47 KiB
47 KiB
MeltMind research tree (Claude maintains it, single writer; started 2026-11-17)
Method (Yanqi): explore many branches; none has to succeed. Each branch gets a cheap gate, learns something and is pruned or expanded. Every probe, counter-example and prune reason is kept here. Goal: state-of-the-art fill prediction.
Rules:
- A branch starts with a gate: the cheapest test that could prune it (about 2 CPU-h or less, or existing outputs only). Gate criteria are written down before the run.
- Statuses: OPEN, PROBING, PRUNED (with reason), PROMOTED (gets a contract), PARKED.
- Every result records what we learned, even when the result is negative.
- At most 2 promoted branches run at a time; gates can run in parallel on free cores.
- Single writer per execution root (MeltAstra on Horsea). One review from Claude per contract.
| ID | Branch | Gate (cheapest falsifier) | Owner | Status |
|---|---|---|---|---|
| B1 | H1: residual on the eikonal volume map (t* - cF, same 19 features) | full run is cheap (<= 20 CPU-h) | spec Mold, run Melt | PARKED (inconclusive under the fixed rule) |
| B2 | H2: geometry-aware operator on the H1 target | pilot <= 2 GPU-h with a cost gate | spec MeltAssist, run Melt | CONFIRMED (field, five folds); last fill significantly WORSE → phase 3 |
| B3 | Arrival-field topology: dry-region H0(-T) maxima and saddles as last-fill / front-meeting diagnostics | on about 50 parts, compute the diagnostics from the true T and from the full-node EIK; check whether they are stable and whether EIK's critical points match the truth | spec Mold, run Melt | OPEN |
| B4 | Filling as a filtration: one scalar potential plus a monotone time map | oracle: the best per-part monotone map from the EIK potential to the true T; the remaining error is ordering error that no time map can fix | Mold / Melt | PRUNED (clock-only on a fixed potential cannot beat A) |
| B5 | Sheet / paired-wall decomposition on real meshes | scope a relaxed general-domain certificate on paper; then try 10 Greif parts and report coverage and false connections | Mold, then Melt | PARKED (coverage unknown: certifier complexity cap) |
| B6 | Intrinsic coordinates (Laplace-Beltrami / spectral) as features | 50 parts: does a spectral feature explain A's residual (partial correlation on the learning folds only)? | MeltAssist lit, Melt run | PRUNED (this fixed HKS recipe: |
| B7 | Order-aware objective (pairwise or rank loss) instead of pointwise L2 | small GBRT ranker on one fold: Kendall tau and last-fill location against A | MeltAssist / Melt | PRUNED (this fixed ranker recipe) |
| B8 | Fresh hold-out from Greif v2 (42 parts, 40 new source tokens) | read-only LOG/YML, field validity and near-duplicate check | MeltAssist | PARKED (missing process provenance: v2 ZIP has 723 VTU and 0 LOG/YML, so no t_set) |
| B9 | Heat transfer / packing targets (cooling time, shrinkage) | read-only inventory of v4/v5 fields | MeltAssist | DONE (inventory) → opens B10 |
| B10 | Freeze time / volumetric shrinkage as targets with the same Gre24-style pipeline (Yanqi: fill first, so low priority) | gate: per-field validity (units, stage, mask) on 20 parts, then a Gre24-style GBRT on one fold against Gre24's reported medians (cooling about 0.095, shrinkage about 0.10) | Melt / MeltAssist | PARKED (priority) |
| B11 | Topological cleaning of predicted fields: persistence simplification removes spurious extrema and saddles before the end-of-fill and event metrics | gate: on the B3 panel, apply a fixed persistence threshold to EIK, and later to the A/H1 full-node predictions, and check event precision, last-fill distance and tau (learning folds only) | Mold spec, Melt run | PRUNED (this fixed rule, all three arms) |
| B12 | Path-integrated flow-resistance features: along the gate-to-node path, the bottleneck (minimum) thickness, the integral of ds/h^2, and the number of junctions crossed. Flow-competition information that Gre24's 19 columns lack | gate: A plus these 3 columns on one grouped fold at the 100 % pool; field L2 and last-fill distance against A, paired CI | Mold spec, Melt run | PRUNED (this fixed recipe) |
| B13 | Region-context features without type matching (a covariate-controlled O1 precursor): means of F and D_w over generic connected regions at K = 8, added to A | gate: one fold. This is a cheap reference only: a null result does not close type-organised context, O1 or real connectivity, which stay independent open hypotheses | Mold spec, Melt run | CONFIRMED (E1b, five folds); E2 (type partition) PRUNED |
| B14 | Flow-splitting index: number of near-shortest path families (5 % of D_f) and non-largest family band-volume share | B12's fixed rule, one fold | Mold spec (c09d428d), Melt run | PRUNED (fixed recipe, after a 16,000 CPU-s rerun; last-fill CI just crosses 0) |
| B15 | Learn the propagation cost, keep the solver: regress local speed f = 1/abs(grad T) (true T, learning folds only) from geometry features, then solve the eikonal with the predicted speed and map it to time with the volume clock | gate: one fold; field L2, last-fill distance and tau against A. The speed must be inferable at prediction time (no test-label velocity). Literature basis: EikoNet / Eikonal DeepONet factorisation (MeltAssist check) | Mold spec, Melt run | PRUNED (fixed recipe: field and last fill both clearly worse) |
| B10x | Freeze/shrinkage validity gate (unparked as a cheap gate while compute is idle) | per-field validity on 20 v5 parts | Melt | PARKED (physical definition and units missing) |
| B16 | Combine the two field gains: B13 context columns + the H1-RES target t* - c*F, same recipe | five folds, test roles (552 / 204); primary new - B13-E1b field upper bound < 0; vs A reported | Melt run | PRUNED (fixed combination recipe; gains do not add) |
