Files
MeltMind/docs/claude/stage-review-SR-20261119-2137.md
T

8.3 KiB
Raw Blame History

Stage review SR-20261119-2137 (Claude chairs; requested by Yanqi)

Review and reading only. No new compute is implied by anything here.

Factual cutoff

Facts as of MeltSol's publication of the formal H2-P paired result, or 23:00 CST 2026-11-19 if that is not published by then (then H2-P is reported as "points only, formal verdict pending"). Runs or results after the cutoff are not used.

Facts at cutoff (details: research-tree.md):

  • Greif grouped protocol (552 parts / 204 families, 5 folds): H2 field 7.08 → 6.58 % vs A, family-robust, small seed spread; last-fill location worse; 34 shared-miss parts, half also missed by A, not explained by tail-geometry attributes. Pruned: c, b, 1b, B-series. Track A (random splits): H2 3.25 % vs A 3.99 %, no paper-level claim.
  • ShapeNet-611 (GINO protocol, paper package): GINO reproduced 7.43 % (3 seeds; paper 7.12 %). H2-P (untuned pressure head, matched budget) 10.01 % (3 seeds), points only until the formal paired verdict.
  • Public SOTA context (Grok): only GINO's Table 2 is verified same-protocol; OTNO 6.70 % and PCNO 7.04 % undetermined.
  • Data: Moldflow Adviser route closed for per-node labels; public Greif–Meyer v4 has LOG/YML; 196-case warpage set (packing removed) under metadata check.

Questions

  1. What does H2 do in 3D? Which ingredients carry over (representation, graph, inputs, training) and which must change or be retrained?
  2. What does Greif actually establish, including the last-fill and independent-family limits?
  3. What does the matched ShapeNet result establish, and not establish, about SOTA?
  4. What follows: is GINO-Greif the right next test? Which benchmarks would genuinely probe generality?
  5. Each: prioritised next tests with a stop criterion.

Roster (independent, at most one page, without reading others first)

Member Status Scope
MeltSol active Executor view; publishes the formal H2-P paired result first (cutoff input); answers 1-5
MeltAssist active Reviewer view: what the reviewed evidence supports and does not; 2, 3, 5
MoldSol active Maths view: 1, 4 (what transfers, what a generality benchmark must contain)
NewMeltAstra reserve (Astra, save quota) Half page, fresh view on 3 and 4
Grok active Public literature: 3 and 4 (benchmarks with independent families / process variation; same-protocol context), sources marked read vs inferred
MeltWin stood down One paragraph: Moldflow route findings only
TRON coordinator Process and routing only
Undermind tool, not a participant No account; may be used for literature by Grok/MeltSol with public queries only (no internal details)
Claude chair Writes own account after collecting the others, then chairs one written discussion round on disagreements only

Process

  1. Independent reports → MeltSol publishes all to Git (Grok's from Slack), links posted.
  2. Discussion checkpoint: Claude lists agreements and disagreements; members answer only the disagreements they are named in (one short reply each).
  3. Claude's synthesis: agreements, unresolved disagreements, prioritised next tests with stop criteria; plain-language summary for Yanqi plus links to each contribution.

Discussion checkpoint and draft synthesis (Claude, after all reports; bundle at 605f2987 + MoldSol / NewMeltAstra appended)

Reports: MeltSol, MeltAssist, MoldSol, NewMeltAstra, Grok, MeltWin (stood down), TRON (process), Undermind (tool-assisted literature note). Factual cutoff: H2-P formal paired result, 21:46:45 CST.

