "overridden_by": null,
+      "passed": true
+    },
+    "decode": {
+      "detail": "weights_decode reference none; candidate microfloat-weight-reconstruction",
+      "overridden_by": null,
+      "passed": true
+    },
+    "form": {
+      "detail": "hidden vs hidden",
+      "overridden_by": null,
+      "passed": true
+    },
+    "geometry": {
+      "detail": "vocab_size 260 equal; hidden_width 64",
+      "overridden_by": null,
+      "passed": true
+    },
+    "head": {
+      "detail": "HEAD-1d: own heads on both sides (2364995d4009 vs 2364995d4009)",
+      "overridden_by": null,
+      "passed": true
+    },
+    "lane": {
+      "detail": "lane=other vs other",
+      "overridden_by": null,
+      "passed": true
+    },
+    "lossy": {
+      "detail": "lossy_codec null on both sides",
+      "overridden_by": null,
+      "passed": true
+    },
+    "panel": {
+      "detail": "suite_token_hash_sha256 equal; 4 shared records; per-record token and mask digests equal; scoring_window equal",
+      "overridden_by": null,
+      "passed": true
+    },
+    "seal": {
+      "detail": "both manifests self-seal and checksums.txt verified; every tensor_content_sha256 recomputed",
+      "overridden_by": null,
+      "passed": true
+    },
+    "stack": {
+      "detail": "lane_identity equal / stack_fingerprint equal -> same_stack",
+      "overridden_by": null,
+      "passed": true
+    }
+  },
+  "high_kld_contexts": [
+    {
+      "allocation_stratum": null,
+      "domain": "code",
+      "index": 2,
+      "max": 0.005562118890622922,
+      "mean": 0.0009589654666779309,
+      "scored_rows": 63,
+      "window_id": "final-0002"
+    },
+    {
+      "allocation_stratum": null,
+      "domain": "arithmetic",
+      "index": 1,
+      "max": 0.005044614761353234,
+      "mean": 0.0008948578801078056,
+      "scored_rows": 63,
+      "window_id": "final-0001"
+    },
+    {
+      "allocation_stratum": null,
+      "domain": "instructions",
+      "index": 3,
+      "max": 0.002863526238924269,
+      "mean": 0.0008777956272695692,
+      "scored_rows": 63,
+      "window_id": "final-0003"
+    },
+    {
+      "allocation_stratum": null,
+      "domain": "garden",
+      "index": 0,
+      "max": 0.0019943248685364476,
+      "mean": 0.0007337451610329949,
+      "scored_rows": 63,
+      "window_id": "final-0000"
+    }
+  ],
+  "kl": {
+    "max": 0.005562118890622922,
+    "mean": 0.000866341033772075,
+    "median": 0.0006639956888964968,
+    "p95": 0.0021510926922378047,
+    "p99": 0.0038031541906571086,
+    "p99_9": 0.005432225354176228
+  },
+  "kl_macro_stratum_mean": null,
+  "kl_micro_token_mean": 0.000866341033772075,
+  "measurement_scope": {
+    "contexts": 4,
+    "covers_full_panel": true,
+    "position_filter": "all",
+    "positions_per_context": 63,
+    "scored_positions": 252,
+    "subset_detail": null
+  },
+  "metric": {
+    "direction": "reference_to_candidate",
+    "direction_label": "KL(reference || candidate)",
+    "higher_is_better": false,
+    "name": "mean_tokenwise_kld",
+    "units": "nats",
+    "value": 0.000866341033772075
+  },
+  "panel": {
+    "context_length": 64,
+    "contexts": 4,
+    "panel_id": "panel--panel",
+    "panel_token_sha256_legacy": "3c9467c85be43d238926f6d18bf825c2e4748814c95b8b0e8b8f1420546c3089",
+    "scored_positions": 252,
+    "scoring_window": {
+      "dropped_positions_total": 0,
+      "min_left_context_tokens": 1,
+      "policy": "every causal prediction position of every window is scored; nothing is dropped",
+      "score_from": 0,
+      "windowed": false
+    },
+    "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36",
+    "tokenizer": {
+      "files": [
+        {
+          "bytes": 1479,
+          "name": "config.json",
+          "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+        },
+        {
+          "bytes": 216,
+          "name": "generation_config.json",
+          "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+        },
+        {
+          "bytes": 6763,
+          "name": "tokenizer.json",
+          "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+        },
+        {
+          "bytes": 236,
+          "name": "tokenizer_config.json",
+          "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+        }
+      ],
+      "files_verified": true,
+      "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+      "maximum_token_id_exclusive": 260,
+      "receipt": null,
+      "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+      "vocab_size": 260
+    },
+    "tokenizer_candidate": {
+      "files": [
+        {
+          "bytes": 1479,
+          "name": "config.json",
+          "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+        },
+        {
+          "bytes": 216,
+          "name": "generation_config.json",
+          "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+        },