| B17 | Last-fill branch: 23 columns = A 19 + B13 G-region 2 + B14 (K, S); learning weight up to 4:1 on the last 10 % of volume (MoldAstra spec b5bd8a4e) | stage 1 screen on the old 5 folds: mean Δlast ≤ -.010 and paired upper < 0 vs E1b; stage 2 confirm only on the sealed fresh holdout (≥ 25 new components, upper < -.010) | Mold spec, Melt run | PRUNED (stage 1 screen failed; stage 2 not run) |
| P1 | O2 local tensor readings | — | — | PRUNED (S1, S2, S3-A: no gain on the current recipe) |
| P2 | Own graph flow-competition solver | — | — | PRUNED (infeasible: splu resource stop, GMRES numerical failure; the idea itself is untested) |
| P3 | O1 with size-matched K = 8 partitions | — | — | PRUNED (eligibility 23 %; lesson: control cut area as a covariate, do not match it) |
Log (newest first):
- 2026-11-20: k4 confirmation on Greif folds 1-4 (bindings c1080ca1 + amendment 652d6d8b, MeltAssist PASS; second-use ledger, 431 reads; result 74eee994). Primary metric = distance to true last-1 % volume support, hit ≤ .10 Rg; family-paired 95 % CI. Primary report on the 320 parts / 132 families never used to select k: H2 full-node argmax 85.05 % → k4 88.51 %, +3.46 pp [+.64, +6.20] → k4 gain CONFIRMED; mean distance .107 → .076 Rg. Common PPS nodes, three arms: A 88.56 % (.074 Rg), H2-argmax 88.04 % (.085), H2-k4 89.09 % (.062); k4 − A +.53 [−3.02, +4.55] → inconclusive; k4 − H2-argmax +1.05 [−1.69, +3.75] → inconclusive. Secondary (distance to the single true global peak): A 21.1 %, H2-argmax 15.1 %, H2-k4 18.8 % hits; A remains better at pinpointing the exact peak. Reading: k4 is a confirmed improvement for H2's endpoint on full nodes; against A on the shared query nodes H2-k4 is level (no evidence of a gap either way on the tail-support metric), whereas the earlier 'last fill worse' finding was on the global-peak metric, where A still leads. Zenodo 170 second-use run pending relay.
- 2026-11-20: Last-fill readout: k4 mesh smoothing before argmax improves H2's endpoint (fold 0, pre-registered). Spec MoldSol 226c2b1e (frozen with my amendment), binding b1955c29 (MeltAssist PASS after two fixes), CPU only, no retraining. Validation (fold 0, 111 parts / 35 families, pooled seeds, hit = official d ≤ .10 Rg): argmax 85.89 %; k1 88.91 % (+3.02, LB +.27); k2 89.58 % (+3.68, LB +.74); k4 90.17 % (+4.28, LB +.94) → frozen winner; centroid 58.18 % (−27.71). Test (fold 0, 121 / 37, one read each): seed 0 hits 82.04 → 84.42 % (+2.38), mean distance .1505 → .1270 Rg; seed 1 83.03 → 92.98 % (+9.95), .0582 → .0374 Rg; combined paired gain +6.16 pp, 95 % CI [+.39, +13.60]. Reading: supports the spurious-maximum hypothesis from the 34-case classification: reading the endpoint from a mesh-smoothed field, not the raw argmax, recovers part of H2's last-fill loss. Field predictions unchanged. Limits: one fold, wide CI; not yet compared against A on the same metric; not yet applied to other folds or the 170 holdout.
- 2026-11-20: NEW-FAMILY HOLDOUT RESULT: H2's field gain TRANSFERS (Zenodo 17831586, 170 cases / 98 ABC families, zero roster/near-duplicate overlap; frozen 5-fold ensembles of A, H2 seed 0, H2 seed 1, no retraining; pre-score PASS a670a4d4/f314ca15 + split PASS 06c35e1a; predictions frozen 202a01a1 before any target read; each Fill_time decoded once; result a236d696, target ledger 0f3cd58d). Family-weighted field relative L2, H2 − A, 98-family paired bootstrap (2,000, seed 2026111408): seed 0 −1.322 pp, CI [−1.721, −0.959]; seed 1 −1.364 pp, CI [−1.769, −0.971] → both FIELD_GAIN_TRANSFERS (pre-registered). Last fill: hit counts equal across the three; distance up for seed 0, slightly down for seed 1 (descriptive, no gate). Reading: on genuinely new part families from a different public release, the learned mesh model's field advantage over the tree model holds and is larger in absolute terms than on the grouped Greif protocol (−0.50 pp); last fill is still not improved. Limits: same solver/material/process as training; single gate per part; one release. Compute: 1a 37,748 CPU-s, 1b ~196 GPU-s, scoring 83 CPU-s.
- 2026-11-20: 34 shared misses classified by true last-3 % region (bounded CPU read, MeltAssist PASS 53c9c543; result by-part.csv 3d6c38e1; hypothesis-generating). Same region / different region / NA (peak > .10 Rg from every last-3 % component): H2 seed 0 4/5/25, seed 1 8/4/22, A 16/4/14 (A on query nodes only). Reading: for most of H2's shared misses the predicted last point is not in the true late zone at all; A's misses are mostly inside the right late region. This corrects my visual note (four pairs suggested 'same end region, different point' was the main case; that is the minority for H2) and is consistent with the S2 diagnostic (H2 more far-from-tail picks, 71 vs 49). Untested hypothesis: H2's predicted field has spurious late maxima away from the true late zone (endpoint = argmax is sensitive to small local errors), rather than a wrong branch choice.
- 2026-11-20: New-family holdout established: 170 cases / 98 families (warpage set, Zenodo 17831586). ZIP verified on Horsea (MD5/SHA-256 49884675…). Geometry vetting under MeltAssist PASS d2a31203 (scope COMBINED_VTU_LOG_SCALE_YML_SURFACE_ALIGNMENT; Points/Cells only, Fill_time sealed): 173 pass (unit uniquely mm from LOG volume ≤ 0.5 %, counts, finite/positive volumes, YML inlet on surface), 23 excluded NO_LOG_SCALE_MATCH. Near-duplicates (pre-frozen < 1 % fingerprint rule) vs our 552 / 204 families: 0 of 95,496 pairs; within the set 0 of 14,878; roster identity overlap 0. Minus 3 cases with V/P switchover < 99 % (10555, 10656, 10941; pre-registered rule) → 170 cases / 98 ABC families: same solver, material, melt/mold temperatures and fill rule as Greif, complete fill, new families. This is the first independent test of 'new parts'. Next: frozen-model evaluation spec (pre-registered before any Fill_time read).
