[qmc] yanwang: map atom-reload gains across 192M surface-code cell-shots

tracks/qmc/solutions/yanwang-66/README.md

@@ -0,0 +1,160 @@
+# Dynamic atom reloading for loss-tolerant surface-code memory
+
+## 评委入口（Start here）
+
+> ### [▶ 直接打开自包含 HTML 技术报告（建议首先阅读）](https://htmlpreview.github.io/?https://github.com/Thatht137/quantum.harness/blob/challenge/qmc-atom-reload-surface-code/tracks/qmc/solutions/yanwang-66/report/report.html)
+>
+> 核心结论、4 张矢量图、独立种子确认、因果配对设计、证据边界与复核命令均集中在此。
+
+[GitHub 可读结果摘要](https://github.com/Thatht137/quantum.harness/blob/challenge/qmc-atom-reload-surface-code/tracks/qmc/solutions/yanwang-66/RESULTS.md) · [HTML 源文件 / 离线下载](https://github.com/Thatht137/quantum.harness/blob/challenge/qmc-atom-reload-surface-code/tracks/qmc/solutions/yanwang-66/report/report.html) · [机器可读汇总](https://github.com/Thatht137/quantum.harness/blob/challenge/qmc-atom-reload-surface-code/tracks/qmc/solutions/yanwang-66/results/summary.json) · [Challenge #66](https://github.com/QuantumBFS/quantum.harness/issues/66)
+
+## Team yanwang
+
+| | |
+|---|---|
+| **Members** | 王介人、何思成、赵志轩 |
+| **Contributor for #66** | 何思成 |
+| **Track** | Quantum Monte Carlo / circuit simulation |
+| **Challenge** | Addresses #66 — How much does atom reloading help loss-tolerant surface-code memory? |
+
+## Headline result
+
+> **Across 179.2 million discovery cell-shots, active atom-reload policies
+> produce 502 provisional helpful and zero harmful FDR classifications among
+> 1,960 paired comparisons. A separate-seed confirmation reaches the registered
+> precision target in all 32/32 comparisons.**
+
+This is a direct paired test, not a comparison between unrelated simulation
+batches. Every reload policy sees the same counter-addressed Pauli,
+measurement, loss and reload-outcome streams for each shot. The result therefore
+isolates the decision to reload from random trajectory luck.
+
+| Quantity | Verified result |
+|---|---:|
+| Discovery cell-shots | `179,200,000` |
+| Discovery paired comparisons | `1,960` |
+| Helpful / no difference / harmful | `502 / 1,458 / 0` |
+| Independent confirmation cell-shots | `12,800,000` |
+| Confirmation precision | `32 / 32` |
+| Precision fraction | `1.0` (required: `0.8`) |
+| Accepted engineering volume | `192,000,000 cell-shots` |
+| Sealed holdout | `0 / 1`, unspent |
+
+Discovery and confirmation are independent evidence streams; the 192 million
+total describes accepted engineering volume and is not pooled as one
+statistical sample.
+
+## Scientific interpretation
+
+- **Strong parameter-region signal.** After 20,000-resample paired intervals
+  and Benjamini-Hochberg correction over all 1,960 comparisons, 25.6% favor
+  active reload and none are classified harmful at the highest verified
+  Discovery phase.
+- **Independent precision success.** All 32 policy-versus-none comparisons in
+  the independent-seed headline slice meet the frozen interval-width target.
+- **Conservative final gate.** The formal deadline disposition remains
+  `inconclusive_at_deadline`: only 81/2,240 Discovery cells and 4/40
+  confirmation cells reached the deliberately high logical-failure count
+  target.
+- **Claim boundary.** The evidence supports the existence of useful reload
+  regions. It does not identify one policy as universally optimal and does not
+  turn finite `d=3,5` data into an asymptotic threshold claim.
+
+## What is new
+
+1. **Dynamic loss is part of the decoder model.** Missing sites generate
+   round-dependent super-stabilizers and an erasure-aware matching graph; the
+   mask is not merely logged beside an unchanged no-loss decoder.
+2. **Reload policies are compared causally.** Counter-addressed common random
+   numbers preserve external events across `none`, `immediate`, `periodic(R)`
+   and `threshold(theta)`.
+3. **Recovery is auditable.** Checkpoints, append-only shot ranges, atomic
+   publication and SHA-256 manifests prevent preempted Slurm jobs from
+   contaminating accepted evidence.
+4. **Decoder inputs are future-proof.** The versioned data contract separates