+        {
+          "bytes": 6763,
+          "name": "tokenizer.json",
+          "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+        },
+        {
+          "bytes": 236,
+          "name": "tokenizer_config.json",
+          "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+        }
+      ],
+      "files_verified": true,
+      "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+      "maximum_token_id_exclusive": 260,
+      "receipt": null,
+      "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+      "vocab_size": 260
+    },
+    "tokenizer_identity_equal": true,
+    "tokenizer_reference": {
+      "files": [
+        {
+          "bytes": 1479,
+          "name": "config.json",
+          "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+        },
+        {
+          "bytes": 216,
+          "name": "generation_config.json",
+          "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+        },
+        {
+          "bytes": 6763,
+          "name": "tokenizer.json",
+          "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+        },
+        {
+          "bytes": 236,
+          "name": "tokenizer_config.json",
+          "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+        }
+      ],
+      "files_verified": true,
+      "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+      "maximum_token_id_exclusive": 260,
+      "receipt": null,
+      "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+      "vocab_size": 260
+    }
+  },
+  "per_context": [
+    {
+      "allocation_stratum": null,
+      "domain": "garden",
+      "index": 0,
+      "max": 0.0019943248685364476,
+      "mean": 0.0007337451610329949,
+      "scored_rows": 63,
+      "window_id": "final-0000"
+    },
+    {
+      "allocation_stratum": null,
+      "domain": "arithmetic",
+      "index": 1,
+      "max": 0.005044614761353234,
+      "mean": 0.0008948578801078056,
+      "scored_rows": 63,
+      "window_id": "final-0001"
+    },
+    {
+      "allocation_stratum": null,
+      "domain": "code",
+      "index": 2,
+      "max": 0.005562118890622922,
+      "mean": 0.0009589654666779309,
+      "scored_rows": 63,
+      "window_id": "final-0002"
+    },
+    {
+      "allocation_stratum": null,
+      "domain": "instructions",
+      "index": 3,
+      "max": 0.002863526238924269,
+      "mean": 0.0008777956272695692,
+      "scored_rows": 63,
+      "window_id": "final-0003"
+    }
+  ],
+  "per_domain": {
+    "arithmetic": 0.0008948578801078056,
+    "code": 0.0009589654666779309,
+    "garden": 0.0007337451610329949,
+    "instructions": 0.0008777956272695692
+  },
+  "per_stratum": null,
+  "receipt_sha256": "1fd8c1201c728b1d06df903e30262834541f8d7f85ce6c612ee4184b703ff254",
+  "reference": {
+    "capture_content_digest": "d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39",
+    "dataset_id": "fidelity--quant-integration.root",
+    "dataset_sha256": "4cdf53f91210943b5c87bcddcfbee61034ed056af87d1b174cec28d7a2405ae0",
+    "form": "hidden",
+    "head": {
+      "bits": 16,
+      "quantized": false,
+      "raw_tensor_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+      "source": "native",
+      "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510"
+    },
+    "label": "reference",
+    "lane": "other",
+    "lane_identity_sha256": "34f640ff350d24093734d97a81c1048ce0443d8b3c7bda0953f8283b8ea9269f",
+    "repository": null,
+    "revision": null,
+    "role": "root",
+    "scope_digest": "attn.o=native:bf16@16|attn.qkv=native:bf16@16|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=native:bf16@16|mlp.gate=native:bf16@16|mlp.up=native:bf16@16|moe.experts=native:bf16@16|norm=native:bf16@16|head=native|kv=bf16",
+    "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd",
+    "weights": {
+      "artifact_ref": null,
+      "checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e",
+      "codec": null,
+      "config_sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7",
+      "declared_bits": null,
+      "declared_head_bits": 16,
+      "index_sha256": null,
+      "model_ref": null,
+      "model_revision": null,
+      "quantized": false,
+      "repository": "/tmp/qfs-complete-tiny-fixture/model",
+      "revision": null
+    }
+  },
+  "schema": "malaiwah.fidelity-comparison-receipt.v1",
+  "self_compare": null,
+  "submission": {
+    "emitted": false,
+    "path": null,
+    "receipt_sha256": null,
+    "refusal": null,
+    "submission_schema": null
+  },
+  "tokenwise": {
+    "bytes": 2144,
+    "path": "tokenwise-kld.npy",
+    "sha256": "9970e64072dd76ae08651c684214da51b94e81fedfc6177c878326d03e55181b"
+  },
+  "top1_agreement": 0.5873015873015873,
+  "top1_discordant_contexts": [],
+  "uncertainty": {
+    "ci95_high": null,
+    "ci95_low": null,
+    "cluster_unit": null,
+    "clusters": null,
+    "method": "none",
+    "samples": null,
+    "seed": null
+  }
+}

