From 74723a149421b96da87cfa08051c84c9e9d37341 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 15 Nov 2026 14:05:17 +0000 Subject: [PATCH] Decision record s8: S3-B(a) stop (splu resource, GMRES numerical) and public-data stage close Co-Authored-By: Claude Opus 5.5 --- .../stage-decision-record-S2-2026-11-15.md | 21 +++++++++++++++++++ 1 file changed, 21 insertions(+) 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 aedea53ec..68d0e9d4e 100644 --- a/docs/claude/stage-decision-record-S2-2026-11-15.md +++ b/docs/claude/stage-decision-record-S2-2026-11-15.md @@ -71,3 +71,24 @@ S3-A runs as approved. In parallel, the route-A falsification tests (O-vol oracl 2. (a) does not beat B0_DW1 on both component-mean relative L2 (paired CI upper bound below 0 in the B0_DW1-to-(a) direction counts as "not better") and Kendall tau-b. Stop the physics-prior line; do not tune the conductance. 3. (b) lands in "no meaningful gain" or worse. The cheap physics-prior feature route ends. 4. If 2 or 3 triggers, the public-data stage for fill time is closed: cheap geometric and physics priors at this data scale do not improve on R2. Effort then moves to the in-house CAE stage (gate-selection pairs, process variation, the acceptance targets), with O1 designed against those targets rather than against Greif v5 fill time. + +## 8. S3-B(a) end and public-data stage close (2026-11-15 evening) +**What happened (MeltAstra receipts, last 4085 stop):** +- Exact event solver (Jacobi-CG): 1130/1324/842 completed; 367 hit the 1800 s wall limit with about 91 % filled. The cost gate failed. +- v0.4 batched (0.5 % M + early re-solve at dead ends), candidate 1 splu: 3 parts passed, then 367 hit the wall limit during factorisation (resource stop). +- v0.4 candidate 2 GMRES + spilu: 1130 and 1324 passed, then 842 failed the frozen post-solve check (flux imbalance 1.013e-8 > 1e-8). That is a NUMERICAL_FAILURE, not resource. The 7 later parts were never attempted, so no solver was selected. +- Pressure calls are dominated by early dead-end re-solves (for example 2807 of 2808 on 1324, in 10272 sub-events, so about 27 % of sub-events). Each new system pays a full factorisation or ILU build. +- No tau gate, no full-617 census, no label scoring was run. Worker CPU used: 6891 of 72000 s. + +**Decision (pre-ruled before the outcome, applied as written):** S3-B(a) ends under stop condition 1. The causes are recorded separately: splu resource, GMRES numerical. No v0.5, no change to Delta V or the routing, no third solver, no loosened tolerance. **This is a feasibility result for this graph-prior implementation, not a statistical test of flow-competition ordering, which remains untested.** + +**Public-data stage for fill time: closed as a resource decision.** +- Every cheap route tried on Greif v5 fill time either failed to improve on R2 (O2 features in S1/S2/S3-A; f = D_w^2; volume remapping) or could not be completed within its budget (S3-B(a)). +- R2 (Gre24-style graph-distance GBRT, 7.3 % component-mean relative L2) stays the working public baseline. +- Not tested, and not counted as failed: O1 with matched partitions; flow competition by any solver-feasible method; continuous eikonal; much larger data; process, gate and material variation. + +**Next (for Yanqi's direction, via TRON):** the in-house CAE stage, designed against the stated industrial goal: predicting new parts from few training sources, with gate selection kept as an application to explore. +- Paired gate-selection runs (Mold's two-gate test), so that gate ranking can be tested directly. +- 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.