- 2026-11-19: Warpage set (Zenodo 17831586) metadata census: candidate new-family fill-time holdout (Claude, HTTP range reads of the 392 LOG/YML entries only; no VTU bytes fetched; docs/claude/data/greif_warpage_log_census.csv sha256 e44ff604…, scripts warp_fetch.py / warp_census.py). 588 ZIP entries = 196 VTU + 196 LOG + 196 YML. All 196: melt 543.15 K / mold 318.15 K (same as main corpus), last filling-phase row 100 % 'Filled' before packing starts, packing profile 0 % (applies only after fill), no short-shot text; V/P switchover median 99.83 %, 193/196 ≥ 99 % (3 between 98.51 and 99 %); gate coordinates + normals present for all. Family overlap: 110 ABC chunk hashes, 0 shared with the 249 v5 hashes, 0 shared abc_id+step. Reading: under the pre-registered rule (complete fill, switchover ≥ 99 %, new families) 193 cases qualify as a same-process new-family holdout, pending (i) overlap check against our actual 552-part roster (v5-based: Zenodo 18598121; corrected 2026-11-20, MeltSol) and (ii) a geometry near-duplicate check once the ZIP is on Horsea. Labels (nodal Fill_time) are listed in the record but untouched. Access note: Zenodo returned 403 to a 'Mozilla/5.0' User-Agent and 206 to curl's default UA from the same host; Horsea's 403s likely came from the UA I had suggested.
- 2026-11-19: H2-P on ShapeNet-611: H2_P_IMPROVEMENT_NOT_ESTABLISHED (worse). Combined entry PASS 1359ba98; three seeds, 100 epochs, matched budget, untuned direct pressure head, molding inputs removed; 111 targets read once per seed. H2-P 10.013 % vs reproduced GINO 7.435 %; pre-registered paired H2-P − GINO +2.578 pp, 95 % CI [+2.327, +2.849] over 111 single-part components (no family structure). Frozen result fcff9e9c (MeltSol); MeltAssist checked. Reading: on this pressure protocol, the H2 architecture as adapted is clearly worse than GINO. This is not weight transfer (trained from scratch) and says nothing about H2's Greif result; it is the planning ruling's 'H2-P behind GINO' branch. Stage review SR-20261119-2137 factual cutoff = this publication.
- 2026-11-19: Moldflow Adviser data route NOT ESTABLISHED for per-node labels (wording corrected 2026-11-19 22:47, RESET-20261119-2247: an STL-import solve is not ruled out; what was not established is direct import of the original BDF and a usable per-node export / ID mapping) (MeltWin on ARON: Adviser 2027 imports STL, no BDF; Grok, official docs: Adviser exports only .mfr for Communicator, API scripts unsupported; Communicator is a viewer with no documented per-node export; per-node XML export is Insight-only). The approved single original-mesh local solve was not used (0 solves, 0 credits); MeltWin stays stood down. Data generation would need Insight/Synergy via another licence (question for Yanqi at the next stage report). New lead instead (Grok): the public Greif–Meyer corpus ships Moldflow LOG + inlet YML from v4 on, and a 196-case warpage set (Zenodo 17831586, Dec 2025; packing pressure removed, so late fill may differ, per TRON/Autodesk V/P docs). MeltSol checks overlap with our 552/204, v4/v5 LOG/YML coverage of B8 and the 94 sealed candidates, and per-case V/P switchover / end-of-fill / short-shot (metadata only). Pre-registered: a warpage case counts as same-process new-family holdout only if fill completes and switchover ≥ 99 %, with the post-switchover tail reported separately; otherwise 'transfer under changed process'.
- 2026-11-19: Track B: GINO reproduced on ShapeNet-611 (pre-registered tolerance). Identity: NeurIPS supplement code + 2-keyword patch (inp/outp_features → in/out_features; no variant accepting the old names found). Paper package, 500/111, 100 epochs, 64³, de-normalised per-part relative L2; each seed read the 111 held-out cars once after its frozen epoch-100 checkpoint. Seeds 0/1/2: 6.948 / 7.764 / 7.591 %; equal-weight mean 7.435 %, inside [6.62, 7.62] → REPRODUCED. 30,351.6 GPU-s (+ 222.8 stopped measurement). Reading: the baseline is usable for same-protocol comparison, but our mean is .31 pp above the paper's 7.12 % and seeds spread .82 pp (one seed alone is above the band), so any ranking must be seed-paired and multi-seed; single-seed comparisons against 7.12 % are not meaningful. Weights in private Drive with SHA read-back.
- 2026-11-19: 1b (100-epoch full schedule) SCREEN_FAIL; CLOSED. Fold 0 validation, 111 parts / 35 components, component-weighted relative L2: 200-epoch seed 0 6.085 %; 100-epoch seed 0 5.936 % (−.149 pp, first gate passed); 200-epoch seed 1 5.877 % (−.208 pp vs seed 0). The schedule effect is smaller than the seed difference, so the second gate fails. One labelled recovery (GPU0 foreign occupant); final run guard STOP at exit, cause undetermined; outputs audited complete (MeltAssist e5cfba65; adoption criterion amended before values were seen, timing item dropped as attribution-only). 5,840.58 GPU-s + 741.16 stopped. Reading: no evidence that the shorter schedule helps beyond seed variation; item 1's overfitting signal is neither confirmed nor refuted. No further folds. Note: the phase-3 guard's exit race (snapshot vs process-list) remains a known limitation of this guard version; the GINO guard has the fix.