raw/candidates/ct-mxfp4/evidence/comparison/tokenwise-kld.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9970e64072dd76ae08651c684214da51b94e81fedfc6177c878326d03e55181b
+size 2144

raw/candidates/ct-mxfp4/evidence/panel-binding.json

@@ -0,0 +1,102 @@
+{
+  "content": {
+    "archive": {
+      "algorithm": "sha256(ustar: sorted regular files; mode=0644; uid=gid=mtime=0)",
+      "bytes": 20480,
+      "compression": "none",
+      "format": "ustar",
+      "sha256": "389706dfa3335a75ec70a40d079995af1a60c05512334125dbd336790f12702f"
+    },
+    "manifest": [
+      {
+        "bytes": 192,
+        "path": "arrays/causal-mask-64.npy",
+        "sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02"
+      },
+      {
+        "bytes": 384,
+        "path": "arrays/final-0000.tokens.npy",
+        "sha256": "d60234f4b3b6044fb21eca18f7df82de9822f388e4d93d19d8a4eb69d6d490af"
+      },
+      {
+        "bytes": 384,
+        "path": "arrays/final-0001.tokens.npy",
+        "sha256": "20188213e4dfaaed04b1396a4ec46a5a2a6236c87d870c64d7102568d585ceea"
+      },
+      {
+        "bytes": 384,
+        "path": "arrays/final-0002.tokens.npy",
+        "sha256": "00341b323c286034a1c3a80ae66318cb6ce67153c08b1eedf6990b7a0310662c"
+      },
+      {
+        "bytes": 384,
+        "path": "arrays/final-0003.tokens.npy",
+        "sha256": "a13ec9b40a39027396ec752ee8595e0613569b78e27ef25d254f4f9a82d8775a"
+      },
+      {
+        "bytes": 4195,
+        "path": "panel.json",
+        "sha256": "fe67acc87ce549ce174fa521a775e9e41c8e1ccbc66155ee673a080fea5e24a6"
+      },
+      {
+        "bytes": 7808,
+        "path": "panel.receipt.json",
+        "sha256": "b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263"
+      }
+    ],
+    "manifest_sha256": "e5f2d6410a6290da1197016c5b57ae522a3570665215ddc23e2a0bd45f912181"
+  },
+  "panel": {
+    "bytes": 4195,
+    "context_length": 64,
+    "contexts": 4,
+    "file": "panel.json",
+    "id": "panel--glm-moe-dsa-tiny.malaiwah.cpu-smoke-v1",
+    "name": "Tiny random-init GLM-MoE-DSA synthetic CPU smoke (not a benchmark)",
+    "positions_per_context": 63,
+    "role": "final",
+    "scored_positions_total": 252,
+    "sha256": "fe67acc87ce549ce174fa521a775e9e41c8e1ccbc66155ee673a080fea5e24a6",
+    "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36"
+  },
+  "receipt": {
+    "bytes": 7808,
+    "declared_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5",
+    "file": "panel.receipt.json",
+    "receipt_file_sha256": "b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263",
+    "receipt_seal_mode": "self-blank"
+  },
+  "schema": "malaiwah.resolved-panel.v1",
+  "tokenizer": {
+    "files": [
+      {
+        "bytes": 1479,
+        "name": "config.json",
+        "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+      },
+      {
+        "bytes": 216,
+        "name": "generation_config.json",
+        "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+      },
+      {
+        "bytes": 6763,
+        "name": "tokenizer.json",
+        "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+      },
+      {
+        "bytes": 236,
+        "name": "tokenizer_config.json",
+        "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+      }
+    ],
+    "files_verified": true,
+    "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+    "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+    "maximum_token_id_exclusive": 260,
+    "receipt": null,
+    "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+    "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+    "vocab_size": 260
+  }
+}

raw/candidates/ct-mxfp4/evidence/root/README.md

@@ -0,0 +1,95 @@
+---
+license: mit
+tags:
+- fidelity
+- kl-divergence
+- fidelity-provenance
+- fidelity-dataset
+- fidelity-root
+x_fidelity:
+  spec: https://github.com/malaiwah/quant-fidelity-suite/blob/main/docs/FIDELITY-DATASET-SPEC.md
+  spec_version: fidelity-provenance/v1
+  role: fidelity-dataset
+  captured_model:
+    repository: /tmp/qfs-complete-tiny-fixture/model
+    revision: null
+    role: root
+  form: hidden
+  lane: other
+  panel:
+    panel_id: panel--panel
+    suite_token_hash_sha256: 525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36
+    contexts: 4
+    scored_positions: 252
+    repository: null
+    revision: null
+  head:
+    policy: native
+    quantized: false
+    bits: 16
+    lm_head_tensor_content_sha256: 2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510
+    lm_head_file_sha256: 07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913
+    final_norm_tensor_content_sha256: null
+    final_norm_file_sha256: null
+    equality_receipt: null
+    replay_permitted: true
+  scope_digest: attn.o=native:bf16@16|attn.qkv=native:bf16@16|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=native:bf16@16|mlp.gate=native:bf16@16|mlp.up=native:bf16@16|moe.experts=native:bf16@16|norm=native:bf16@16|head=native|kv=bf16
+  seal:
+    manifest: fidelity-dataset.json
+    dataset_sha256: null
+    checksums: checksums.txt
+    checksums_sha256: '0000000000000000000000000000000000000000000000000000000000000000'
+    note: 'written before the seal: dataset_sha256 and checksums_sha256 are digests of the tree this file
+      is inside, so they cannot appear in it. Read them from fidelity-dataset.json.'
+  interop:
+    compatible_with:
+    - kimi-k3-distribution-fidelity/1
+    k3_compat_emitted: false
+  registry:
+    dataset: malaiwah/quant-fidelity-registry
+    schema_version: quant-fidelity-registry/v1
+    artifact_id: null
+    measurement_ids: []
+    snapshot: null
+  fidelity_dataset:
+    repository: null
+    revision: null
+    dataset_sha256: null
+    capture_content_digest: d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39
+    form: hidden
+    role: root
+  measurements: []
+---
+# Local quant integration root
+
+A **root** fidelity dataset in **hidden form**, produced by `engines/tools/hf_capture.py` from `/tmp/qfs-complete-tiny-fixture/model`.
+
+## The cut
+
+the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits after it). Same cut as engines/tools/hidden_replay.py and as Festr's kimi-k3 hidden-replay qualification.
+
+## What is here
+
+| | |
+|---|---|
+| records | 4 |
+| scored positions | 252 |
+| context length | 64 |
+| hidden width | 64 |
+| vocab size | 260 |
+| capture bytes (hidden form) | 33348 |
+| the same capture in logit form would be | 262080 bytes (8x) |
+| scope policy | native |
+| lane | other |
+
+## Does the model still generate sensibly
+
+`The capital of France is` -> `�` (p=0.0067, entropy 5.548 of a possible 5.561 nats). Recorded, not enforced. One extra window through the same schedule; its hidden state was discarded and is not part of this dataset.
+
+## How to check it
+
+```
+bin/fidelity-dataset verify <this directory>
+```
+
+Full specification: <https://github.com/malaiwah/quant-fidelity-suite/blob/main/docs/FIDELITY-DATASET-SPEC.md>

