diff --git a/docs/claude/stage-decision-record-S2-2026-11-15.md b/docs/claude/stage-decision-record-S2-2026-11-15.md index 68d0e9d4e..a0028abe1 100644 --- a/docs/claude/stage-decision-record-S2-2026-11-15.md +++ b/docs/claude/stage-decision-record-S2-2026-11-15.md @@ -92,3 +92,16 @@ S3-A runs as approved. In parallel, the route-A falsification tests (O-vol oracl - Process variation, which Greif v5 cannot provide. - Acceptance outputs agreed with Yanqi before any model comparison. - O1 and any flow-competition prior designed against those targets, with a CAE-provided pressure solve available as a reference instead of our own graph solver. + +## 9. Reproduction and factor stage (2026-11-16/17) +- **Gre24 reproduction:** + - v4 (the paper's own data), median RRMSE under max(y): GBRT 0.0396, FNN 0.0452, against the paper's 0.0338 / 0.0339; GBRT's fraction below 0.1 is 94.9 % against the paper's 95.5 %. + - v5 gives nearly the same summaries. + - Close but not strict: the paper's scoring code and several details are unavailable. Independent audit by MeltAssist, `271c47a6`. +- **R4(a), shared element speed:** EIK 9.56 % vs GRAPH 9.86 % (-3.1 %, CI excludes 0). The historical B0 is 10.38 %, attributed only to the whole distance construction. +- **R4(b)/(c), INTERIM_100 (552/204):** + - Training-free EIK 9.59 %; all learned arms 7.00 to 7.43 %. + - Each single factor (t_set as input, geometry construction, query weights, all nodes, learner) changes the error by about 0.1 to 0.2 points. + - The faithful Gre24 GBRT (A, 7.08 %) is about the same as the historical R2 (7.30 %). + - **Reading:** the distance-feature GBRT family has plateaued around 7 % on our grouped protocol. Further gains need new information, not more tuning. H3 (the lowest point estimate) is not adopted as a base, to avoid selecting on the test set. +- **Decision:** the next experiment is the minimal O1 test (MoldAstra specification 14Drirmj…; arms B/T/G/C; T vs G primary; 50 % pool; delta = 1 %) on the frozen A base, chosen for fidelity rather than score. Execution contract by MeltAstra; it runs after the R4bc queue finishes.