+   syndrome + loss/reload history + metadata from the final logical label, so
+   future learned decoders can use the same benchmark without label leakage.
+
+## Why we trust the baseline
+
+1. **Frozen physics.** The round timeline, reload semantics, policy thresholds,
+   noise matrix and stopping rules were fixed before the large grid.
+2. **Independent oracles.** Exhaustive small-geometry graph checks and a
+   separate policy state-machine oracle test the simulator outside its own
+   implementation path.
+3. **Causal isolation.** Prefix-causality and label-poison tests reject access
+   to future loss or final logical outcomes.
+4. **Negative controls.** Wrong-answer, timeout, environment-escape,
+   background-process and hard-coded candidates all fail closed.
+5. **Exact replay.** The accepted candidate reached
+   `185.979924557991` validated decoded shots/s and reproduced immutable shard
+   hashes under the locked environment.
+6. **Independent seed.** The confirmation stream is separate from Discovery
+   and passes its precision gate in every comparison.
+
+## Model and policies
+
+- Rotated surface-code memory at `d={3,5}`, memory-X/Z and `T={d,2d}`.
+- Per-round data Pauli errors, measurement flips and stochastic atom loss.
+- Site state machine:
+  `ACTIVE -> LOST_UNDETECTED -> LOST_DETECTED -> RELOADING -> ACTIVE`.
+- Policies: `none`, `immediate`, `periodic(1|d|2d)` and
+  `threshold(0.02|0.05|0.10)`.
+- Baseline decoder: loss-history-conditioned erasure-aware MWPM.
+- Statistics: paired 95% bootstrap intervals and BH FDR `q=0.05`.
+
+Reload restores a fresh carrier, not the unknown quantum state held by the
+lost atom. This distinction is enforced in the event model and report.
+
+## Included artifacts
+
+| Path | Role |
+|---|---|
+| `report/report.html` | Standalone judge-facing report with embedded vector figures |
+| `report/report.json` | Structured four-section challenge report |
+| `RESULTS.md` | Compact scientific result and claim boundary |
+| `results/summary.json` | Machine-readable public headline numbers and hashes |
+| `results/*/*.parquet` | Verified compact cell/comparison aggregates |
+| `src/reload_qec/` | Simulator, policy, decoder, analysis and artifact pipeline |
+| `tests/` | Core, Discovery, confirmation and sensitivity contracts |
+| `research/reference/` | Independent graph and policy oracles |
+| `research/DATA_SCHEMA.md` | Decoder-ready input/label separation and boundary semantics |
+| `research/MODEL.md` | Frozen physical and timing model |
+| `research/SCIENCE_GATE.md` | Deadline gate audit |
+| `slurm/` | Reproducible SCNet/xh5 batch entry points |
+
+Large per-shot arrays stay on the cluster. Their original artifact names and
+digests remain in the committed analysis manifests; the compact files included
+in this PR have directly runnable subset checksum lists.
+
+## Compact verification
+
+From `tracks/qmc/solutions/yanwang-66` in the locked Python environment:
+
+```bash
+python3 -m pytest \
+  tests/test_experiment_core.py \
+  tests/test_discovery_contract.py \
+  tests/test_confirmation_contract.py
+
+(cd results/discovery-phase-3 && sha256sum -c included-checksums.sha256)
+(cd results/confirmation-phase-5 && sha256sum -c included-checksums.sha256)
+
+python3 ../../../../skills/report/render_report.py report
+```
+
+The report was generated by the repository's `challenge-report` / `report`
+pipeline and is self-contained after rendering.
+
+## Honest boundary
+
+The highest verified Discovery artifact is a post-deadline supplement. It
+strengthens the public signal but does not retroactively change the frozen
+deadline disposition. Cost sensitivity was not started because the registered
+Discovery prerequisite did not complete, and the sealed holdout was not
+queried. No `d=7,9` experiment was run, so no asymptotic threshold is claimed.
+
+The result is nevertheless concrete: it advances Challenge #66 from an open
+engineering question to a reproducible benchmark with broad positive
+finite-size signals, independent precision confirmation, compact public
+aggregates, and a clear route to a decisive future stopping run.

