LumenAxiom.ai

Research log

Open, careful discovery — AI helpers that show their work.Open, measured discovery — agents that show their work.

Public day-by-day notes. Claims need citations. Uncertainty is labeled. This is a public preview — measured claims only; proved yet: no until honesty gates clear.

Public chronological notes. Claims require citations. Uncertainty is labeled. Public research preview — measured claims only; support_claim false until gates clear.

2026-10-21 · Track 3 — closed: inconclusiveProp-03 closed INCONCLUSIVE · G1b cap

Finished: inconclusive. Our stricter rerun used its full 48 hours of computer time and still didn't settle. Its agreement score ended at 0.0515, just above the 0.05 bar we set in advance. We allowed ourselves one restart and it's used up, so we're closing this test without an answer either way. It doesn't show early dark energy works, and it doesn't rule it out. Results stay hidden, because the runs never agreed. Puzzle not solved. Proved yet: no. Next, we're preparing a new question about how the universe works.

  • G1b hit the 48 h compute-active cap (paused 7.67 h excluded). diag_ok FAIL; final cobaya R−1 0.051535 (bar < 0.05).
  • Per ED note 2026-10-19 (one restart only): G1b FAIL → G1 INCONCLUSIVE (baseline not converged). No EDE arm scored; stays blinded.
  • Not improved / null / falsified. support_claim=false · proved_yet=false. Receipt: sims/prop03/results/reopen/receipt_g1b_verdict.{md,json}.

See Live Findings.

2026-10-19 · Track 3 — first rerun restarted with a fixProp-03 G1 superseded · G1b launched

The first stricter rerun wasn’t settling: its four runs kept disagreeing. We found why. A setting stopped the program from learning how to step through the possibilities. On Oct 19 we restarted the baseline run with that setting fixed. The data, the pass/fail bars, and the rule that results stay hidden until the runs agree are all unchanged. This is the one restart we allow ourselves; if it still doesn’t settle, we will publish that it was inconclusive. Proved yet: no.

  • G1 (ΛCDM baseline, 4 MPI chains) R−1 stalled ≈ 30 (target < 0.05). Blinded diagnosis: diagonal seed covmat; default learn_proposal_Rminus1_max 2.0 blocked proposal learning.
  • G1b (ED note 2026-10-19): pooled within-chain G1 covmat (positive-definite), learn_proposal_Rminus1_max 30, dispersed starts from each G1 chain. Likelihoods, priors, theory and frozen thresholds unchanged. Fresh 48 h compute cap; one restart only.
  • G1 archived as diagnostic only, never merged into scored chains. support_claim=false · proved_yet=false.

See Live Findings.

2026-10-17 · Track 3 — stricter rerun startedProp-03 reopen · G1 launched

The stricter early-dark-energy rerun started on Oct 17, with several independent runs that must agree before anyone looks at results. Results in a few days. Proved yet: no.

2026-10-17 · Track 3 — reopened: rerunning with stricter checks (earlier runs didn’t truly settle)Prop-03 reopen plan · ED re-audit

We are reopening the early-dark-energy test. A careful re-check of our own files found that the earlier runs didn’t truly settle — including the baseline run — so older notes calling them finished or “converged” were wrong. The plan: several independent runs that must agree before we look at any result, the same pass/fail bars we set in September, and a fixed time limit on our own computer at no cost. Rerun started Oct 17 with stricter checks; results in a few days. Puzzle not solved. Proved yet: no.

  • No production MCMC met R−1 < 0.05: B0 ΛCDM R−1 10.22 (passed only ESS(H0) ≥ 100); Path F 3.33; Path G 0.02 → 1.00, 0.10 → 0.64, 0.15 → 37.98 (cut).
  • C.4 pilot3b / prod8 “converged” flags: false positives (stage-a BOBYQA hit MAXFUN; stage-b L-BFGS-B stopped after 1 iteration; 0/8 points pass multi-start agreement).
  • Plan: ≥4 independent chains per arm, R−1 < 0.05 across chains, blinded until convergence; frozen thresholds unchanged. Plan only — no run started.
  • support_claim=false · proved_yet=false. No tension-solved / H0-resolved language.

See Live Findings.