raw/candidates/ct-mxfp4/evidence/root/capture/hidden_0000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7d5a505a8bdb71b45df052427ee2f917c390c4eb751dc586bfc213192c273e5
+size 8337

raw/candidates/ct-mxfp4/evidence/root/capture/hidden_0001.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fbd2c627ad45d2c1c39520ab9ba17f5d8aa1d1e73952682a488b732764574a6
+size 8337

raw/candidates/ct-mxfp4/evidence/root/capture/hidden_0002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eaf7e5903afff4ebeff8ba42353c3ebd58cc120c009d5038ea9393242bf82e4d
+size 8337

raw/candidates/ct-mxfp4/evidence/root/capture/hidden_0003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b7729fe05d7c7ef4b80be6eb94f511e262765e417037cceb7d98ce80c17685d
+size 8337

raw/candidates/ct-mxfp4/evidence/root/capture/manifest.json

@@ -0,0 +1,159 @@
+{
+  "capture_content_digest": "d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39",
+  "context_length": 64,
+  "coverage": {
+    "complete": true,
+    "declared_records": 4,
+    "index_range": [
+      0,
+      3
+    ],
+    "missing_indices_sample": [],
+    "present_records": 4,
+    "shard_of": null,
+    "subset_detail": null
+  },
+  "created_utc": "2026-09-07T16:52:44Z",
+  "dtype": "BF16",
+  "dtype_lossless": true,
+  "form": "hidden",
+  "format_version": 1,
+  "hidden_width": 64,
+  "receipt_sha256": "be259be5d93938a69a87b8f141ccd8f5afe4f90dfb1089586d6a49ae7225d5a0",
+  "records": [
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 0,
+      "document_id": "embedded-synthetic/garden.txt",
+      "domain": "garden",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.034,
+      "file": "hidden_0000.safetensors",
+      "index": 0,
+      "key": "hidden_states",
+      "payload_sha256": "80897a67cfdc22f410255f2463193e7d765de5500d62da2c13aacbd3b9507083",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "a7d5a505a8bdb71b45df052427ee2f917c390c4eb751dc586bfc213192c273e5",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "80897a67cfdc22f410255f2463193e7d765de5500d62da2c13aacbd3b9507083",
+      "token_ids_json_sha256": "08f8d8d449f7f679b24ada75dc4db5b0108508e16ccefb0f61cad6144593e315",
+      "token_ids_sha256_legacy": "9d5378f0316be57e37f80d2bae7895ba4a1140c302c87ff8cce2fbd9afc83b77",
+      "window_id": "final-0000",
+      "window_index": 0
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 1,
+      "document_id": "embedded-synthetic/arithmetic.txt",
+      "domain": "arithmetic",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.027,
+      "file": "hidden_0001.safetensors",
+      "index": 1,
+      "key": "hidden_states",
+      "payload_sha256": "1fc13aaf069c53d5592198daeacfc5f652abf5b9ba552213859596c32e7c4046",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "9fbd2c627ad45d2c1c39520ab9ba17f5d8aa1d1e73952682a488b732764574a6",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "1fc13aaf069c53d5592198daeacfc5f652abf5b9ba552213859596c32e7c4046",
+      "token_ids_json_sha256": "5bc88e4e47faf077b2fc85ae0758eee5308003ff0fefa5d6691a7f24f22ecc46",
+      "token_ids_sha256_legacy": "e0f95114b71a46cb489539dc38909fb67fb378f25fafbb513e32690cc7f11b37",
+      "window_id": "final-0001",
+      "window_index": 1
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 2,
+      "document_id": "embedded-synthetic/code.txt",
+      "domain": "code",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.028,
+      "file": "hidden_0002.safetensors",
+      "index": 2,
+      "key": "hidden_states",
+      "payload_sha256": "98b8d9e1ccb196898fda27cb5f625f348aa0046121c39484a4af7cb042e1be92",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "eaf7e5903afff4ebeff8ba42353c3ebd58cc120c009d5038ea9393242bf82e4d",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "98b8d9e1ccb196898fda27cb5f625f348aa0046121c39484a4af7cb042e1be92",