- 2026-11-19: Shared-miss attribute study CLOSED (descriptive; MoldAstra prereg d62b919e; deliveries
44c1d5a1+ 40f9e4ce; MeltAssist PASS 77639066-series: 7029a6e6 and combined 2930943d; 0 GPU-h, 109 CPU-s; true test-role fields read once under a bounded CPU-only scope, hypothesis-generating only). 34 shared-miss parts / 28 families vs 518 / 193 controls. Only clear difference: A also misses 17/34 vs 32/518 (family-weighted 41.1 % vs 6.7 %, +34.3 pp, family-block CI [17.0, 53.1]). Main tail fragments (1.29 vs 1.28), near-tie share (22.1 % vs 21.3 %) and weighted gate distance (2.72 vs 3.06 Rg) show no difference (CIs cross 0); single gate everywhere; wall thickness, physical branches and semantic type remain NA. Reading: the shared misses are largely parts that are hard for both model families; the true-tail geometry measures tried here do not explain them. Candidate explanations left (untested): missing inputs (thickness, process), label quality on those parts, part semantics. Ruling: close the study; no new last-fill loss; the 34-part list is kept as a named hard subset to report separately in future evaluations (Track B / new families). Any fix must be confirmed on data not used here. - 2026-11-19: H2 family/seed diagnostic (descriptive, pre-registered 97f29e53; result 2f8d8787; MeltAssist PASS 77639066; 0 GPU-h). 552 / 204. Family-weighted field L2: A 7.083 %, H2 seed 0 6.579 %, seed 1 6.594 %; both seeds improve in all five folds and no leave-one-family-out flips the sign, so the field gain holds at family level. Tail misses (> .10 Rg from the true tail): both seeds 34 parts / 28 families, seed 0 only 37, seed 1 only 28, neither 453; the reference count expected if the two seeds' misses were independent is 71 × 62 / 552 ≈ 8. Reading (corrected wording, MeltAssist): the shared misses are repeated descriptive evidence that gives a candidate list for a mechanism study. They do not establish a mechanism, rule out seed effects, or carry formal significance. (My Slack reply had called it 'systematic, not seed noise'; that was too strong.) Next: pre-registered descriptive attribute study of the 34 parts (MoldAstra attributes, MeltSol), 0 GPU-h.
- 2026-11-19: Planning ruled (four independent answers at 2c0dd614; ruling in planning-discussion-2026-11-19.md). Short: 1b + a zero-GPU family/seed/fold paired table from frozen scores. Medium chain: GINO reproduction on ShapeNet-611 → H2-P vs GINO on the same protocol → GINO as a Greif fill-time baseline. Long: new families with full process inputs, whole-family holdout and paired gate/process interventions; blocked on data rights (question to Yanqi).
- 2026-11-19: Stage review ruled (answers at 52c74d25; ruling in stage-review-2026-11-19.md): Track B (GINO ShapeNet protocol) is the main line, starting with the data-mapping check and a GINO reproduction gate (7.12 %); 1b reruns in parallel inside its existing authorisation with a co-located cap and narrow review; no new last-fill loss; Track A closed at the four-split descriptive result; five process rules adopted (contract → real guard CPU simulation, concurrency-measured caps with waiting booked separately, four-state routing, scoped restrictions in handoffs, delta-only reviews of unchanged components).
- 2026-11-19: Item b (H2 + frozen B13 context inputs) DOES_NOT_PASS its fixed screen; PRUNED (this recipe). Recovery postfit (PASS 029a77a2) COMPLETE, 552 / 204, evidence
192ce24a. b − H2 seed 0: field −.00006, CI [−.0018, +.0017] (needed upper bound < −.001 and |effect| > seed spread .00015: neither met); last fill −.044 Rg, CI [−.130, +.035]; τ-b +.0002, CI crosses 0. Descriptive strata (no gate): 510 parts with valid B13 context −.0008, CI [−.0026, +.0009]; 42 historically missing +.0089, CI [+.0026, +.0156] (worse where the context is absent and zero-filled). Reading: B13's tabular context adds nothing measurable to H2; the representation appears to already carry the regional information that helped A. Phase-3 scored set is now complete: H2 confirmed on field; seed noise small for field, large for last fill; c and b pruned; 1b parked; Track A shows H2 < A on our random splits without a paper-level reproduction. Stage review opened (docs/claude/stage-review-2026-11-19.md). - 2026-11-19: Item c (last-fill-aware objective inside H2) FAILS its pre-registered screen; PRUNED (this recipe). Recovery postfit (PASS 59a6255f, one labelled infrastructure recovery after a pre-worker guard KeyError, 0 targets read in the stop) COMPLETE, 552 / 204. c − H2 seed 0: last fill +.0193 Rg, CI [−.0559, +.0948] (gate needed ΔD ≤ −.010 with upper bound < 0, and |effect| > the same-domain seed spread .0482: not met); field +.00029, CI [−.0013, +.0019] (upper bound above +.001: fails); τ-b +.00037, CI crosses 0. Secondary (reported separately, no gate): last-1 % tail error E_tail +.066, CI [+.044, +.087], i.e. worse tail placement; peak certificate 0 for every arm. Reading: this location-aware objective did not move the last-filled point, and by the secondary metric it moved the late region the wrong way. That closes this recipe, not the question: the S2 diagnostic (more far-from-tail picks) still stands as H2's main weakness. 630.73 GPU-s postfit including the stop.
- 2026-11-19: Track A (Gre24-style random-split protocol) scored, and seed-1 noise reference scored (MeltSol, evidence fee71ea6; original bindings 958abd8c/a5fa4a6e and 7b4da089, each run once). Track A: splits 0/1/2/4 (split 3 stopped on budget, disclosed; 14 preparation failures disclosed; our own random splits), 1228 paired test visits / 571 unique parts. Split-averaged median per-part RMSE/max: A 3.99 % → H2 3.25 %; per-part paired Δ −.00742, CI [−.00844, −.00646]. Reading: on the same re-created random splits H2 clearly beats A. This is not a SOTA claim against Gre24's table: our A on these splits gives 3.99 %, not the paper's 3.31-3.38 %, so the plan's reproduction gate is not met and the gap between H2 3.25 % and the paper number is not interpretable (four splits, different seeds, random splits leak families). Seed-1 vs seed-0 (H2, grouped protocol, 552/204): field +.00015, CI [−.00143, +.00204]; last fill −.048 Rg, CI [−.122, +.024]; τ +.00004, CI crosses 0. Reading: the seed spread is small for the field (about 1/30 of H2's −.005 gain, so the field confirmation is not a seed artefact). For last fill it is about .05 Rg, comparable to H2's +.080 last-fill loss, so single-seed last-fill differences are noisy. c and b will be read against the same-domain seed-1 reference, as their bindings specify.