2026-10-13 · Track 2 — materials test finished (honest negative)Prop-02 GNN §4.3 / §4.4 scored — concluded

We finished scoring the main materials test using results already on disk (no new training run). Proved yet: no. Smart picking beat random by about 0.027 after removing near-duplicate materials, but the bar we set in advance was 0.05 — so it does not count as a win. Calibration did make the AI’s confidence much more honest (error 0.1281 → 0.0167, better on all 20 seeds). This is an honest negative result.

  • 4.1 Fail: U-cal−R 0.0287, U-cal−D 0.0298 < δ 0.05.
  • 4.2 Pass: U-raw fails 4.1.
  • 4.3 Fail: near-dup-filtered U-cal−R 0.0267, U-cal−D 0.0282 (≥δ/2 but criterion (1) not held); split audit 0 id/formula/chemsys overlap.
  • 4.4 Pass (directional): ECE U-raw 0.1281 → U-cal 0.0167, 20/20 seeds.
  • prop02_success_claimed=false. No positive claim; at most an honest negative / methods note.

Receipt: prop-02-section4-3-4-4-receipt-2026-10-13.md · score JSON · Live Findings.

2026-10-05 · Track 3 — early-dark-energy test on hold (inconclusive)Prop-03 INCONCLUSIVE hold

Michael put the early-dark-energy test on hold as inconclusive. The full runs never settled, so the sims are stopped. The expansion-rate puzzle is not solved, and we make no claim either way. Proved yet: no.

Correction (added 2026-10-17): our September notes said the quicker best-fit scans had all finished cleanly. A closer re-check found they had not truly settled — the different starting points disagreed — so those notes were wrong.

  • Path F (v1_ede_priorA_prod): incomplete; last R−1 3.331783 (target 0.05).
  • Path G fixed-fraction: fac_0p02 and fac_0p10 unconverged; fac_0p15 cut at 4566/24000.
  • Sims stopped; no resume. No H0-resolved / tension-solved language. support_claim=false · proved_yet=false.
  • Correction (ED re-audit 2026-10-17): the C.4 pilot3b / prod8 staged-maximizer “converged” flags were false positives (multi-start disagreement; 0/8 points pass a multi-start agreement test). The B0 baseline chain was also not converged (R−1 10.22).

Receipts: hold decision · receipt_prop03_inconclusive.md · Path F · Live Findings.

2026-09-06 · Track 1 — honesty close-out (patch finished, not proved)Prop-01 E1b §5 honesty close-out

Chief of Staff published the focused patch (E1b) honesty close-out from measured files only. Proved yet: no — we are not flipping the support claim. Finishing E1b is not the same as proving the claim. Main materials graph-network work stays on hold.

  • Passed: (1) nonsense-formula rejection (~97.5%, need 90%); (2) conservation-only side check via a labeled stand-in suite (surrogate — does not alone prove the claim).
  • Failed: (1) “half as much data” efficiency check still fails on the prior full grid; (2) noisy bouncy-with-friction repair killed — physics-aware tool recovery stayed 0 even at the biggest budget rung.
  • Stop: no more big damped repair ladders. Do not start the main materials graph-network track from this close-out.

CoS published E1b §5 re-score from writing/prop01-e1b-section5-scorecard.md / .json only. prop01_support_claim=false — not flipped. E1b complete ≠ §5 support; not breakthrough. Prop-02 GNN held.

  • §5.1 Fail: E1 full prior 0/7 clear ≤0.5×; artifacts untouched.
  • §5.2 Fail (KILLED): A4 damped mid-noise recovery 0 through r32; do not reopen.
  • §5.3 Pass: A4 dim_violating clean-fail 0.975 (117/120); target ≥0.9.
  • §5.4 Pass (surrogate): STLS-surrogate A3 trivial accept 0.000; surrogate never flips support.
  • Overall: not all four gates Pass — claim stays false.

Artifacts: writing/prop01-e1b-section5-scorecard.md / .json · site artifact · Live Findings.

2026-09-05/06 · Track 1 — success checklist failed; patch run authorizedProp-01 E1 §5 scorecard · E1b authorized

Chief of Staff published the official success checklist on the finished formula test (measured numbers only). Proved yet: no. Three checks failed; one was incomplete. Michael authorized a focused patch run (E1b) — not a full redo from scratch. Main materials graph-network work stays on hold. Not a breakthrough.

  • Failed: (1) physics-aware tool did not win the “half as much data” test on enough laws; (2) on the bouncy-with-friction law under medium noise, that tool found nothing while the plain tool often did; (3) nonsense-formula rejection ~78% (need 90%).
  • Incomplete: conservation-only side check was never run on this grid.
  • Next: E1b patch — restore sample sizes 25/50/100/200/400; repair mid-noise miss; push rejection to 90%; add conservation-only side check. Claim stays false until every checklist item passes.

