diff --git a/docs/claude/stage-review-S1-to-S2-2026-11-14.md b/docs/claude/stage-review-S1-to-S2-2026-11-14.md new file mode 100644 index 000000000..5782a66ba --- /dev/null +++ b/docs/claude/stage-review-S1-to-S2-2026-11-14.md @@ -0,0 +1,61 @@ +# Stage review: S1 → S2 (Claude, 2026-11-14) + +This is the habit agreed with Yanqi: summarise, revisit the original intent and the prior work, discuss, then advance. This is the integrating draft. MeltAstra (execution differences), MoldAstra (mathematics and grouping boundaries) and Grok (targeted gaps only) add to it, and I fold their additions in. + +## 1. Original intent (Yanqi, 2026-11-09…13) +The team is developing injection-molding software and wants a surrogate model that is practically useful. Splitting by local geometry type was proposed *to simplify the input and improve generality*. It was not meant to compress information or to speed up inference. Hele-Shaw reduction is not the target; full 3D labels are. Work starts on public data, and the in-house CAE comes later for controlled extensions. Both geometric routes get a full attempt, and neither is presented as a proven advantage. + +## 2. What S1 supports and what it does not +S1 = model-free input coverage on a 24-source LOPO panel. Results are in codex/s1-real-scoring 90b8c66. + +**Supported, within S1's scope:** +- The O2 representation (v1.1a) is computable at real-part scale. It is numerically verified: bit-exact sampled-vs-full gate, Type A rigid-motion invariance, 0 invalid. +- The edge-direction block carries slightly more source-transferable joint structure than a bound-preserving permutation (about −0.9 %, interval excludes 0). + +**Not supported:** +- That true O2 node invariants make unseen-source inputs more familiar. The paired difference is +5 %, i.e. *behind*. + +**Not tested at all:** +- prediction error, sample efficiency, partitions/interfaces (O1), connectivity, other materials or processes, remesh stability on real parts. + +**Not a valid conclusion:** +- That the edge advantage against the plain shuffle (−19 %) is real structure. 86 % of plain-shuffle slots are algebraically impossible. +- Anything about the topology/partition route. + +## 3. Against prior work (what we still lack) +| Prior work | What it already shows | Our gap | +|---|---|---| +| Gre24 (eikonal + GBRT, Greif corpus) | Distance features predict fill time across geometries | No R2 baseline yet on our grouped split. S2 must include it, or any neural claim is untested. | +| Lee26 (GAT, plate family) | Generalisation within a family; cross-family pressure error rises 11.9 → 36.1 % (confounded) | Our split is grouped by source plus near-duplicates; the near-duplicate rule is still pending. | +| Transolver / GINO / RIGNO | Strong whole-part and region models; RIGNO has data-efficiency curves | We have no whole-part learning curve yet; the N arm is the first one. | +| Huang SNI / NEST | Local operators + Schwarz on new geometry (other physics) | O1 has no running test yet. | +| Cadmould AI Solver (industry) | A neural filling solver trained on very large proprietary data | Our lever is sample efficiency; that needs S2 learning curves. | + +## 4. Why O1 (type/topology partitioning) stays open +- S1 changed no partitions or interfaces. Arm C kept only per-node edge mean/max. +- O1's hypothesis is about *aggregation and interface exchange in connected neighbourhoods organised by type*. Nothing has tested it yet. +- Its infrastructure is defined: + - O1-v1 admission; + - exact capacity LP with a basic solution; + - partition acceptance tests. +- What is still missing: + - an O1 type definition built on the validated descriptors; + - region tables beyond 3 parts. +- MoldAstra's bypass pair (constant Δp gives T_closed − T_open = A·R·L/(2Δp); fixed flow gives none) is the cleanest discriminating test of connectivity versus type. It needs paired solves from the in-house CAE. + +## 5. S2: smallest identifiable question (design v0, d8aa0f42) +Do edge-direction readings reduce arrival-time error, or the number of training sources needed, on unseen source groups? + +Same whole-part backbone and the same g12/g23 in both arms, with R2 in the same table. + +**Entry conditions:** +1. Near-duplicate rule and frozen grouping (MoldAstra). +2. d/t_part over the full corpus, and its cost. +3. Frozen margin and seed table. +4. A pilot measuring t_step and memory. +5. Yanqi's explicit go for training. + +## 6. Points for discussion +- Should S2 also include an *O2-free* N arm (without g12/g23), so that the whole O2 feature block can be judged against R2? It costs one more arm. +- When does O1 get its own S2 arm, and what is the minimal O1 type definition to freeze first? +- Is single-material/process Greif data enough for the research claim, or should the in-house CAE pilot (bypass pair plus gate edits) start in parallel?