+      "token_ids_json_sha256": "51b956fbf40161ce10fe3da6c1e186368a567676467030535468064aea0d7d07",
+      "token_ids_sha256_legacy": "482b890c7b0e2f7400257c43680093b7f2b90d6df5361c6de3ae97ddfdd2504e",
+      "window_id": "final-0002",
+      "window_index": 2
+    },
+    {
+      "allocation_stratum": null,
+      "attention_mask_sha256": "a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02",
+      "context_index": 3,
+      "document_id": "embedded-synthetic/instructions.txt",
+      "domain": "instructions",
+      "dtype": "BF16",
+      "elapsed_seconds": 0.027,
+      "file": "hidden_0003.safetensors",
+      "index": 3,
+      "key": "hidden_states",
+      "payload_sha256": "01646a47a79cc47760fc0645c30bc9e37a930c26ac84cd3d1a1d5077a2f48df8",
+      "prediction_positions": 63,
+      "raw_chunks_retained": false,
+      "request_id": null,
+      "role": "final",
+      "scored_rows": 63,
+      "semantic_class": null,
+      "sha256": "9b7729fe05d7c7ef4b80be6eb94f511e262765e417037cceb7d98ce80c17685d",
+      "shape": [
+        63,
+        64
+      ],
+      "size_bytes": 8337,
+      "source_cluster_id": null,
+      "tensor_content_sha256": "01646a47a79cc47760fc0645c30bc9e37a930c26ac84cd3d1a1d5077a2f48df8",
+      "token_ids_json_sha256": "8653267d2a6dc67550759c29ccb80a7a5529d927e6c367c681f5018d55497a2d",
+      "token_ids_sha256_legacy": "12e01edd3900ff2ca82621feb5504f24b003a53c0afe98a9ee0b490eb93141d4",
+      "window_id": "final-0003",
+      "window_index": 3
+    }
+  ],
+  "run_name": "hf-capture",
+  "runtime_manifest": "../runtime/capture-runtime.json",
+  "runtime_manifest_sha256": "60510dc79ef1eb967e64492e56911dec31e885c167c9b0d2ffa7d083e27a941b",
+  "schema": "malaiwah.fidelity-capture-manifest.v1",
+  "scored_rows_per_context": 63,
+  "semantic_point": "after_final_rmsnorm_before_lm_head",
+  "tensor_key": "hidden_states",
+  "total_scored_rows": 252,
+  "total_size_bytes": 33348,
+  "vocab_size": 260
+}

raw/candidates/ct-mxfp4/evidence/root/checksums.txt

@@ -0,0 +1,20 @@
+df28a18b7e9c0d39ac079235136da25818038cc09314a890063058e94bf1c226  README.md
+a7d5a505a8bdb71b45df052427ee2f917c390c4eb751dc586bfc213192c273e5  capture/hidden_0000.safetensors
+9fbd2c627ad45d2c1c39520ab9ba17f5d8aa1d1e73952682a488b732764574a6  capture/hidden_0001.safetensors
+eaf7e5903afff4ebeff8ba42353c3ebd58cc120c009d5038ea9393242bf82e4d  capture/hidden_0002.safetensors
+9b7729fe05d7c7ef4b80be6eb94f511e262765e417037cceb7d98ce80c17685d  capture/hidden_0003.safetensors
+3880f9a0ee35848dcae0ce89149c9b720d066f788b6d724c0710e69aa3251b40  capture/manifest.json
+777b038f747bed25d1f5db8adbfd2a78c7762fb5c870ff8c3c785a984d48a168  head/head.json
+07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913  head/weight.safetensors
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0000.npy
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0001.npy
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0002.npy
+a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02  panel/masks/context-0003.npy
+b0ffc18b0de3b30bb797855a55097083222f6c651bf207dad58c3a7f02384263  panel/panel-receipt.json
+68efa258ea3c5d4ba759c3c15aadd6e8c984711831bea3f335903109345cdcef  panel/panel.json
+08f8d8d449f7f679b24ada75dc4db5b0108508e16ccefb0f61cad6144593e315  panel/tokens/context-0000.json
+5bc88e4e47faf077b2fc85ae0758eee5308003ff0fefa5d6691a7f24f22ecc46  panel/tokens/context-0001.json
+51b956fbf40161ce10fe3da6c1e186368a567676467030535468064aea0d7d07  panel/tokens/context-0002.json
+8653267d2a6dc67550759c29ccb80a7a5529d927e6c367c681f5018d55497a2d  panel/tokens/context-0003.json
+60510dc79ef1eb967e64492e56911dec31e885c167c9b0d2ffa7d083e27a941b  runtime/capture-runtime.json
+769c103cf430360bf1b79a4d027f5bd073ce3f61204d9da77adf3ad494c27e73  validation/structural-validation.json