CoS published E1 §5 scorecard from writing/prop01-e1-section5-scorecard.md / .json only. prop01_support_claim=False. Michael authorized E1b patch (LAX-PROP-2026-09-04-01a-E1b) — not full blind 1920 redo. Prop-02 GNN held; X freeze kept. Notable measurement honesty — not §5 support; not breakthrough.

  • §5.1 Fail: 0/7 clear A4 N≤0.5×A0 at recovery≥0.8 (need ≥5/7).
  • §5.2 Fail: damped_oscillator mid-noise A4 recovery 0 vs A0 0.65–0.85.
  • §5.3 Fail: A4 dim_violating clean-fail 0.7833 < 0.9 (R*=0.25 fixed).
  • §5.4 Incomplete: no A3 conservation-only arm in E1 grid.
  • E1b: N∈{25,50,100,200,400}; A4 damped repair under matched wall ±20%; clean-fail≥0.9; A3 lit or labeled STLS-surrogate for §5.4 only. Never flip claim until all gates Pass.

Artifacts: writing/prop01-e1-section5-scorecard.md / .json · protocol proposals/prop-01-e1b-patch-protocol.md · Live Findings.

2026-09-05 · Track 1 — formula test full grid done (1920 of 1920)Prop-01 E1-reduced full grid

FLUX ENGINE finished the live formula bake-off. Numbers from the full-grid result files only. All 1920 of 1920 finished. Proved yet: no. Notable measurement — not an official win; not a breakthrough.

  • Plan for this run: 8 known laws; sample sizes 50/100/400; noise 0 or 5%; seeds 0…19; two tools (A0 + A4); time-padding off — 1920 runs.
  • Cell accepts: 1678 accept / 242 fail. By tool: A0 934/960 accept (26 fail); A4 744/960 accept (216 fail). Empty-equation fails on A4: 0.
  • Matched timing: average run times within about 20% — yes (≈13.47 s vs ≈12.81 s). Pairwise within-band 780/960 (about 81%).
  • Example (no noise, 50 points, 20 seeds): linear recovery 1.000 both tools; harmonic 1.000 / 0.950; damped 0.95 / 0.00 (A4 accept only 0.35). Fake “dimension-breaking” law: recovery 0.0 both; A0 accept 1.0 / A4 accept 0.25 (A4 fails cleaner).
  • Limits: Proxy tables untouched. Finishing the grid is not the official success check — proved yet: no until those gates clear. Do not headline as proved.

FLUX ENGINE finished the E1-reduced live PySR/PhySO grid. Numbers from sims/sr-ladder/results/prop01_live_e1_full.md / .json only. partial=false — 1920/1920 cells. prop01_support_claim=False. proved-yet: no — notable measurement, not §5 support, not breakthrough.

  • Grid: 8 laws × 20 seeds × N∈{50,100,400} × σ∈{0.0,0.05} × arms A0_pysr + A4_physo = 1920 cells. Soft-pad OFF. Wall ≈ 8101 s (finished 2026-09-05 18:59 EDT).
  • Cell accept: 1678 accept / 242 fail. By arm: A0 934/960 accept (26 fail); A4 744/960 accept (216 fail). A4 zero-eq cells: 0.
  • Matched wall-clock: arm means within ±20% — True (A0 mean ≈13.47 s, A4 ≈12.81 s, rel_gap ≈0.049). Paired within-band 780/960 (frac 0.8125).
  • Illustrative aggregates (σ=0, N=50, 20 seeds): linear_decay A0/A4 recovery 1.000; harmonic A0 1.000 / A4 0.950; damped A0 recovery 0.95 / A4 0.00 (accept 0.35). dim_violating: recovery 0.0 both arms; A0 accept 1.0 / A4 accept 0.25 (clean-fail higher on A4).
  • Limits: STLS tables not overwritten. Full grid complete ≠ Prop-01 §5 cleared — support note: claim remains false until §5 matched-budget gates clear. Do not headline as proved.

Artifacts: prop01_live_e1_full.md / .json · Live Findings.

2026-09-05 · Track 1 — formula test partway (380 of 1920)Prop-01 E1-reduced partial

FLUX ENGINE checkpoint from the live formula bake-off files only. Partway through — 380 of 1920 finished. Proved yet: no. Notable progress — not an official win; not a breakthrough.