tracks/qmc/solutions/yanwang-66/RESULTS.md

@@ -0,0 +1,73 @@
+# Challenge 66 result record
+
+## Headline
+
+Across the highest verified public artifacts, dynamic reload policies produce
+**502 provisional helpful and zero harmful FDR classifications** among 1,960
+paired policy-versus-none comparisons. An independent-seed confirmation reaches
+the target confidence-interval precision in **32/32 comparisons**.
+
+This is promising evidence for parameter-dependent benefit, not a claim that a
+single policy wins everywhere. The registered cell-level stopping rule remains
+incomplete, so the deadline disposition is `inconclusive_at_deadline`.
+
+## Discovery Phase 3
+
+| Quantity | Value |
+|---|---:|
+| Physical groups | 280 |
+| Cells | 2,240 |
+| Policy-versus-none comparisons | 1,960 |
+| Cumulative cell-shots | 179,200,000 |
+| Helpful, provisional | 502 |
+| No significant difference, provisional | 1,458 |
+| Harmful, provisional | 0 |
+| Cells meeting the failure target | 81 |
+| Lockstep groups completing the target | 4 |
+
+The 20,000-resample paired intervals are corrected across all 1,960
+comparisons with Benjamini-Hochberg `q=0.05`. Phase 3 was verified after the
+deadline and is reported as a supplement; it does not rewrite the frozen
+deadline decision.
+
+## Independent-seed confirmation Phase 5
+
+| Quantity | Value |
+|---|---:|
+| Headline groups | 8 |
+| Cells | 40 |
+| Comparisons | 32 |
+| Cumulative cell-shots | 12,800,000 |
+| Comparisons meeting precision | 32 / 32 |
+| Precision fraction | 1.0 (required: 0.8) |
+| Cells meeting the failure target | 4 / 40 |
+
+The precision gate is a concrete success: every candidate-versus-none
+comparison has reached its registered interval width. The remaining blocker is
+the deliberately conservative count of observed logical failures, not unstable
+confidence intervals.
+
+## What the result supports
+
+- Active reload deserves parameter-region optimization rather than a blanket
+  always/never rule.
+- Common-random-number pairing resolves policy differences efficiently.
+- A round-level erasure-aware surface-code simulator can be made deterministic,
+  causal, restartable and independently auditable on public HPC systems.
+- The decoder-ready interface cleanly separates observation history from the
+  final logical label.
+
+## Claim boundary
+
+- The study does not identify a universally best reload policy.
+- `d=3,5` finite-size data do not establish an asymptotic loss threshold.
+- Cost sensitivity was not run because its registered discovery prerequisite
+  did not finish.
+- The sealed holdout remains unused at `0 / 1`.
+
+The machine-readable source for every number above is `results/summary.json`.
+The two result subdirectories contain the verified aggregate Parquet tables,
+their original analysis summaries and checksum manifests. The original
+`analysis-checksums.sha256` files also name large cluster-only artifacts;
+`included-checksums.sha256` provides a directly runnable check for the compact
+files committed in this PR.