raw/candidates/ct-mxfp4/evidence/root/fidelity-dataset.json

@@ -0,0 +1,389 @@
+{
+  "capture": {
+    "capture_content_digest": "d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39",
+    "dtype": "BF16",
+    "dtype_lossless": true,
+    "form": "hidden",
+    "head_not_separable_reason": null,
+    "head_separable": true,
+    "hidden_width": 64,
+    "lossy_codec": null,
+    "manifest_file": "capture/manifest.json",
+    "manifest_file_sha256": "3880f9a0ee35848dcae0ce89149c9b720d066f788b6d724c0710e69aa3251b40",
+    "records_count": 4,
+    "scored_rows_total": 252,
+    "semantic_point": "after_final_rmsnorm_before_lm_head",
+    "tensor_key": "hidden_states",
+    "total_size_bytes": 33348,
+    "vocab_size": 260
+  },
+  "coverage": {
+    "complete": true,
+    "declared_records": 4,
+    "index_range": [
+      0,
+      3
+    ],
+    "missing_indices_sample": [],
+    "present_records": 4,
+    "shard_of": null,
+    "subset_detail": null
+  },
+  "created_utc": "2026-09-07T16:52:44Z",
+  "dataset": {
+    "author": {
+      "handle": null,
+      "is_registry_maintainer": false,
+      "name": "malaiwah",
+      "role": "capture-author",
+      "url": null
+    },
+    "base_capture": null,
+    "id": "fidelity--quant-integration.root",
+    "license": "mit",
+    "name": "Local quant integration root",
+    "qualification": null,
+    "repository": null,
+    "revision": null,
+    "role": "root",
+    "structural_status": "sealed"
+  },
+  "dataset_sha256": "4cdf53f91210943b5c87bcddcfbee61034ed056af87d1b174cec28d7a2405ae0",
+  "determinism": {
+    "cold_start_per_run": true,
+    "distinct_evidence_hash_count": 1,
+    "evidence_hashes": [
+      "d3a8907faca213c298b4ff9e77dd5f041bba4353603447a30457e33c33386d39"
+    ],
+    "evidence_kind": "hidden_state_tensor_sha256",
+    "identical_across_runs": null,
+    "note": "a single cold capture; cross-run determinism is NOT asserted here. Capture a second cold run and run `fidelity-dataset compare --self-compare` -- an exactly 0.0 result is the SC-1 reproduction confirmation.",
+    "repeat_noise": null,
+    "repeats": [],
+    "run_count": 1
+  },
+  "disclosures": [
+    {
+      "affects_comparability": false,
+      "code": "rtn_format_fixture",
+      "detail": "Local RTN-format fixture; optimizer-not-run; no model-quality or serving-kernel equivalence claim.",
+      "severity": "info"
+    },
+    {
+      "affects_comparability": false,
+      "code": "reduced_run_count",
+      "detail": "this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus `fidelity-dataset compare --self-compare`, whose exactly-0.0 result is the SC-1 reproduction confirmation.",
+      "severity": "caveat"
+    }
+  ],
+  "format_version": 1,
+  "generation_sanity_probe": {
+    "degenerate": false,
+    "enforced": false,
+    "entropy_nats": 5.548019647470335,
+    "expect": null,
+    "logit_spread": 0.986328125,
+    "note": "one extra window through the same schedule; its hidden state is discarded and never enters the dataset.",
+    "prompt": "The capital of France is",
+    "schema": "malaiwah.generation-sanity-probe.v1",
+    "status": "recorded",
+    "top1_id": 113,
+    "top1_probability": 0.006720104384265371,
+    "top1_text": "�",
+    "top_k": [
+      {
+        "id": 113,
+        "probability": 0.006720104384265371,
+        "text": "�"
+      },
+      {
+        "id": 237,
+        "probability": 0.006412368399474785,
+        "text": "�"
+      },
+      {
+        "id": 193,
+        "probability": 0.006035639491843906,
+        "text": "\u0001"
+      },
+      {
+        "id": 236,
+        "probability": 0.0055386141320063614,
+        "text": "�"
+      },
+      {
+        "id": 44,
+        "probability": 0.0054740873228682424,
+        "text": "I"
+      }
+    ],
+    "uniform_entropy_nats": 5.560681631015528,
+    "vocab_size": 260
+  },
+  "head": {
+    "applied_in_capture": false,
+    "bias": null,
+    "bits": 16,
+    "compat_tensor_key": "weight",
+    "dtype": "BF16",
+    "equality_receipt": null,
+    "file": "head/weight.safetensors",
+    "file_sha256": "07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913",
+    "final_norm": {
+      "applied_at_replay": false,
+      "applied_in_capture": true,
+      "dtype": null,
+      "file": null,
+      "file_sha256": null,
+      "shape": null,
+      "tensor_content_sha256": null,
+      "tensor_key": null
+    },
+    "head_json": "head/head.json",
+    "present": true,
+    "quantized": false,
+    "raw_tensor_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+    "shape": [
+      260,
+      64
+    ],
+    "source": "native",
+    "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+    "tensor_key": "lm_head.weight"
+  },
+  "interop": {
+    "adapted_from": null,
+    "compatible_with": [
+      "kimi-k3-distribution-fidelity/1"
+    ],
+    "divergences": [
+      {
+        "field": "head",
+        "id": "D-1",
+        "reason": "every capture declares its OWN head identity by tensor content digest; the comparator refuses a hidden-form comparison across differing heads."
+      },
+      {
+        "field": "runtime.lane",
+        "id": "D-2",
+        "reason": "lane is required and gated at compare time so registry BIAS-006 is checked before a number exists, not at submission."
+      },
+      {
+        "field": "panel.scoring_window",
+        "id": "D-3",
+        "reason": "scoring_window is part of panel identity, not a comparator flag: a score_from=1024 number is a different panel, not a variant."
+      },
+      {
+        "field": "comparison_kind",
+        "id": "D-4",
+        "reason": "self-compare is a first-class mode asserting exactly 0.0."
+      },
+      {
+        "field": "estimator.accumulation_dtype",
+        "id": "D-5",
+        "reason": "required, fp64 default; an fp32 estimator lands in a different comparability class on the same data and that must be visible."
+      },
+      {
+        "field": "coverage",
+        "id": "D-6",