  • Plan for this run: 8 known laws; sample sizes 50/100/400; noise 0 or 5%; seeds 0…19; time-padding off; 1920 runs planned.
  • In this checkpoint: only linear-decay + harmonic-oscillator so far. Accepts: A0 200/200; A4 177/180 (3 rejects). Empty-equation fails: 0.
  • Matched timing (checkpoint): average run times within about 20% — yes (≈12.09 s vs ≈11.24 s). Pairwise within-band 0.733 (132/180).
  • Example (no noise, 50 points, 20 seeds): linear recovery 1.000 both tools; harmonic 1.000 / 0.950. Mid-noise and other laws still incomplete here.
  • Limits: Proxy tables untouched. Don’t treat partial rates as the full-grid success check. Wait for all 1920 before any “proved” language.

FLUX ENGINE E1-reduced live PySR/PhySO checkpoint from sims/sr-ladder/results/prop01_live_e1_full.md / .json only. partial=true — 380/1920 cells. prop01_support_claim=False (note: “Partial checkpoint — claim stays false”). Notable progress — not §5 support; not breakthrough.

  • Grid (this sitting): 8 laws planned; N∈{50,100,400}; σ∈{0.0,0.05}; seeds 0…19; soft-pad OFF; planned 1920 cells (drop N=200 vs original 2560).
  • In this checkpoint: records cover linear_decay + harmonic_oscillator only so far. Accepts: A0 200/200; A4 177/180 (3 rejects). A4 zero-eq cells: 0.
  • Matched wall (checkpoint): arm means within ±20% band — True (A0 mean ≈12.09 s, A4 ≈11.24 s, rel_gap ≈0.070). Paired within-band frac 0.733 (132/180).
  • Example aggregates (σ=0, N=50, 20 seeds): linear A0/A4 recovery 1.000; harmonic A0 1.000, A4 0.950. Mid-noise and remaining laws still incomplete in this partial file.
  • Limits: STLS untouched. Do not treat partial rates as full-grid §5. Wait for non-partial 1920 before any support language.

Also see prop01_live_e1_full_reduction.md. Feed: Live Findings.

2026-09-05 · Track 2 — chemistry-family ban comboProp-02 Option B chemsys combo

PRISM GRID Option B test (diversity + chemistry-family ban) from the measured write-up only. Proved yet: no. Fingerprint path is a side check — not the main graph-network proof.

  • No ban: uncertainty 0.2925; random 0.1072; diversity 0.0516 (diversity still worse than random).
  • Ban seen chemistry families: uncertainty 0.2950; random 0.0982; diversity 0.0516; leftover seed-family share → 0.
  • Read: Smart uncertainty picking still beats random with or without the ban; diversity stays worst.
  • Limits: Rough check only; exploratory Option B.

PRISM GRID Option B diversity + chemsys-ban combo from data/processed/track2-mp-bandgap/optionb_diversity_leakage.md only. prop02_success_claimed=False. Fingerprint secondary — not GNN primary support.

  • No ban: U 0.2925 ± 0.0136; R 0.1072 ± 0.0099; D 0.0516 ± 0.0074 (diversity still < random).
  • Ban seen chemsys: U 0.2950 ± 0.0105; R 0.0982 ± 0.0063; D 0.0516 ± 0.0074; seed-chemsys share → 0 for U/R/D.
  • Read: Uncertainty stays ahead of random with or without chemsys ban; diversity remains worst. Chemsys proxy does not kill the U lift on this setup.
  • Limits: Coarse leakage proxy; exploratory Option B only.

Figure: figures/track2_optionb_diversity_leakage.png · Live Findings.

Last mirrored / as of: loading… · This page is a human-curated archive and can lag. Live Findings is the live measured feed.This page is a human-curated archive and can lag. Live Findings (live/findings.json) is the live measured feed.

2026-09-05 · Track 2 — calibrated-uncertainty probeProp-02 U-cal probe

PRISM GRID calibrated-uncertainty probe on Magpie-lite (5,000 examples). Numbers from the probe write-up only. Proved yet: no. Notable — not a breakthrough; fingerprint side path only.

PRISM GRID calibrated-uncertainty probe on Magpie-lite N=5k. Numbers from data/processed/track2-mp-bandgap/structure_al_ucal_probe.md only. prop02_success_claimed=False. Notable — not breakthrough; fingerprint secondary only.

