Upload article.md with huggingface_hub

article.md

@@ -0,0 +1,76 @@
+---
+title: "FINAL-Bench Quantum: An Open, Neutral Benchmark for Quantum-Computing Methods"
+thumbnail: 
+authors:
+- user: SeaWolf-AI
+tags: [quantum, quantum-error-correction, benchmark, qec, decoders, vqe, qram]
+---
+
+# FINAL-Bench Quantum: An Open, Neutral Benchmark for Quantum-Computing Methods
+
+Quantum-computing results are notoriously hard to compare. A "logical error rate," a
+"query fidelity," or an "approximation ratio" only mean something next to a clearly stated
+protocol — yet papers use different codes, noise models, hardware, and shot counts. **FINAL-Bench
+Quantum** is our attempt at a single, fair yardstick: an open, neutral benchmark suite where
+methods compete in five events under identical, published protocols, and where every number is
+labeled as either *measured here* or *quoted from a source*.
+
+🔗 **Leaderboard:** `huggingface.co/spaces/FINAL-Bench/quantum-bench-leaderboard`
+
+## Two tracks, one rule: say exactly what you measured
+
+- **Track A — Verified.** Methods run here on one frozen, public test set, reported with 95%
+  confidence intervals. These are directly comparable.
+- **Track B — Reported.** Numbers quoted from their papers/sources. Codes, noise models, and
+  hardware differ, so they are **not** directly comparable — and we say so.
+
+We make **no quantum-advantage claims**. Where a result is a simulation, we call it a simulation;
+where it is real hardware, we name the chip and job ID. Where two results are within their
+confidence intervals, we call it a statistical tie rather than crowning a winner.
+
+## The five events
+
+| Event | What it measures | Verified highlight |
+|---|---|---|
+| ① QEC Decoder | logical error rate on a rotated surface code (Stim, circuit noise) | a maximum-likelihood decoder (Tesseract) leads; the BP-class decoders (BeliefMatching, BP+OSD) and a neural ensemble are a statistical tie just behind |
+| ② Optimization | Max-Cut quality (cut / optimum) | QAOA reaches ~0.88 approximation ratio at depth p=3 on a 14-node instance |
+| ③ VQE | molecular ground-state energy vs FCI | H₂ to **chemical accuracy** (0.22 mHa); a real-hardware 4-qubit run on IBM Heron r2 |
+| ④ QRAM | quantum-memory query fidelity | a 4-cell select-QRAM on **IBM Heron r2 at 0.92 mean query fidelity**; a QEC-protected variant reaches 99.6% in simulation |
+| ⑤ Simulation | how large a circuit a classical method handles | **100,000-qubit** Clifford (stabilizer) sampling; 30-qubit exact statevector on a GPU |
+
+## A few results we found interesting
+
+**Decoders are near a floor on idealized syndromes.** On a distance-5 rotated surface code, the
+strong classical decoders (BP+OSD ≈ 0.0095, BeliefMatching ≈ 0.010) and a regularized neural
+ensemble (≈ 0.0098) all sit within each other's 95% CIs — only the brute-force maximum-likelihood
+decoder (Tesseract, ≈ 0.0092) is lower, and it runs at ~79 ms/shot, far too slow for real-time use.
+The practical lesson: on *binary* syndromes there is little headroom; a fancier architecture does
+not beat the floor.
+
+**…but soft information changes the game.** When the decoder is given *analog* (soft) readout
+values instead of thresholded bits, a small neural decoder cuts logical error by **40–46%** versus
+hard-decision decoding on a repetition memory — matching the soft-optimal decoder. This is the same
+mechanism behind Google's AlphaQubit gains: the advantage is in the *information*, not the network.
+
+**Real hardware, honestly bounded.** On IBM Heron r2 processors, a repetition-code memory shows the
+error-correcting distance boundary directly — distance-3 fails to correct two injected errors
+(~97% logical error) while distance-5 corrects them (0.5–3%), a 29–175× reduction across three
+backends (z = 7–552σ). This is a *repetition code* with offline majority-vote decoding — **not**
+surface-code below-threshold, not fault-tolerant QEC, and not a quantum-advantage claim.
+
+## Submit your method
+
+The **Submit** tab takes a method name, links (GitHub / Hugging Face), an email, and an optional
+file. Submissions are stored privately, reproduced under the event's fixed protocol, and the
+submitter is emailed about inclusion. Entries are labeled by origin (country flag + author/group),
+exactly like every other entry.
+
+## Why neutrality is the point
+
+A leaderboard is only useful if you trust it. That means including strong competitors even when
+they beat the host's own entries, quoting sources faithfully, and refusing to round a simulation up
+into a hardware claim. FINAL-Bench Quantum is built around that discipline. Several entries come
+from our own group (VIDRAFT, 🇰🇷); they are shown under the same protocol, confidence intervals, and
+honesty boundaries as Google, IBM, NVIDIA, USTC, Riverlane, and the rest.
+
+*A methods paper is in preparation. Feedback and submissions are welcome.*