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 be0fe9604..d53b1c813 100644 --- a/docs/claude/stage-decision-record-S2-2026-11-15.md +++ b/docs/claude/stage-decision-record-S2-2026-11-15.md @@ -50,3 +50,24 @@ S3-A runs as approved. In parallel, the route-A falsification tests (O-vol oracl - If positive unfilled volume remains with no reachable front, record it as blocked. Do not drop the d = 0 mass and do not declare the part filled. - This yields a conservative graph-ordering prior. To call it a Hele-Shaw discretisation, the edge conductance must state how it includes face scale, length, thickness and viscosity, and what it means under remeshing. A truly closed cavity with trapped air needs a gas-pressure or venting model. - With one gate, linear conductance and a uniform front pressure, scaling Q only changes the clock. The new, testable information is how flow splits between branches, and the resulting order. + +## 7. Stage close (after S3-A, MeltMind 1e49c157, Mold reading) +**Constrained (tested, negative or no gain, within the limits stated):** +- O2 local-tensor readings as fill-time predictors on Greif v5. S1: no coverage support for the node invariants. S2: the edge readings were harmful in the neural backbone (-1.29 %). S3-A: no meaningful gain inside R2 (E4 +0.36 %, CI [-0.18, +0.90] %; G2 +0.10 %, descriptive). The line is closed for this target and corpus. +- The neural whole-part backbone as configured (width 128, 42 epochs, direct t*). It fails the gate against R2, the gap is mostly shape (D3), and longer training or scale calibration will not close it (D2/D3). +- A stronger distance weighting f = D_w^2 (20.2 %, against 9.1 % for D_w^1), for this graph approximation and clock only. +- Volume-consistent remapping of R2: no net gain (-0.09 %, CI includes 0). + +**Measured references (oracles; not deployable, not floors):** true full-domain order plus training clock gives 1.48 %. The LOG time map is within 1 pp of linear for 411/624 parts, with a tail up to 34 pp. Part 996 shows that the last-filled timing can sit in the post-switch-over phase. + +**Untested:** O1 (region or branch organisation with interfaces); flow-competition (branch-split) ordering; continuous eikonal versus Dijkstra; process, gate and material variation (impossible on public data); acceptance targets (asked through TRON, not yet answered); whether neural models do better with far more data. + +**Next route: the filling order from flow competition, S3-B(a) then (b).** CPU only, no training without a separate contract. +- (a) A graph filling prior (Mold's definition; conductance g_ij = dual-volume/L_ij^2 x D_w^2, fixed before the run). Scored like the naive batch against B0_DW1 and R2. +- (b) Only if (a) passes: its order as a single extra R2 feature, using the six-region table against R2. + +**Stop conditions (fixed now):** +1. (a) is ill-posed in practice: "blocked" or unfinished on more than 5 % of the 617 parts, or more than 5 % of the volume. Report and stop the line. +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.