tracks/qmc/solutions/yanwang-66/environment/requirements.lock

@@ -0,0 +1,58 @@
+stim==1.15.0 \
+    --hash=sha256:f0fb249f1a2897a22cbe4e0c2627abf49188cbbf19b942d4749972d1c3bdf12c
+pymatching==2.2.2 \
+    --hash=sha256:166abeb45525cbbcb656e66e4400ff24a11b2cf29495dc7b213885809f45d237
+numpy==2.1.3 \
+    --hash=sha256:bc6f24b3d1ecc1eebfbf5d6051faa49af40b03be1aaa781ebdadcbc090b4539b
+scipy==1.14.1 \
+    --hash=sha256:fef8c87f8abfb884dac04e97824b61299880c43f4ce675dd2cbeadd3c9b466d2
+pandas==2.2.3 \
+    --hash=sha256:c124333816c3a9b03fbeef3a9f230ba9a737e9e5bb4060aa2107a86cc0a497fc
+pyarrow==18.1.0 \
+    --hash=sha256:e31e9417ba9c42627574bdbfeada7217ad8a4cbbe45b9d6bdd4b62abbca4c6f6
+matplotlib==3.9.3 \
+    --hash=sha256:d3c93796b44fa111049b88a24105e947f03c01966b5c0cc782e2ee3887b790a3
+pytest==8.3.4 \
+    --hash=sha256:50e16d954148559c9a74109af1eaf0c945ba2d8f30f0a3d3335edde19788b6f6
+psutil==6.1.1 \
+    --hash=sha256:97f7cb9921fbec4904f522d972f0c0e1f4fabbdd4e0287813b21215074a0f160
+jsonschema==4.23.0 \
+    --hash=sha256:fbadb6f8b144a8f8cf9f0b89ba94501d143e50411a1278633f56a7acf7fd5566
+attrs==26.1.0 \
+    --hash=sha256:c647aa4a12dfbad9333ca4e71fe62ddc36f4e63b2d260a37a8b83d2f043ac309
+contourpy==1.3.2 \
+    --hash=sha256:3859783aefa2b8355697f16642695a5b9792e7a46ab86da1118a4a23a51a33d7
+cycler==0.12.1 \
+    --hash=sha256:85cef7cff222d8644161529808465972e51340599459b8ac3ccbac5a854e0d30
+fonttools==4.63.0 \
+    --hash=sha256:d76ac49f929aecaf82d83250b8347e099d7aecba0f4726c1d9b6df3b8bb5fe18
+iniconfig==2.3.0 \
+    --hash=sha256:f631c04d2c48c52b84d0d0549c99ff3859c98df65b3101406327ecc7d53fbf12
+jsonschema-specifications==2025.9.1 \
+    --hash=sha256:98802fee3a11ee76ecaca44429fda8a41bff98b00a0f2838151b113f210cc6fe
+kiwisolver==1.5.0 \
+    --hash=sha256:2517e24d7315eb51c10664cdb865195df38ab74456c677df67bb47f12d088a27
+networkx==3.6.1 \
+    --hash=sha256:d47fbf302e7d9cbbb9e2555a0d267983d2aa476bac30e90dfbe5669bd57f3762
+packaging==26.2 \
+    --hash=sha256:5fc45236b9446107ff2415ce77c807cee2862cb6fac22b8a73826d0693b0980e
+pillow==12.2.0 \
+    --hash=sha256:8bd7903a5f2a4545f6fd5935c90058b89d30045568985a71c79f5fd6edf9b91e
+pluggy==1.6.0 \
+    --hash=sha256:e920276dd6813095e9377c0bc5566d94c932c33b27a3e3945d8389c374dd4746
+pyparsing==3.3.2 \
+    --hash=sha256:850ba148bd908d7e2411587e247a1e4f0327839c40e2e5e6d05a007ecc69911d
+python-dateutil==2.9.0.post0 \
+    --hash=sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427
+pytz==2026.3.post1 \
+    --hash=sha256:dd95840dd199baea12d9cc096a1d452caa6596a1c1e4b5f3dbd1541855d5e815
+referencing==0.37.0 \
+    --hash=sha256:381329a9f99628c9069361716891d34ad94af76e461dcb0335825aecc7692231
+rpds-py==2026.6.3 \
+    --hash=sha256:9c1255b302953c86a486b81d330d5ee1d5bd937691ce271b6be0ef0e299eaab7
+six==1.17.0 \
+    --hash=sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274
+tzdata==2026.3 \
+    --hash=sha256:dc096730c87af6cab1b171c9d532be840741ff5d459015e7f6947bd7d7e54931
+typing-extensions==4.12.2 \
+    --hash=sha256:04e5ca0351e0f3f85c6853954072df659d0d13fac324d0072316b67d7794700d

tracks/qmc/solutions/yanwang-66/pyproject.toml

@@ -0,0 +1,26 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "reload-qec"
+version = "0.1.0"
+description = "Challenge 66 dynamic atom-reload surface-code benchmark"
+requires-python = ">=3.11"
+dependencies = [
+    "numpy==2.1.3",
+    "pandas==2.2.3",
+    "pyarrow==18.1.0",
+    "pymatching==2.2.2",
+    "scipy==1.14.1",
+    "stim==1.15.0",
+]
+
+[tool.setuptools]
+package-dir = {"" = "src"}
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]