Agreements (all reports consistent)

  1. What H2 is in 3D. A mesh message-passing architecture (local messages on real mesh edges, mass-weighted aggregation, per-component global tokens, per-node decoder) that can be re-instantiated on tetrahedral volumes (Greif) and triangle surfaces (ShapeNet). Input layer, normalisation, backbone and head were retrained from scratch for ShapeNet; no weight or fill-prior transfer has been tested.
  2. What Greif establishes. On our grouped protocol (552 parts / 204 families, 5 folds) H2 lowers family-weighted field error 7.08 → 6.58 / 6.59 % (two seeds), in every fold and robust to leaving out any family. Hard last-fill location is worse (.377 → .457 / .408 Rg). 34 shared misses (half also missed by A) are unexplained by tail-geometry attributes. Not established: new families from another source, process or gate interventions, any last-fill benefit, any paper-level comparison.
  3. What ShapeNet establishes. GINO reproduced on its own protocol (7.43 %, 3 seeds; paper 7.12 %). The untuned H2-P is clearly worse: +2.58 pp, paired CI [+2.33, +2.85]. Only GINO's Table 2 is a verified same-protocol public number; OTNO / PCNO stay context only: their author code de-normalises pressure, but the ordered case IDs, query layout and checkpoint/seed rule are not verified as matching. No SOTA claim is possible or made.
  4. Next test. GINO-Greif (matched protocol, 2 folds × 2 seeds, 212 parts / 78 families, paired GINO − H2) is unanimously first. Its reviews have passed (combined pre-score + custody delta fbaef6e2). Pending Yanqi's decision whether to launch; before execution MeltSol adopts the exact PASS per the frozen contract, binds the formal ShapeNet ruling and the live phase-3 budget, and passes the card-admission gate. Both arms predict the residual on the same FP64 volume clock (MoldSol's question).
  5. Generality needs data we do not have. A real benchmark needs independent families with an official family split, explicit and varying gates, material/process variation, per-node fill time and a public licence. We have not found a verified, fully matching public injection-molding benchmark (Grok; not proof that none exists). Best available probes: family holdouts from ABC chunk hashes in v4/v5, and the 196-case warpage set if its families are new and its fill phase is shown to match.

Points resolved in this checkpoint (no further round needed)

  • Packing removal and the fill phase (Grok: should not change fill, inferred; TRON/MeltSol: switchover precedes full fill, so late fill can change). Not a matter of opinion: the pre-registered rule stands (case counts as same-process only if fill completes and switchover ≥ 99 %, post-switchover tail reported separately); the logs decide.
  • Remeshing variability (NewMeltAstra): adopted for future intervention benchmarks: a design ranking is declared unjudgeable where the true gate/process effect is smaller than the remeshing variability of the reference solver.
  • Grid mapping / near-wall aliasing (MoldSol): this round reports only the existing fixed binary near-wall per-part relative L2 already in the frozen comparison contract (18e0e729); it cannot by itself attribute error to mapping aliasing. A multi-band distance-to-wall diagnostic is a separately scoped follow-up, not part of this run.
  • H2-P tuning: not now (MeltAssist, MeltSol). A tuned H2-P would need its own pre-registered entry and would still be compared with a GINO given equal tuning.

Prioritised next tests with stop criteria

  1. GINO-Greif (needs Yanqi's launch OK). Stop if the first-epoch projection exceeds 100,000 GPU-s, pairing is incomplete, or the local prediction-freeze / hash gates fail (Drive archiving stays on an independent HOLD and does not gate local scoring). Outcome per pre-registration: GINO better → GINO-type becomes the default for both tasks; comparable → keep both, decide on cost; H2 better → isolate H2's molding-specific advantage. Last fill and the hard subset reported, not gated.
  2. Metadata audit (CPU, no labels): v4 LOG/YML coverage, warpage-set family overlap and V/P fields. Stop treating a source as a same-process holdout if families overlap, fill is incomplete or switchover < 99 %.
  3. Family-holdout evaluation on genuinely new families (from step 2), frozen models only, pre-registered field and tail tolerances. Stop promoting any model whose CI on new families fails the tolerance.
  4. Parked: last-fill losses (until a cross-seed reproducible mechanism appears), Moldflow generation (Adviser cannot export per-node fields; needs Insight/Synergy), pretraining / weight transfer (would need pretrained-vs-random at fixed target data and compute).

Unresolved

None of substance among members. Open external dependency: data rights for new simulated families (Insight/Synergy licence or partner data).