- 2026-11-19: All phase-3 training terminal, no scores yet (TRON read-only, 07:10 and 07:30 CST). seed-1 v2: five folds PASS, 105,759.84 GPU-s; postfit PASS 7b4da089 already adopted, postfit not started. b: five folds PASS, 95,685.04 GPU-s on the ledger (9,647.52 of it charged waiting), postfit binding not yet written. For c, seed-1 and b: no prediction, scoring or test-target read; 7,200 GPU-s postfit reserve each. Horsea has no compute processes. Native routing is still missing; order once it returns: c postfit review → Track A shared-slot narrow review → seed-1 postfit → b postfit binding + review.
- 2026-11-19: Item c training terminal (TRON read-only, 05:20 CST): five folds PASS, TRAIN_COMPLETE_POSTFIT_REVIEW_PENDING, 105,362.42 GPU-s spent + 7,200 postfit reserve; no prediction, scoring or test-target read. Native routing is down (tool update removed TRON's Codex session send), so c's postfit review and Track A's shared-slot delta review (
3e62e08c) are held, with no Slack substitute and no TRON exception. seed-1 (4 PASS / 1 active) and b (5 active) keep running. - 2026-11-19: Item b launched under the shared capacity rule (MeltAssist narrow delta PASS 9c91a641, MeltSol fixed 1d2ce8d0, start audit published
3438020b): folds 0/1 on GPU0, fold 2 on GPU2, folds 3/4 pending; first-round projections about 143.4k GPU-s, cap 172,908. Track A (958abd8c) is starved: its reviewed gate needs a whole idle GPU2, which the shared queues (c folds 2/3, b fold 2, b folds 3/4 pending) will not leave free for a long time. Ruling: move Track A to one shared-rule slot through a narrow delta (lease + guard, cap re-sized from co-located epoch times + 25 %), reviewed by MeltAssist before launch; scored binding, splits and metric unchanged. - 2026-11-18: Phase 3 item 1 (data sufficiency) COMPLETE: UNIDENTIFIED_OR_OVERFIT_SIGNS (pre-registered rule; validation and test). Corrected reading (my first reading had the g sign backwards; MeltAstra correction): g = error(U) − error(2U), so the negative g values mean the 2U schedule had higher error (test 50 %: .0688 → .0728; 100 %: .0668 → .0696; validation 100 %: .0584 → .0613), i.e. overfitting signs, not a gain from longer training. s_2U = (error50 − error100)/log(n100/n50) > 0 means more data helps: validation +.0033, CI [+.0011, +.0055]; test +.0032, CI crosses 0. One fold / seed; budget and schedule changed jointly; the rule's classification stands. Track D trigger not formally met; the longer-schedule candidate is withdrawn (direction was wrong); overfitting control is the better-aimed candidate after seed 1.
- 2026-11-18: S2 last-fill diagnostic (descriptive, post hoc, no gate; MoldAstra spec 05344bac; result
98519a6c). 552 / 204. Full-node peak class A / H2: near the true peak 153 / 129; inside the true last-1 % tail but away from the peak 350 / 350; outside the true tail (far) 49 / 71 (component-weighted 9.4 % / 16.4 %); exact tie at another true peak 0 / 2. The γ > 2ε certificate holds for 0 / 552 in every combination (errors are too large relative to peak margins for the sufficient condition, so it is uninformative). Reading: H2's last-fill loss is mostly more far-from-tail picks plus fewer near-peak picks, not exact-tie flips; my hypothesis (b) (metric fragility) is at most a partial explanation. For about 63 % of parts both models land inside the true late region but not at its peak, so the set-based endpoint matters. This supports item c (location-aware objective) as designed; nothing is decided by this diagnostic. - 2026-11-18: Phase 3 item 1 (data sufficiency, v2) running. Design revised from e24fff7c: six fresh fits (25 % / 50 % at 200 epochs; 50 % and 100 % each at U = 32k and 2U = 64k steps with their own complete schedules), 100 % 200-epoch reused from S1; U vs 2U = joint budget+schedule change. Pre-run review PASS (MeltAssist cbd67318); cap 120,000 GPU-s. S2 published (
2338dfe9) with independent terminal audit PASS (d4999e4a); folds 1-4 alone: field Δ -.00495, CI [-.00765, -.00248]. - 2026-11-18: H2 CONFIRMED on field (pre-registered pooled rule), last fill WORSE. 552 / 204 test parts, pre-score review PASS 9c163702; postfit launched by TRON as a labelled one-time exception (MeltAstra unreachable). Field H2 .06579 vs A .07083: Δ -.00504, CI [-.00754, -.00283] (about 7 % relative); vs B13-E1b Δ -.00391, CI [-.00631, -.00164]. τ +.0035, CI [+.0010, +.0062]. Last fill H2 .4566 vs A .3767: Δ +.080, CI [+.007, +.156] (worse). Boundary: the pooled rule was registered after fold 0 was seen, so fold 0 is not independent; folds 1-4 alone to be reported descriptively. Reading: the learned representation is the lever for the field (better ordering, lower error), but it puts the last-filled point in the wrong place more often than A. Phase 3 is triggered (data sufficiency → seed 1 → c last-fill in H2 → b → seed 2 / ensemble); c is now the priority question.
- 2026-11-17: Gate calibration v2 FAIL_OUTCOME_EQUIVALENCE (pre-registered, review PASS before scoring). Label-derived minus LOG gates, 545 paired parts / 202 components: field A +.00083, CI [+.00039, +.00134]; E1b +.00077, CI [+.00035, +.00125] (both exceed ±.001); last/Rg CIs exceed ±.010. Cost 94,261 / 104,639. Consequence: the v2 fresh holdout is BLOCKED_INPUT for gate-dependent methods; the shift is of the same size as the effects we test (E1b gain .0011). The A and E1b shifts are similar, so a paired method difference might cancel, but that is an observation from this result, not a pre-registered test, and will not be used. Route forward: real gate metadata for v2 (question to Yanqi), or a different fresh source. Phase-3 item c's fresh-family stage is blocked accordingly.