+        "reason": "declared vs present record counts with shard_of; fixes our own manifest that claimed 5,120 captures in a 512-file repository."
+      },
+      {
+        "field": "panel",
+        "id": "D-7",
+        "reason": "the panel is a separately referenceable object, which is what makes a candidate capture publishable standalone."
+      },
+      {
+        "field": "capture.lossy_codec",
+        "id": "D-8",
+        "reason": "lets a llama.cpp .kld be ingested without laundering a lossy capture as exact."
+      },
+      {
+        "field": "capture.records[].attention_mask_sha256",
+        "id": "D-9",
+        "reason": "our packed and streaming lanes vary mask construction; a single-request capture path makes it invariant and has no equivalent field."
+      },
+      {
+        "field": "capture.head_separable",
+        "id": "D-10",
+        "reason": "hidden form is the default; logit form must declare head_separable:false with a reason."
+      }
+    ],
+    "k3_compat_emitted": false,
+    "note": null
+  },
+  "panel": {
+    "contamination": {
+      "checked": false,
+      "hits": null,
+      "method": "not established",
+      "receipt": null
+    },
+    "context_length": 64,
+    "contexts": 4,
+    "panel_file": "panel/panel.json",
+    "panel_file_sha256": "68efa258ea3c5d4ba759c3c15aadd6e8c984711831bea3f335903109345cdcef",
+    "panel_id": "panel--panel",
+    "panel_receipt_file": "panel/panel-receipt.json",
+    "panel_receipt_sha256": "3432d5ccee92559c0999bc80d976183faaffc1accf7c00d5fdca151b0e357ba5",
+    "panel_token_sha256_legacy": "3c9467c85be43d238926f6d18bf825c2e4748814c95b8b0e8b8f1420546c3089",
+    "remap_file": null,
+    "repository": null,
+    "revision": null,
+    "scored_positions_total": 252,
+    "scoring_window": {
+      "dropped_positions_total": 0,
+      "min_left_context_tokens": 1,
+      "policy": "every causal prediction position of every window is scored; nothing is dropped",
+      "score_from": 0,
+      "windowed": false
+    },
+    "suite_token_hash_sha256": "525cb6c625096ddb0ab19771a252e83223de98cc984733ef7da25a5eb076db36",
+    "tokenizer": {
+      "files": [
+        {
+          "bytes": 1479,
+          "name": "config.json",
+          "sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7"
+        },
+        {
+          "bytes": 216,
+          "name": "generation_config.json",
+          "sha256": "1d59e0fb61ca550ea84bd37a6bdf553ab364bc2689b17ba69f6a1adbc9d61e34"
+        },
+        {
+          "bytes": 6763,
+          "name": "tokenizer.json",
+          "sha256": "8557198de5989f11dc18c955f6bc98143d2f301c141a8dcab7461a14f78591da"
+        },
+        {
+          "bytes": 236,
+          "name": "tokenizer_config.json",
+          "sha256": "e0b489082ac322f131a544ca0833983e1cecb99c2817b2e72184e5ffbdb6ce63"
+        }
+      ],
+      "files_verified": true,
+      "id": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "identity_sha256": "d352eac9d82f64010d4094b2dc548d67bf6020df8f7fd4f529bb66f22ceaa9b1",
+      "maximum_token_id_exclusive": 260,
+      "receipt": null,
+      "repository": "malaiwah/glm-moe-dsa-tiny-random-bf16",
+      "revision": "45563636ef723acfb826755493447dc40c7a0c37",
+      "vocab_size": 260
+    }
+  },
+  "runtime": {
+    "backend_identity_sha256": null,
+    "file": "runtime/capture-runtime.json",
+    "file_sha256": "60510dc79ef1eb967e64492e56911dec31e885c167c9b0d2ffa7d083e27a941b",
+    "lane": "other",
+    "lane_identity_sha256": "34f640ff350d24093734d97a81c1048ce0443d8b3c7bda0953f8283b8ea9269f",
+    "lane_inferred": false,
+    "runtime_reader_sha256": null,
+    "source": "native",
+    "stack_fingerprint_sha256": "a83c4de13e06b8727c3a05959e470708b1fa396b0fcbd78c2f9deac666fea6cd"
+  },
+  "schema": "malaiwah.fidelity-dataset.v1",
+  "scope": {
+    "assignments": [
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "embed_tokens",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "attn.qkv",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "attn.o",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "mlp.gate",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "mlp.up",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "mlp.down",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "moe.experts",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "norm",
+        "treatment": "native"
+      },
+      {
+        "bits_per_weight": 16,
+        "format": "bf16",
+        "layer_range": null,
+        "tensor_class": "lm_head",
+        "treatment": "native"
+      }
+    ],
+    "head_policy": "native",
+    "kv_cache_dtype": "bf16",
+    "policy": "native",
+    "scope_digest": "attn.o=native:bf16@16|attn.qkv=native:bf16@16|embed_tokens=native:bf16@16|lm_head=native:bf16@16|mlp.down=native:bf16@16|mlp.gate=native:bf16@16|mlp.up=native:bf16@16|moe.experts=native:bf16@16|norm=native:bf16@16|head=native|kv=bf16"
+  },
+  "seal": {
+    "checksums_file": "checksums.txt",
+    "checksums_format": "sha256sum",
+    "checksums_sha256": "654ea33f52cbbdf4643fb7abdfc11ebd3e5dc73bd9d448556d5cc1a46a13ae06",
+    "covers": [
+      "every published file except checksums.txt and fidelity-dataset.json, via checksums.txt",
+      "fidelity-dataset.json itself, via the self-blanked seal",
+      "every capture tensor a second time by CONTENT, via capture.capture_content_digest"
+    ],
+    "excludes": [
+      "checksums.txt",
+      "fidelity-dataset.json"
+    ],
+    "external_anchor": null,
+    "method": "self_blanked_canonical_json_sha256",
+    "seal_field": "dataset_sha256"
+  },
+  "upstream_receipts": [],
+  "weights": {
+    "artifact_ref": null,
+    "checkpoint_identity_sha256": "800be70604bef0c3adfb429774f44756d60bf23d141e82d24a032ce25ce0451e",
+    "codec": null,
+    "config_sha256": "7d5880677655ece868c12425a204f539243244af121bc8ad09b881a7546930d7",
+    "declared_bits": null,
+    "declared_head_bits": 16,
+    "index_sha256": null,
+    "model_ref": null,
+    "model_revision": null,
+    "quantized": false,
+    "repository": "/tmp/qfs-complete-tiny-fixture/model",
+    "revision": null
+  }
+}

