docs: final GPU readout in README + model card; add POM relation section

fix: paper_data exporter prefers GPU readout; stamp val_total into GPU evals; stats bundle

feat: GPU readout figures + stats scripts + val-total macros

feat: GPU A/B readout — per-genre prospective figure + refreshed paper_data macros

docs: tidy README + TODO — accurate quick-start, paper package, green tests, READY gate

fix: pre-existing test failures — draft_engine None-MW crash + stale gate test

feat: HF stats-collector job — consolidated paper_stats_bundle.json

feat: paper figures from existing artifacts (ablation bars, POM UMAP, trajectory)

feat: paper_data.yaml exporter for autopaper (49 data_macros)

chore: READY gate state, material_profiles v11_5 refresh, source-population AI review, repair-pairs benchmark

feat: v11.6 repair/enrichment scripts + independent-candidate prep + integrity tests

chore: v11.6 + v11.5 + candidate-20260716a artifact reports and manifests

feat: low-evidence gating in substitution_v11 + wisemoor description provenance

docs: triplet expansion 8->20 + GPU A/B run in README and TODO

fix: GPU job install step — don't shadow PyPI with torch index-url

feat: expand substitution-triplet benchmark 8 -> 20

feat: HF GPU two-arm A/B training job

docs: mark representation-validation sequence complete in TODO

chore: two-arm A/B checkpoints + frozen-benchmark readout -> HF

feat: preregistered frozen-benchmark evaluator (substitution triplets + prospective)

feat: --structural-source flag in eval_v6 for arm-matched evaluation

docs: open design decisions + training-readiness gate in README

feat: wire genuine 256-dim OpenPOM into training (--structural-source)

feat: gate flips READY_FOR_TRAINING (all 5 checks pass)

feat: full molequles multi-tag target space (575-dim ablation arm)

feat: curated perfumer tiers for 15 residual CAS (coverage 99.80%)

feat: prospective formula benchmark (40 in-silico formulas, blinded labels)

feat: complete representation ablation + discordant substitution triplets

feat: link molequles material-profiles odor text into training descriptor matrix

feat: freeze representation validation gate

docs: add representation validation todo

chore: record genre challenge v2 archive manifest

release: freeze blinded genre challenge v2

feat: add leakage-resistant genre baseline benchmark

v11.5 verdict + paper fold: substitution = bounded

v11.5 powered re-run: effective n=14→28, retriever=0/28 [0.000,0.121], physchem=0/28 [0.000,0.121], random=0/28 [0.000,0.121], verdict=underpowered

Complete catalog recovery and enrichment

v11.5 paper fold: substitution = bounded

v11.5 training decision: not-warranted, data is the lever

v11.5 powered re-run: effective n=28, full=1/28 [0.006,0.177], physchem=0/28 [0.000,0.121], verdict=underpowered

v11.5 enrichment: VP/BP 67→340, structure 1731→1741, eval-pairs newly representable=3, leak-rejected=0, formula/MW-rejected=15

v11.3 eval harness: remove unused profile lookup import

v11.3 paper fold: substitution section = inconclusive, profile-coverage limited

v11.3 training scope: not-warranted, independent data coverage is the lever

v11.3 verdict: advantage ambiguous at effective n=14, CIs full=[0.076,0.476] physchem=[0.000,0.215]

v11.3 powered retriever: full=3/14 [0.076,0.476], physchem=0/14 [0.000,0.215], random=1/14 [0.013,0.315], n_excluded=27

v11.3 eval set expanded: n=83 authoritative pairs, shape distribution=41/39/3

results(v11): report preregistered retrieval assessment

fix(v11): preregister mode-aware retrieval evaluation

v11 assessment: retriever top5=0.133, descriptor kept, formula scoped