- 2026-11-17: H2-S1 (fold 0 test, 121 parts / 37 components, one seed). Pre-score review PASS (MeltAssist 07514084). Field H2 .06747 vs A .07292: Δ -.00546, CI [-.01035, -.00098], about 7.5 % relative on fold 0 only (not comparable with E1b's five-fold 1.6 %: different statistical domain; no ranking or attribution to representation/data). Last fill +.110, CI [-.134, +.393] (no support); τ CI crosses 0. 7,786 full-node / 145 query predictions negative (kept, not clipped). S2 gate passed. Pre-registered before S2 results: H2 is CONFIRMED if the pooled five-fold (552 / 204 test) field upper bound vs A < 0; vs B13-E1b reported, not gating. Single fold / seed: no causal or cross-split claim yet. Data-sufficiency test (Mold e24fff7c) follows S2 under its own cap.
- 2026-11-17: B17 PRUNED (stage 1). 552 / 204, pre-score review PASS (MeltAssist 235cfc9f, after one score-assembly fix). Last fill Δ (new - E1b) -.0033, CI [-.0389, +.0298]; it needed mean ≤ -.010 and upper < 0. Stage 2 not run. Cost 4,745 / 10,800. Learned: the B14 / B16 last-fill near-misses do not survive a dedicated, tail-weighted recipe. Last fill stays unsolved by the tabular feature route; it is not closed for learned representations (H2) or for other targets. Gate calibration v2 (outcome-level, 72k CPU-s) running; the fresh holdout remains useful for an independent deployment check of E1b.
- 2026-11-17: B17 opened (MoldAstra spec b5bd8a4e, accepted as written). Stage 1 = screening only on the already-used 552 parts; stage 2 = one unsealing of the fresh holdout. B14 / B16 near-misses are one hint, not two. Fresh holdout: 94 candidates sealed (a2dac8d1 / manifest a5a133a2; 72 source groups, 42 / 40 new tokens); gate extractor must first reproduce the LOG gates on the 552. The 10 GB evidence is copied to /home and hash-verified (not a remote backup).
- 2026-11-17: B16 PRUNED (fixed combination recipe). 552 / 204 test parts, pre-score implementation review PASS (MeltAssist, 8cdd3ff4). New - E1b field +.00014, CI [-.00051, +.00079]; last fill -.039, CI [-.090, +.004]; τ -.0011, CI [-.0019, -.0004] (worse). New - A field -.00099, CI crosses 0. Learned: the H1-RES target does not add to B13 context. Last fill is a near-miss again (B14 too). Two near-misses on the same data are a hint, not evidence: any last-fill branch must be confirmed on fresh parts. Next: (1) v2 extra parts (94, 42 from new families) prepared as a sealed fresh holdout; (2) H2 data-sufficiency test (Mold e24fff7c) after H2-S1; (3) CAD pretraining (Mold 8114bc3c) parked behind H2 (it needs a learned model worth pretraining); (4) Mold to spec one last-fill-targeted branch, confirmation only on the fresh holdout. Literature (MeltAssist 1-3/3): video-model OOD physics failure has limited support (Kang et al. ICML 2025, controlled 2D); PDE pretraining sample efficiency has evidence (Poseidon, CoDA-NO); generation on irregular meshes has precedents, but there is none for 3D volume filling or calibrated uncertainty; CAD sets give geometry only.
- 2026-11-17: B13-E2 PRUNED (fixed type-context recipe). 552 / 204 test parts, implementation review post hoc PASS (MeltAssist). Raw T-G field +.00028, CI [-.00075, +.00118]; last fill +.019, CI crosses 0; τ CI crosses 0. Field means G .06970, T .06998, C .06973. The adjusted T-G interval is unavailable (13.7 % of replicates have the zero point outside support, above the 5 % rule); the T-C and C-G adjusted intervals cross 0. Learned: at this recipe the B13 gain comes from regional averaging of F / D_w* itself, and which rule draws the 8 regions (generic, type-organised or shuffled type) makes no detectable difference. Not an equivalence claim. It does not close O1 with other readouts or connectivity. Cost 6,631 / 14,400.
- 2026-11-17: B15 PRUNED (fixed recipe). Fold 0 validation, 111 / 35. Field .0918 vs A .0649 (+.0270, CI [+.0160, +.0388]); last fill +.250, CI [+.111, +.408]. Both clearly worse, so learned speed + eikonal solve + volume clock lands near the training-free EIK level. Learned (narrowed 2026-11-18): this recipe failed; a plausible reading is that pointwise speed errors integrate along paths, but several factors changed at once, so it does not show pre-solve learning is generally inferior. It does not close learned potentials or a learned moving interface. Cost 6,354 / 8,000. B16 opened: B13 context + H1-RES target, five folds, primary vs E1b.
- 2026-11-17: B14 PRUNED (fixed recipe). Fold 0 validation, 111 parts / 35 components, coverage 131/132 learning and 35/35 validation components. Field (new - A) -0.000091, CI [-0.000328, +0.000153]; last fill -0.0364, CI [-0.0738, +0.0019]. Both cross 0, so it fails the B12-B15 rule. Cost 13,951 / 16,000 CPU-s. Learned: path-family features do not move the field; the last-fill CI only just crosses 0, a weak hint at most and not a finding. It does not refute the channel/topology hypothesis in general. Any reuse must be a new, pre-registered last-fill-targeted branch (e.g. combined with B13 context), not a rescoring of this one. E2 geometry is complete (552/552, 3,868 CPU-s); postfit is being implemented. B15 prediction continuation running.
- 2026-11-17: B13-E1b CONFIRMED (Melt, local commit 044b99eb; review ZIP 04a9fc8e). All 552 test parts / 204 components, each scored once; the 42 G-invalid parts kept as native NaN. Field relative L2 (new - A) is -0.001128, CI [-0.001999, -0.000260] (A .07083 -> .06970, about 1.6 % relative). τ-b +0.00146, CI [+0.00031, +0.00277]. Last fill +0.0109, CI crosses 0. Four folds improve on the mean; fold 1 is slightly worse with its CI crossing 0. This is the first five-fold-confirmed increment over A: small, and a generic region-context effect only. It is not evidence for type organisation (O1) or a physical mechanism. E2 (T-G deployment gate, cap 14,400 CPU-s) is released. A combination of B13 and H1-RES is a later candidate. B1 erratum filed by Melt: the field condition passed, and PARK came from last fill and τ.