raw/candidates/ct-mxfp4/evidence/root/head/head.json

@@ -0,0 +1,34 @@
+{
+  "applied_in_capture": false,
+  "bias": null,
+  "bits": 16,
+  "compat_tensor_key": "weight",
+  "dtype": "BF16",
+  "equality_receipt": null,
+  "file": "head/weight.safetensors",
+  "file_sha256": "07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913",
+  "final_norm": {
+    "applied_at_replay": false,
+    "applied_in_capture": true,
+    "dtype": null,
+    "file": null,
+    "file_sha256": null,
+    "shape": null,
+    "tensor_content_sha256": null,
+    "tensor_key": null
+  },
+  "format_version": 1,
+  "note": "the head is shipped verbatim from the checkpoint so a third party can replay logits = hidden @ head^T without the weights",
+  "present": true,
+  "quantized": false,
+  "raw_tensor_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+  "receipt_sha256": "2da3cfbfc607296b3bb32eb6fec1f122e4ce30fe9ea96653e5bb1b4ebab6514f",
+  "schema": "malaiwah.fidelity-head-identity.v1",
+  "shape": [
+    260,
+    64
+  ],
+  "source": "native",
+  "tensor_content_sha256": "2364995d4009b32279b09045e48bf7cdceefcdaebdba3f70b003919a236f9510",
+  "tensor_key": "lm_head.weight"
+}

raw/candidates/ct-mxfp4/evidence/root/head/weight.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07b6cafe9bd551f79e76e98d2fd2441927a2244bd30e93b528b525bf4db1b913
+size 33365

raw/candidates/ct-mxfp4/evidence/root/panel/masks/context-0000.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/root/panel/masks/context-0001.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/root/panel/masks/context-0002.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192

raw/candidates/ct-mxfp4/evidence/root/panel/masks/context-0003.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38fba8bfd9cd25a81b0e37fa144e8e13e84f769ae7f609ceebc94a64f4f5d02
+size 192