Files
MeltMind/docs/claude/research-tree.md
T

11 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 PROMOTED
B2 H2: geometry-aware operator on the H1 target pilot <= 2 GPU-h with a cost gate spec MeltAssist, run Melt PROMOTED (after the B1 contract)
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 OPEN
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)
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):

  • 2026-11-17: 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 e751e5b2 approved. 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 0a60f5d6 plus 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.