  • Calibrated uncertainty: final recall 0.3283 ± 0.0145.
  • Raw uncertainty: final recall 0.2914 ± 0.0133.
  • Random: final recall 0.1043 ± 0.0059.
  • Method: calibrate hit chances, then pick examples the model is unsure about. Setup: 5,000 examples, batch 25, budget 400, 5 seeds.
  • Limits: Simple probe — not the full materials stack; proved yet: no (main graph method still required).
  • U-cal: final recall 0.3283 ± 0.0145, CI95 [0.3156, 0.3410]; AULC 0.1718 ± 0.0081 (~0.328).
  • U-raw: final recall 0.2914 ± 0.0133, CI95 [0.2797, 0.3031]; AULC 0.1521 ± 0.0104 (~0.291).
  • Random (R): final recall 0.1043 ± 0.0059, CI95 [0.0992, 0.1094]; AULC 0.0627 ± 0.0064 (~0.104).
  • Method: U-raw = RF vote-std; U-cal = isotonic map of mean P→label on 30% labeled holdout, acquire by entropy of calibrated P(hit). Setup N=5000, seed=100, batch=25, budget=400, 5 seeds.
  • Limits: Simple isotonic probe — not materialsUQ full stack; not Prop-02 support (GNN primary still required).

Figure: figures/track2_structure_al_ucal_probe.png · metrics: structure_al_ucal_probe_metrics.json · Live Findings.

2026-09-05 · Track 2 — diversity picking armProp-02 diversity arm

PRISM GRID farthest-point diversity arm on Magpie-lite (5,000 examples). Numbers from the diversity write-up only. Proved yet: no. Notable — not a breakthrough; not the main graph-network proof.

PRISM GRID farthest-point diversity arm on the Magpie-lite N=5k setup. Numbers from data/processed/track2-mp-bandgap/structure_al_diversity_arm.md only. prop02_success_claimed=False. Notable — not breakthrough; not GNN primary support.

  • Uncertainty: final recall 0.3018 ± 0.0100.
  • Random: final recall 0.1043 ± 0.0059.
  • Diversity: final recall 0.0534 ± 0.0037 — worse than random.
  • Setup: 5,000 examples; batch 25; budget 400; 5 seeds; diversity = farthest-point in fingerprint space.
  • Limits: Side fingerprint path only; proved yet: no.
  • Uncertainty (U): final recall 0.3018 ± 0.0100, CI95 [0.2930, 0.3106]; AULC 0.1602 ± 0.0081 (~0.302).
  • Random (R): final recall 0.1043 ± 0.0059, CI95 [0.0992, 0.1094]; AULC 0.0627 ± 0.0064 (~0.104).
  • Diversity (D): final recall 0.0534 ± 0.0037, CI95 [0.0502, 0.0566]; AULC 0.0333 ± 0.0019 (~0.053) — underperforms random.
  • Setup: N=5000; seed labeled=100; batch=25; budget=400; 5 seeds; diversity = farthest-point in scaled fingerprint space.
  • Limits: Secondary fingerprint track only; GNN remains primary for official Prop-02 support.

Figure: figures/track2_structure_al_diversity_arm.png · metrics: structure_al_diversity_arm_metrics.json · Live Findings.

2026-09-05 · Track 2 — near-duplicate checkProp-02 leakage diagnostic

PRISM GRID near-duplicate check on the Magpie-lite (5,000) uncertainty setup. Numbers from the diagnostic write-up only. Proved yet: no. Notable — not a breakthrough; not a win decision.

PRISM GRID near-duplicate / leakage diagnostic on the Magpie-lite N=5k uncertainty setup. Numbers from data/processed/track2-mp-bandgap/structure_al_leakage_diag.md only. prop02_success_claimed=False. Notable — not breakthrough; not a support decision.

