diff --git a/docs/claude/stage-decision-record-S2-2026-11-15.md b/docs/claude/stage-decision-record-S2-2026-11-15.md new file mode 100644 index 000000000..befede01e --- /dev/null +++ b/docs/claude/stage-decision-record-S2-2026-11-15.md @@ -0,0 +1,33 @@ +# S2 stage decision record: what is established, what is not, and routes beyond "tune the network / add features" (Claude, 2026-11-15) + +Inputs: S2 custody result 1dea89d6, post-hoc D1-D4 daed8e87 (with Mold's final reading), the S3-A contract b0390231, the S1 result, and the design record. Grok's reading of the original papers and Mold's independent O1/scale/learning-object review will be added here when they arrive. Everything below that is not marked established is a hypothesis. + +## 1. Established (617 parts / 209 components, Greif v5, automatic process, one gate per part) +1. A Gre24-style distance-feature GBRT (R2, graph distances, not eikonal) is the strongest model tested: component-mean relative L2 of 0.073. +2. The neural route as configured (Transolver, width 128 chosen for feasibility, 42 epochs, direct t* target) reaches 0.108. That fails the pre-registered usefulness gate against R2 (CI entirely negative) and cannot be told apart from the naive B0 (0.104; the interval includes 0). +3. Inside that backbone the 4 edge-direction readings are harmful (Delta = -1.29 %, CI [-2.12, -0.56] %). +4. N's gap to R2 is mostly in shape: even a per-part oracle multiplier leaves H = 0.0137, above R2's total of 0.0073. +5. B0 = V*_f x t_set, a volume-fraction prior along graph distance, has the best last-filled location of all four arms (0.34 Rg), even though R2 wins on L2. +6. N improves with more data (-15.9 % from 25 to 100 %) and is still improving slightly at 42 epochs. + +## 2. Not established, or possibly hidden by how we have been framing the work +- **Irreducible error is unknown.** The labels resolve about 1 % of volume per solver step, and the 2 % non-identifiability stop was withdrawn. How much of R2's 7.3 % can be removed at all has not been measured. +- **The primary metric may not match practical use.** Relative L2 on arrival time weights late, large-T nodes. The model that is best on L2 (R2) is not best on last-filled location (B0). What a molding engineer needs (fill pattern, last-filled region, weld lines, short-shot risk, required pressure) was never written down as the acceptance target. +- **The learning object.** All arms regress T(x) pointwise. Under an approximately constant volumetric flow rate, T(x) = t_fill x (fraction of the cavity volume that fills before x). Time would then be determined almost entirely by the *filling order*, with the scale fixed by physics. B0 is exactly this with graph-distance order, which would explain why it is competitive and why it locates the last-filled region well. **Whether the automatic process in Greif v5 is close to constant flow has not been checked.** +- **No physics in the distance.** R2's graph distance ignores thickness, yet the flow conductance of a thin cavity scales like h^3 (Hele-Shaw). A thickness-weighted (conductance) path distance is a geometry-only quantity we have never tried. +- **Scale of the data.** Gre24-style feature models are strong at a few hundred geometries; whole-part neural operators usually rely on far more data. S2 does not tell us whether the neural route fails here because of the data size, the recipe, or something fundamental. +- O1 (type or region organisation with interfaces) is still untested; nothing in S1 or S2 bears on it. + +## 3. Routes beyond tuning the network or adding features +A. **Change the learning object to filling order, with the time from a volume map (new; cheapest to falsify).** Predict an ordering potential phi(x). Map it to time through the volume cumulative: T_hat(x) = t_fill x F_V(phi(x)), where F_V(phi) is the volume fraction with a smaller phi. This is monotone and volume-consistent by construction, so errors become pure ordering errors. + - Minimal falsifying comparisons, using only existing predictions and the label cache (post-hoc, CPU). The S2 query draws are mass-proportional, so the empirical distribution of the draws estimates F_V over the eligible volume. + - (O-vol oracle) T_hat = c x F_V(rank of the true T). This tests the constant-flow time map itself. If the oracle relative L2 is not well below 7.3 %, route A is falsified as a full-time model: the flow-rate profile matters and the map must be learned. Report c = T_end/t_set fixed from training parts, and the true-T_end variant separately as an oracle. + - (Reorder test) T_hat = c x F_V(rank of R2's prediction), and the same for N and B0. If re-mapping R2's own ordering through the volume map beats R2, the volume-consistent object is better than direct regression without any training. +B. **Physics-structured distance (new).** Compute a conductance-weighted path distance on the tet graph from the existing exact wall distance d(x) (edge cost proportional to 1/h(x)^2 or 1/h^3, with h = 2d as a proxy, chosen before the run). Test it as a one-factor addition to R2 and as the ordering potential for route A, so that geometry carries the flow physics instead of the model having to learn it. +C. **Hybrid residual network (S3-B, already proposed).** The network corrects R2's prediction, or under route A, the ordering potential. Gate: it must beat R2 by the margin. +D. **Settle the acceptance target with Yanqi.** Agree which outputs matter in practice (last-filled region, weld-line or front-meeting locations, short-shot or pressure risk, full-field T) before choosing between models that rank differently on different metrics. +E. **Data and process (in-house CAE).** Public data cannot test process, gate or material interventions; the bypass pair and process sweeps remain the controlled extension. +F. **O1** with a matched generic partition and a control cut, unchanged as the next design item. + +## 4. Order +S3-A runs as approved. In parallel, the route-A falsification tests (O-vol oracle and reorder test) are cheap post-hoc CPU work on existing artifacts and are proposed next, together with the route-B distance definition, which must be fixed before any run. Route D needs Yanqi's input and should be asked now, because it decides what "useful" means for every later comparison.