Files
MeltMind/docs/claude/design-state-2026-11-14.md
T
ClaudeandClaude Opus 5.5 628a10bad3 docs: Claude design-state record and Greif v5 census tools
Consolidated MeltMind design decisions, S1 result review and Drive index;
stdlib-only ZIP range-read census (624 logs) with SHA256SUMS. Review
material only; not for main.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-14 08:52:20 +00:00

7.7 KiB

MeltMind design state: consolidated record (Claude, 2026-11-14)

Sources: Slack C0C78UFKGRM thread 1794219487.255049 (2026-11-09 to 2026-11-14), the MeltMind codex/* review branches named below, and the Drive documents listed in README.md. This file summarises; where it differs from a frozen source, the source wins.

0. Goal and scope

  • Goal: a practically useful filling surrogate for unseen parts. Geometry organisation is a means toward cross-part generalisation and sample efficiency, not a goal in itself. Inference speed is secondary.
  • Labels: 3D Moldflow labels (Greif v5). Physical reduction (Hele-Shaw) is not the target.
  • Two exploratory candidates:
    • O1: type-aligned grouping/aggregation.
    • O2: local geometric tensor representation, the continuous form of "local frames".
  • Other routes, kept as references and baselines:
    • R1: strong whole-part models;
    • R2: Gre24-style distance/eikonal features with GBRT;
    • R3: in-house 2.5D solve plus learned 3D correction;
    • R4: explicit global pressure coupling.
  • Topology/local-type partitioning (O1) remains an independent candidate. The S1 node-representation result does not count against it.
  • Public data limits: Greif v5 has one material, one random gate per part and an automatic process. It cannot identify gate, process or material interventions, short shots, or reconnection effects; those wait for the in-house CAE.

1. Data contract (Greif v5 on Horsea)

  • data.zip is 32,757,812,726 bytes, official MD5 verified. 624 parts; 249 filename source keys (a leak-prevention key, not verified independent CAD families).
  • Geometry comes from the VTU tet subdomain.
    • 617 parts are pure tet. 7 are mixed (triangles plus tets): the triangles carry no field values, are not a coincident surface, and in 00001048 form a separate second body. They are kept as an auxiliary component, flagged mixed_topology, and are not part of the d(x) wall.
  • Units, by numerical cross-check only: Points in mm, Fill_time in s, Pressure in Pa, Temperature in K. Velocity unit unknown, so velocity is unused.
  • Frames are joined by physical time, never by block number. Masks are F (Fill_time <= t), V (finite) and M = F AND V. No nan_to_num. Missing-value age varies, so no fixed frame lag applies.
  • First target: static Fill_time on tet nodes.

2. Descriptors

  • d(x): exact distance to the tet-subdomain boundary faces. Validated analytically (box, plate, rod) and against brute force. It is exact only for the given triangulation; 2d is not thickness.
  • t_part = 4 x the lumped-mass-weighted median of d over all tet nodes, including d = 0. This is a scale proxy. R = 2.5 t_part.
  • O2-A (analytic rod): the in-plane split shrinks at first order with transverse refinement and flips sign with a mirrored tet split. This is directional discretisation bias that converges. It is not established on real parts under remeshing.
  • E6 (three real parts): a four-point kernel compared with the old 3-hop field. Part 778's near-cut share rose from 25.9 % to 40.05 %. Hard type labels are therefore ill-posed near degenerate spectra.

3. O2 representation (frozen O2-rep v1.1a, MoldAstra spec; implemented at codex/o2-rep-v1-1a f2e5ed0)

  • mu = lambda+/tau+; g12 = mu1 - mu2, g23 = mu2 - mu3.
  • B_k = (mu_k - mu_{k+1}) (P_1 + ... + P_k). Continuous at repeated eigenvalues and adds no information beyond M/tau.
  • Edge and gate readings q(u) = (u^T B1 u, u^T B2 u). Edge readings are mesh-edge features (rigid-motion invariant, not remesh invariant).
  • FP64 constants:
    • tau_min = 1e-10;
    • eps_psd = 1e-12 tau0 (clip small negative eigenvalues, then renormalise);
    • edge length tolerance 1e-12 L_med;
    • gate tolerance 1e-6 L_med.
  • clipped is an audit flag only, never a model input.

