Update Exit Lens model card from legacy README

README.md

@@ -1,28 +1,85 @@
 ---
+language:
+- ko
+license: cc-by-nc-4.0
 library_name: onnx
 tags:
-  - onnx
-  - browser
-  - metadata-classification
+- text-classification
+- multi-head-classification
+- onnx
+- browser
 ---
 
-# saibiindex/exitlens
+# Exit Lens Classifier
 
-`saibiindex/exitlens` contains the current Exit Lens metadata classifier artifacts for the Onexit web app.
+Version: `2026-06-27-v2-3model-onnx`
 
-This repository is not a final-answer generation model. The ONNX files only provide metadata predictions used by the Exit Lens runtime. Final user-facing responses remain controlled by the GitHub app's rule pipeline, `SafetyGate`, `FragmentComposer`, and `TemplateValidator`.
+Selected artifacts: `general/model.onnx`, `cult/model.onnx`, and `risk/model.onnx` (`ONNX`)
 
-## Current Structure
+Model SHA256 values are listed in the current role checksums table below.
 
-The default structure is a three-model orchestrated classifier:
+These verified ONNX artifacts are intended for browser-side metadata classification with rule/fragment safety fallback.
+
+## Current repository
+
+This Hugging Face repository is `saibiindex/exitlens`. It supersedes the legacy single-model repository `world08/exit-lens-classifier` for browser runtime downloads. The legacy repository is retained only as read-only model-card and reference material.
+
+The current package has three role-specific ONNX models:
 
 - `general`: ordinary conversation, light support, service-help turns, and hard-negative examples.
-- `cult`: high-risk group context, deceptive recruitment, exit or distancing context, and related coercive-control metadata.
+- `cult`: high-risk group context, deceptive recruitment, exit or distancing context, and coercive-control metadata.
 - `risk`: self-harm, unknown substances, privacy or money pressure, sexual violence or boundary risk, violence, source integrity, and other safety-risk metadata.
 
-The legacy Hugging Face repository `world08/exit-lens-classifier` is a read-only migration source for model-card/reference material only. It is not the default runtime path for the current app. Any legacy single-model classifier material should be treated as historical context.
+## Korean-language model only
+
+These models are intended exclusively for Korean-language input. They were not designed or evaluated for English-language users or multilingual use. If your application will not operate in Korean, do not use these models. Inputs in other languages may produce unreliable or meaningless classifications.
+
+## License and attribution
+
+This model is released under the [Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/).
+
+- Commercial use is prohibited.
+- Attribution is required for every permitted use, including personal study, education, research, experiments, and non-commercial projects.
+- Attribution should identify **Exit Lens Classifier**, `world08`, `saibiindex/exitlens`, and the `pseudo-index` project, and should include a link to this model repository and the CC BY-NC 4.0 license.
+- Review and comply with the full license terms before redistributing the model or derivative material.
+
+## Project scope and change policy
+
+These models were built specifically for the maintainer's personal `pseudo-index` project and the Onexit web app. They are not general-purpose classifiers, a standardized public service, or a drop-in solution for unrelated products. The maintainer may modify, replace, retrain, relabel, restrict, or remove the model at their discretion, without guaranteeing backward compatibility or continued availability.
+
+## Not an LLM or counseling service
+
+These models are multi-head classifiers, not large language models. They do not conduct conversations, provide counseling, generate final advice, or make professional decisions. Do not ask these models for counseling and do not deploy them as a substitute for a counselor, clinician, lawyer, emergency service, or other qualified professional.
+
+## Purpose
 
-## Files
+Exit Lens Classifier is a set of supporting risk-signal classifiers for counseling inputs related to high-risk groups and coercive or deceptive situations. It is not a final counseling AI and does not directly decide what a user should do.
+
+Each role model predicts four metadata heads:
+
+- `primaryDomain`
+- `templateId`
+- `urgency`
+- `slotProfile`
+
+Model output is not a user-facing answer. Final responses in Onexit remain controlled by `SafetyGate`, `ResponsePlanner`, `FragmentComposer`, and `TemplateValidator`.
+
+## Safety and limitations
+
+- The models and their maintainer do not take responsibility for a user's deployment choices, interpretations, decisions, actions, generated downstream content, or resulting outcomes. To the extent permitted by law, users and operators assume responsibility for their use of the model and for compliance with applicable laws, policies, and professional duties.
+- The models are provided "as is," without warranties of accuracy, fitness, reliability, safety, or availability.
+- These models do not replace legal, medical, mental-health, or counseling professionals.
+- Self-harm, violence, sexual violence, and personal-information harm require safety-gate handling and urgent guidance before ordinary classification output.
+- Do not use the classifiers to condemn or disparage religions in general.
+- Do not present unsupported criminal allegations as facts.
+- Review the package before publication to ensure that model files, metadata, and this card contain no private or sensitive source text.
+- Training artifacts are intended for storage/evaluation. Only exported and validated ONNX packages may be marked browser-ready.
+
+## Package files
+
+`manifest.json` and each role `manifest.json` record file roles and model checksums. `sha256sums.txt` covers every uploaded package file except itself. Legacy single-model files from `world08/exit-lens-classifier` are kept under `legacy/` as read-only reference material. Current browser downloads use `saibiindex/exitlens`.
+
+## Current 3-model package files
 
 ```txt
 README.md
@@ -47,7 +104,7 @@ legacy/
   README.md or other copied legacy reference files when available
 ```
 
-## Role Checksums
+## Current role checksums
 
 | Role | Model file | SHA256 | Size |
 | --- | --- | --- | ---: |
@@ -71,13 +128,7 @@ Get-FileHash cult/model.onnx -Algorithm SHA256
 Get-FileHash risk/model.onnx -Algorithm SHA256
 ```
 
-## Runtime Boundary
-
-These ONNX files run in a browser ONNX runtime as optional metadata classifiers. They do not decide final risk by themselves, do not provide legal findings, and do not answer users directly. The application keeps deterministic rules, safety gates, response planning, vetted fragments, and template validation as the final authority.
-
-The model roles are scoped to conversation metadata. They must not be used to attack a religion or protected belief category as a whole, and they should not be described as a definitive labeler for people, organizations, or legal facts.
-
-## Legacy Source
+## Legacy source
 
 Legacy model-card/source files copied into `legacy/`: `README.md`, `label_map.json`, `model_manifest.json`, `special_tokens_map.json`, `tokenizer.json`, `tokenizer_config.json`, `vocab.txt`