- 2026-11-17: B13-E1 PARK_G_COVERAGE (no fit and no score). The fold-1 test (31/41 components = .756) and fold-3 learning (101/128 = .789) roles fell below the 80 % complete-component gate. This is not a performance failure and not a five-fold confirmation. B13-E1b (separately named) is opened: the same frozen recipe on all 552 parts, with native NaN for the two G columns where G is invalid (the fold-0 B13 recipe already kept its 4 G-invalid parts this way), no coverage gate, and the same decision rule. A deployed surrogate must score every new part. The E1 PARK receipt stays. B14 per-part cap raised to 300 s, with one rerun for indices 16, 21 and 22.
- 2026-11-17: Rulings (Claude).
- B13-E1: scored on each fold's test (552 / 204, each component once), not on validation. Validation overlaps across folds (507 evaluations over 117 components) and fold-0 validation was the screening data. A's hashed test predictions are reused; the fold-0 model is reused; only folds 1–4 are new fits. Gate: pooled field upper bound < 0. Cap 10,800 CPU-s, ≤ 4 workers.
- B13-E2: spec accepted (MoldAstra, Drive 1rwDjGi…, SHA 182e0312) with three amendments: E1's test roles; the deployment decision is raw full-roster T−G under the B12/B13 rule, and adjusted T−G / T−C only label the mechanism; cap 14,400 CPU-s. Starts only if E1 confirms.
- B15: reallocated within 7,200 CPU-s (preparation 3,000 / fit 1,500 / predict 2,400 / score 300).
- B14: unparked with a 16,000 CPU-s total (preparation ≤ 13,500), no further extension.
- Run order: E1 first, then B15 and B14.
- 2026-11-17: B1 / H1-RES final: PARK. Grouped protocol, 100 % pool, 552 parts / 204 components.
- Field L2 (H1 - A): -0.00064, CI [-0.00114, -0.00016]. A small but real field improvement.
- Last-fill: +0.0065, CI crosses 0. Tau: -0.0001, CI crosses 0.
- Under the fixed rule it is neither EXPAND (no end-of-fill or order gain) nor PRUNE.
- Learned: a residual on c*F buys a small field gain only. Its size is similar to B13's on fold 0, so combining them is a candidate later, not now.
- Cost note: the topology diagnostics took about 11.3k CPU-s of the scoring; future contracts should budget for that.
- 2026-11-17: B10x PARKED (units and physical definition missing from the VTU attributes; 20/20 read). B14 PARKED_RESOURCE (projection 12672 CPU-s against the 7200 cap; not a refutation of the channel hypothesis). B15 started.
- 2026-11-17: B13 EXPAND (fold 0). Field relative L2 (new - A): -0.00163, CI [-0.00333, -0.00011]; that is .06323 against .06486, about 2.5 % relative.
- Last-fill distance/Rg: +0.082, CI [-0.036, +0.235], not improved and possibly worse.
- G coverage passes both 80 % gates; parts without a valid G keep NaN in the 2 new columns, and nothing was dropped. No Kendall tau in the frozen implementation (a gap, kept).
- Learned: generic connected-region context carries a small but real increment for field accuracy, with no evidence of an end-of-fill gain.
- Next steps: E1 (all 5 folds at 100 %, add tau) to confirm, and E2 (type-organised against generic regions, covariate-controlled) to test O1 properly.
- 2026-11-17: B12 PRUNED (fixed recipe), 2700 CPU-s; details in Melt's report. B13 running.
- 2026-11-17: H2 pilot (GPU0, fold 0, 1 epoch): 95 s per epoch, 11 s per validation pass, about 1.3 GB VRAM, finite throughout. Projected cost of the full 5 folds x 200 epochs: 35.8 GPU-h, above the 16 GPU-h cap, so PARKED_RESOURCE under the pre-registered gate.
- Ruling (no results seen; a resource decision only): stage it.
- H2-S1: fold 0 only, 200 epochs, about 7 GPU-h.
- If H2-S1 is not clearly worse than A on fold 0 (paired component CI of field L2 with upper bound < +0.005, or a last-fill improvement), run H2-S2: folds 1 to 4 on 4 GPUs in parallel.
- The total H2 GPU budget is raised to 40 GPU-h; the GPUs were idle.
- Ruling (no results seen; a resource decision only): stage it.
- 2026-11-17: MeltAssist literature check (EikoNet, Eikonal DeepONet, LENO, level-set advection operator): three distinct object classes (a solver for fixed media; an operator across media; transport of a moving interface). Future branches separate learning the potential directly / learning the propagation cost and keeping the solver (opened as B15) / learning the moving interface. B4 does not close learned potentials.
- 2026-11-17: B1-v2 scoring hit the post-fit CPU cap (a budget stop); +36,000 CPU-s labelled extension for that scoring only. H2 CPU tests pass (MeltAssist review); the pilot then the full run (4 GPUs) are authorised. B14 and B10x running in parallel.
- 2026-11-17: B6 PRUNED (fixed HKS). Partial r = .044 / .053 / .062 at t = .1 / 1 / 10 over lambda_1, all below .10 (learning folds only; rank of the 19 columns plus intercept is 19 of 20).
- Learned (scope per MeltAstra): a linear partial correlation on the learning side, controlling for the 19 columns, is small. Neither nonlinear increments nor out-of-sample effects were tested.
- 2026-11-17: Synthesis after B3-B11. Clock (B4), loss (B7), cleaning (B11), spectral context (B6) and the sheet certifier (B5, parked) are all ruled out at these recipes. What remains open is new information about how flow competes: B12 (path resistance) and B13 (region context); learned representation H2 after B1's decision.
- 2026-11-17: B11 PRUNED for all arms (fixed rule, 0.020 persistence).
- A: precision +.0034, recall -.102, last-fill distance +.043 Rg.
- H1: precision +.0024, recall -.103, last-fill distance +.009 (within gate). Tau changed slightly in both.
- No change to the whole-part maximum plateau in any of the 100 fields; the query representative last point changed for 12 A and 15 H1 parts (kept separate).