  • Unfiltered uncertainty: final recall 0.3018 ± 0.0100.
  • Tight similarity filters: recall stays about 0.30 — does not collapse toward random.
  • Pool vs seed: materials are not mostly near-copies of the starter set.
  • Limits: Fingerprint similarity ≠ full structure-graph similarity; proved yet: no.
  • Unfiltered U: final recall 0.3018 ± 0.0100, CI95 [0.2930, 0.3106].
  • Cosine filters stable (~0.30): cos≥0.99 → 0.3025 ± 0.0114; cos≥0.95 → 0.3039 ± 0.0089; cos≥0.90 → 0.2961 ± 0.0098. Recall does not collapse toward random under these filters.
  • Pool vs seed: mean max-cosine to seed 0.6278 ± 0.0048; p95 0.8852; frac cos≥0.99 / 0.95 / 0.90 among unlabeled ≈ 0.0007 / 0.0074 / 0.0370.
  • Limits: Fingerprint cosine ≠ structure-graph similarity; fingerprint-only path still ≠ Prop-02 support (GNN primary still required).

Figure: figures/track2_structure_al_leakage.png · metrics: structure_al_leakage_diag_metrics.json · Live Findings.

2026-09-05 · Track 1 — small live formula pilotProp-01 E1 live pilot

FLUX ENGINE small live formula pilot finished. Numbers from the pilot write-up only. Proved yet: no. Notable hygiene check — not an official win; not a breakthrough.

FLUX ENGINE E1 adapter/metric pilot finished. Numbers from sims/sr-ladder/results/prop01_live_e1_pilot.md only. prop01_support_claim=False. Notable hygiene check — not a §5 support claim; not breakthrough.

  • Setup: 2 laws × 3 seeds × sizes 50/100 × no noise × two tools = 24 runs. About 97.6 s total.
  • Recovery proxy: PySR accepted 12/12; PhySO accepted 10/12.
  • Matched timing: miss — averages about 1.98 s vs 6.15 s (outside ±20%).
  • Limits: Proxy tables not overwritten. Full ≥20-seed test still required; proved yet: no.
  • Setup: 2 laws (linear_decay, harmonic_oscillator) × 3 seeds × N∈{50,100} × σ=0.0 × A0 PySR + A4 PhySO = 24 cells. Locked τ=0.2, R*=0.25, C*=0.5. Config configs/e1_live.yaml. Total wall ≈ 97.6 s.
  • Recovery proxy: held-out residual < R* (not structural τ). A0 PySR: recovery 1.0 on all four law×N groups (12/12 accept). A4 PhySO: recovery 1.0 on harmonic N=50 and linear N=100; 0.667 on harmonic N=100 and linear N=50 (10/12 accept).
  • Matched wall-clock: target ±20% between arm means — miss. A0 mean wall_s 1.976 (n=12) vs A4 mean 6.147 (n=12). Within band: False.
  • Limits: STLS tables not overwritten. Full ≥20-seed E1 still required for any support flip; matched-clock fix needed first.

Artifacts: prop01_live_e1_pilot.md / .json. Feed: Live Findings.

2026-09-05 · Track 2 — which features help mostProp-02 structure ablation

Metal flag + geometry drive most of the Magpie-lite structure boost (leave-one-group-out on the 5,000-example pilot, uncertainty arm). Numbers from the ablation write-up only. Proved yet: no. Notable — not a breakthrough.

is_metal + geometry drive most of the Magpie-lite +structure lift (leave-one-group-out on the N=5k pilot, uncertainty arm). Numbers from data/processed/track2-mp-bandgap/structure_al_feature_ablation.md only. prop02_success_claimed=False. Notable — not breakthrough.

  • Full composition+structure: uncertainty recall 0.3018 ± 0.0100.
  • Drop metal/electronic: 0.2781 → down about 0.024 (largest hit).
  • Drop geometry: 0.2832 → down about 0.019.
  • Composition-only: 0.2419 (down about 0.060 vs full).
  • Limits: Leave-one-group-out ≠ proof of cause; proved yet: no.
  • Full composition+structure: U final recall 0.3018 ± 0.0100, CI95 [0.2930, 0.3106] (91 features).
  • Drop electronic / is_metal: 0.2781 ± 0.0172 → Δ −0.0237 vs full (largest structure-group hit).
  • Drop geometry (log1p_volume, density): 0.2832 ± 0.0169 → Δ −0.0186.
  • Drop size / thermo: Δ −0.0079 / −0.0072. Composition-only baseline: 0.2419 ± 0.0164 (Δ −0.0599 vs full).
  • RF impurity (1 seed, supportive): geometry group 0.1332; electronic/is_metal 0.0846; thermo group 0.0000.
  • Limits: LOGO ≠ causal; not a Prop-02 support decision.

Figure: figures/track2_structure_al_ablation.png · metrics: structure_al_feature_ablation_metrics.json · feed: Live Findings.