4. O1 (type grouping) status: not frozen

  • Admission rule O1-v1:
    • J = all regions after V/512 cleaning;
    • K = max(ceil(V/min V_j), 8);
    • admit only if K <= 512 and N/K >= 200;
    • mixed mode, no region quotas.
  • The screen used the old 3-hop type field and covered 3 parts only. One part was admitted (666); 778 and 850 failed on nodes per group.
  • Decided order:
    1. O1 type definition and any merge/admission change, frozen after E6;
    2. calibrate every stage on part 890 plus one uncached part;
    3. checkpointed region tables for a panel;
    4. exact capacity LP (basic solution, identical in every arm) on admitted parts.
  • O1-v2 (tying cleaning to the resolution floor) was not adopted.
  • Partition matching must be verified, not assumed: K, connectivity, volume within the tolerance, and pool-level shape and boundary-fraction tests.

5. S1: model-free coverage diagnostic (M2 v1.2 sampled)

  • Panel: 24 parts, one per source group, from node-count tertiles (seed 20261114). Folds: LOPO-24. Commits: codex/s1-panel-precheck 81f56af, codex/s1-sampled-draw 94011fd.
  • Sampling: 2,000 mass-PPS draws per part, with replacement (seed [2026111400, prefix]), on d > 0 nodes.
    • Mid-rank weights (1/P) c E / K_pE.
    • k = 8 distinct physical neighbours; D_p = RMS over p.
    • Ratio estimators over draw counts.
  • Arms:
    • A: d/t, gate distance / R_g.
    • B: A plus g12, g23, gate q1, gate q2.
    • C: B plus edge mean/max of q1 and q2.
    • Controls: A+S (whole-block shuffle of the B columns), B + plain edge shuffle, B + zero-pattern-stratified normalised shuffle (secondary, not a gate).
  • Seeds: permutations 1401-1405; query bootstrap 1406 (200 replicates); source bootstrap 1407 (2,000 replicates).
  • Decision: the five-seed mean source-level interval. That interval is a source-reweighting interval over fixed LOPO paired differences, not a coverage guarantee for new sources.
  • Result (codex/s1-real-scoring 90b8c66; five-seed means, rank-distance units):
Contrast Difference Interval Reading
B - (A+S), whole +0.0026 [+0.0014, +0.0041] behind (about +5 % of A+S)
B - (A+S), near-wall +0.0017 [+0.0004, +0.0032] behind
C - plain edge shuffle -0.0150 [-0.0162, -0.0138] ahead, but 86 % of plain-shuffle slots break the algebraic bounds
C - stratified shuffle (secondary) -0.00055 [-0.0009, -0.0002] ahead, about 1 %

All five seeds point the same way for every contrast. Cross-dimension contrasts are descriptive only.

  • Reading:
    • No support for O2 node invariants on input coverage.
    • The edge-direction readings carry a small amount of extra source-transferable joint structure.
    • S1 ranks candidates and does not refute them. Nothing here concerns prediction.

6. Rules carried forward

  • One factor per comparison; identical descriptors in every arm; a matched generic partition plus a control cut for O1; cost-accounted distance baselines are always reported.
  • Decision table for S2 learning curves. Margin m is a candidate 10 %. With interval [L, U]:
Region Outcome
U < 0 harmful
L <= 0 <= U < m no meaningful gain
L <= 0, U >= m insufficient evidence
0 < L, U < m positive but below margin
0 < L <= m <= U positive, size undetermined
L > m meaningful gain

Single-seed results are exploratory only.

  • Process:
    • fixed-commit packages;
    • supervision of 30 min / 1.25 GiB per field;
    • the 5e8 visit count is an alarm only;
    • over-budget parts are listed, never silently replaced;
    • no post-hoc tuning of thresholds;
    • no training without an explicit go from Yanqi.

7. Open (as of 2026-11-14 16:30 CST)

  • Yanqi's decision on O2 continuation and S2 scope. Recommendation: build the S2 baseline stack (R1 + B0/R2) first, and keep only the O2 edge readings as one feature group.
  • O1 type definition and region-table preparation (see §4).
  • Velocity units and the export path have no written source.