- Learned: persistence cleaning at this threshold costs about 10 % recall for under 0.5 % precision on learned predictions too. Spurious events are not simplification-removable noise for any arm. Commit
97cefdb2.
- 2026-11-17: B7 PRUNED (fixed ranker recipe). Fold 0 validation, 111 parts / 35 components.
- Kendall tau-b: ranker .9305 vs A .9339, difference -.0034, CI [-.0059, -.0004], so slightly worse.
- Last-fill distance/Rg: ranker .370 vs A .295. The mean is worse, but the paired 95 % CI crosses 0, so it is not significantly worse.
- Learned: this ranker recipe does not beat A on order or end of fill. Because the training support differs (up to 2048 points per part against A's full tet mesh), the effect of changing the loss alone cannot be identified. Raising the priority of representation is a research judgement, not a finding. The confound remains that the ranker saw sparse samples while A trained on the full tet mesh, so this does not close order-aware or topological objectives in general. Commit
1884fa3b.
- 2026-11-17: B1-v1 completed naturally, incomplete and without decision value. The B1 recovery (5 predictions, no refit) ran; v2 scoring is under way. B11 A/H1 full 50 built. B6 running.
- 2026-11-17: B11, EIK arm: PRUNE_THIS_FIXED_RULE_ON_THIS_ARM.
- Events before/after cleaning: TP/FP/FN 244/16956/416 → 215/13505/445.
- Precision +.0015 (gate +.010 not met); recall -.044 (allowed -.020; breached); query last-fill distance +.063 Rg (allowed +.010; breached); tau unchanged.
- Learned: about 80 % of EIK's spurious events outlive the 0.02 persistence threshold, so they are persistent structural critical points, not short-lived noise. Cleaning removes some true events and moves the end of fill. The eikonal potential's topology differs from the true field's at the structural level, which fits B4 (the order itself is wrong, not just the clock).
- The A/H1 arms are still pending B1's prediction freeze.
- 2026-11-17: B5 PARKED (coverage unknown). On 10 learning parts, 5 had no candidate; on the other 5, the first candidate hit the fixed 70000-operation complexity cap and came out UNKNOWN. 0 certified and 0 rejected; only 5/10 parts have a certified union that is complete and below .20, short of the >= 6/10 threshold.
- Learned: the relaxed paired-wall certifier is computationally infeasible on real Greif meshes at this cap. That is not a refutation of sheet decomposition.
- Reopen only with a new certifier design (no cap extension under this gate). Cost 133.7 CPU-s. Commit
6aa7baf6.
- 2026-11-17: B1 first-cell cost gate PASS (fit projection 3610 CPU-s); fits 01-04 running.
- 2026-11-17: B3 result (50 learning components):
- The true-field H0(-T) events are stable under recording-precision perturbation: precision/recall .954/.921 (0.001 grid) and .947/.915 (half grid), above the 0.80 gate.
- EIK against the truth: 244 TP, 16956 FP, 416 FN, so precision .014 and recall .370; median match distance .049 Rg.
- Learned: the eikonal field has a very large number of spurious topological events. This supports a stable diagnostic on this panel; it is not a weld-line validation.
- New branch B11 (topological cleaning).
- 2026-11-17: B4 PRUNED (clock-only). The post-hoc best monotone map of the fixed D_f potential gives a component-mean relative L2 of .0916, against A's .0664 on the same full-tet domain.
- Learned (narrowed per MeltAstra): A is not a member of the family of monotone scalar clocks on this fixed D_f, so that family alone cannot explain A's advantage. It does not attribute the advantage solely to order changes.
- This does not prune new potentials, branch information or learned ordering. Commit a6998f72; push pending.
- 2026-11-17: B1 contract
e751e5b2approved. Branch decision rule (grouped protocol, 100 % pool, paired component bootstrap, signs as H1 - A):- EXPAND: upper bound of the relative-L2 difference < 0.001, and either the last-fill distance/Rg difference has its upper bound < 0 or the Kendall tau difference has its lower bound > 0.
- PRUNE: lower bound of the relative-L2 difference > 0 and neither of those improves clearly.
- PARK otherwise.
- Topology metrics and v4 results are descriptive only.
- B8/B9 publication fixed at
1153cd8c.
- 2026-11-17: B9 full-library schema scan (MeltAssist; 1253/1253 files parsed, 0 failures, no numeric decoding). Freeze_time and shrinkage are present in every v4 and v5 part. Temperature, pressure, velocity and deflection are present in 621 of 629 v4 parts (the 8 missing ones are listed) and in all 624 v5 parts. Trajectories have 99 to 110 frames. v5 also carries cell residual stress. This shows the schema exists; the values have not yet passed any eligibility check. Opened B10 as PARKED, behind fill-time priority.
- 2026-11-17: B8 PARKED (MeltAssist; 4 Range reads, 77,714 B). Reopen if the companion LOG/YML is obtained or the VTU carries equivalent pre-solve metadata. Narrower reopen path (Claude): the 42 parts may still serve as a fresh hold-out for scale-free metrics only (Kendall tau, last-fill location, H0(-T) events), which need no t_set. They then need only field validity and the near-duplicate check, not process provenance. Narrowed by MeltAssist: the models still need the gate as an input, through D_i, D_f and F. Scale-free scoring is therefore not gate-free inference. Without equivalent metadata in the VTU or an original gate source, reopening needs a separate task built only from the available fields; recovering the gate from the minimum of the true T would be label leakage.
- 2026-11-17: MoldAstra B1 short spec + H0(-T) metric (Drive 1WHFdSP7…, SHA 91ecc62f…); one-page maths judgement (Drive 1I_H_UxU…: an order-preserving clock change cannot move ranking events; in a simplified constant-pressure counter-example, along-path loading gives a 4x fill-time difference that vanishes at constant flow); gate specs for B3/B4/B5 (Drive 1LUaoJn9…) approved as written. MeltAstra B1 contract draft
0a60f5d6plus a one-page physics finding (gate reachability and early wet islands are more direct than wet-region H0 merges; the volume quantile is a shared progress variable, not the full hydraulic state). B3/B4/B5 move to PROBING once the B1 contract is frozen. - 2026-11-17: tree started; B1/B2 promoted on Yanqi's approval; B3-B9 opened.