pino-pimt-representation-ab

PINO PIMT — Representation A/B (Morgan vs genuine OpenPOM vs OpenPOM+all-tags)

Representation ablation for the PINO Physics-Informed Mixture Transformer (PIMT). Three arms trained on the identical corrected corpus (empirical_dataset_v10_alltags.jsonl, molecule-disjoint split, seed-fixed) differing only in the structural embedding block and (for the all-tags arm) the readout width:

Selected via --structural-source morgan|openpom_256|pom_alltags. CAS absent from the genuine-POM asset receive a real Morgan block zero-padded to 256 so the input dim is constant. The genuine-POM asset load fails closed (no silent Morgan substitution).

v2 corrected-label readout (20 epochs, T4 GPU) — authoritative

Trained/evaluated on the v10 corpus with regenerated pyramid targets after a label-corruption defect was found and fixed (see disclosure below). Evaluated once per arm on the frozen benchmarks (labels evaluation_only):

Full per-item detail: frozen_eval_morgan_v2.json, frozen_eval_openpom_256_v2.json, frozen_eval_pom_alltags_v2.json (sha256-pinned). Prospective family-profile agreement is only defined for the 138-dim readout, so it is n/a for the 575-dim all-tags arm.

Honest readout. The three arms are closely matched; no arm dominates. Genuine OpenPOM edges substitution triplets (0.45 vs 0.35/0.35) and the all-tags arm reaches the lowest validation loss, but all arms remain below chance (0.5) on substitution triplets and prospective family agreement is moderate (~0.50). These are genuine negative/mixed results, reported as such rather than spun as a win. The corrected labels substantially improved objective-head convergence (objective loss ~0.04–0.08 vs ~0.37 degenerate under the corrupt labels), which was the primary purpose of the v2 retrain.

Superseded v1 readout (corrupt labels — lineage only, do not cite)

The first GPU run trained on the v9 corpus whose pyramid targets were silently corrupted (30.2% of records collapsed to base-only; tier mass 4.05/2.53/16.67). Its headline numbers (val 0.3959/0.3600, triplets 0.30/0.45, prospective 0.5378/0.6121) were computed against those defective targets and are superseded by the v2 readout above. Retained here for provenance only.

Prior readouts (lineage, superseded)

• Model-free ablation (representation_ablation_*): genuine OpenPOM beats Morgan on every frozen predictive task (5-fold paired bootstrap).
• CPU run (6 epochs, n=8 triplets): OpenPOM ahead on all three — directional only.

Label-corruption disclosure

A defect in the pyramid-target generator (scripts/generate_pyramid_targets.py, now superseded) caused an OAV-decay classifier to default 1,366 of 2,420 CAS to base-only, collapsing 30.2% of v9 pyramid_targets to the base tier. The defect was caught in pre-submission review, the targets were regenerated with a corrected, width-agnostic fold (scripts/regenerate_pyramid_targets_fold.py), producing the v10 corpus (base-only 6.8%, tier mass 4.55/4.19/3.20), and all three arms were retrained from scratch (this v2 readout). A float32 rounding artifact (1.00000012 > 1) in 49 all-tags records additionally trips the CUDA BCE kernel; BCE targets are clamped to [0,1] in src/pino/train.py (621906d). The v9 corpus and its derived results must not be used.

Lineage

• Corpus: 5,708 formulation/trajectory records; 604 pure single-molecule anchors; molecule-disjoint train/validation (0 shared active CAS).
• Triplets: 20 discordant substitution triplets admitted (8 initial + 12 via material_profiles trade-name mining), structural discordance + authoritative preference provenance.
• Prospective: 40 preregistered in-silico formulas from a 51-material hobbyist palette; intended family profiles sequestered (labels.sequestered.json).
• Code: mattbitzesty/pino-source-code. Demo Space: mattbitzesty/pino.

Known limitations

• Substitution-triplet accuracy is below chance (0.5) for all arms at n=20 — a small-n, genuine negative.
• Prospective family agreement is in-silico only (no human sensory panel).
• The single-molecule objective head collapses out-of-domain to a "sweet" attractor; it is intentionally not exposed in the demo and must not be used for single-molecule descriptor lookup.
• The substantivity head is unsupervised (no ground-truth persistence labels).

License: MIT.