docs(validation): re-score prospective N=28 on the current engine (DE-54) — no OOD improvement#102
Conversation
…-54) Closes the README 'predates FLUX-1, not re-scored on the current engine' note on the 2026-06-01 prospective set with an actual number, and records the result. Re-ran the 28-drug public-clone prospective set through the current engine (HEAD f8a4663: FLUX-1 + CLF leak-free + UGT single-path + RBP honest-1.0) via scripts/score_prospective_candidates.py. Result: Meta AAFE 3.2084 -> 3.2861 (+0.078), engine 4.3017 -> 4.5506 (+0.249), ML 3.2704 -> 3.2704 (per-drug bit-identical). Deep inside the N=28 CI [2.42, 4.37] -> no improvement, statistically indistinguishable. In-domain (N=16) meta 3.1988 -> 3.3177. The comparison is clean: the ML track being per-drug bit-identical proves the old artifact shares this local numerics stack, so the delta isolates engine code. Mechanism: prospective is under-predicted (geobias 0.50); FLUX-1 correctly raised extraction -> lowered Cmax -> removed the E-cap's compensating over-estimate -> deepened the under-call. Residual OOD deficit is absorption (fa) dispersion (DE-42 floor), not extraction. Adds the re-scored artifact + DE-54 (dead-ends) + a 2026-07-05 experiment-log entry + a diagnosis section 8 confirmation. Headline 2.743 untouched.
… prospective section Keeps the published 3.21 headline; adds that re-scoring the N=28 set on the current engine gives 3.29 (no improvement, within CI), with the error-cancellation mechanism (extraction fix deepens the fa-limited OOD under-call).
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 34ff70d904
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| **2026-06-03 — the prospective set is also co-calibration-locked; the meta damps the engine to ~18% everywhere (dead-ends.md DE-43).** The natural follow-up — "the *prospective* N=28 set (3.21) isn't in the meta co-calibration, so a first-pass lever foreclosed retrospectively should help it" — was tested and **falsified**. A `F = fa·Fg·Fh` decomposition shows the prospective catastrophes (mirdametinib, sevabertinib, kinase inhibitors) are **fa-first** (absorption starved: low Peff or low-solubility → `particle_radius=50µm` → `ka ≪` gut transit) then **gut-CYP3A `Fg`-second** — *not* the fa-saturated pure-CYP3A mode. Two levers (absorption scalar 5.25×; gut-CYP3A 0.5×) each improve the **engine** track on prospective (4.11→3.75 / 4.11→4.00; mirdametinib engine fold 58→13) — but the fixed-weight **meta-learner damps the change to ~18–19% pass-through, identically on prospective and retrospective**, so the production meta nets neutral-to-negative (absorption scalar: prospective −0.069 *vs* retrospective +0.082 = **net −0.012**; gut-CYP3A: +0.020, inside the N=28 CI). **The meta is robust to engine errors by construction, which symmetrically blocks engine improvements — on every benchmark.** This is the unifying mechanism behind the error-cancellation dead-ends (§2): the engine is structurally not a headline lever, retrospective *or* prospective. The only remaining levers are per-drug **measured-F routing** (a capability, not a global move) or a **meta-architecture change** (AD-gated engine weighting — itself likely foreclosed per the meta-learner dead-ends DE-23/24/25 and the non-generalizing AD signal DE-41, and N=28-underpowered to validate). | ||
|
|
||
| **2026-07-05 — confirmed by the actual full-set re-score (dead-ends.md DE-54).** The 2026-06-03 prediction above was, until now, an extrapolation from a decomposition; it has been tested directly. The N=28 prospective set, re-scored end-to-end on the post-FLUX-1 engine (+CLF/UGT/RBP; HEAD `f8a4663`), moves **Meta AAFE 3.21 → 3.29** and **engine 4.30 → 4.55** — the engine's +0.25 damped to +0.08 at the meta, exactly the ~18% pass-through predicted, and *worse* not better. The comparison is clean: the ML track is per-drug bit-identical (max abs diff 0.0), so the 2026-06-01 artifact shares this local numerics stack and the delta isolates engine/predict code. Mechanism: prospective is under-predicted (geobias pred/obs 0.50); FLUX-1 correctly raised extraction → lowered Cmax → *removed a compensating over-estimate* (the pre-fix `E`-cap held Cmax too high), deepening the net under-call. The residual is absorption-side `fa` dispersion (the DE-42 floor), not extraction. Artifact: `data/validation/prospective_N28_current_engine_2026-07-05.json`. |
There was a problem hiding this comment.
Fix the pass-through ratio claimed by DE-54
Using the AAFE deltas cited in this same sentence, the current re-score does not show the predicted ~18% damping: 0.0777 / 0.2489 ≈ 31% (and the log-AAFE change is ≈43%). Since this paragraph records DE-54 as a full-set confirmation of the DE-43 mechanism, future research decisions will be gated on an overstated mechanism unless the text either computes the pass-through on the intended per-drug/weight basis or stops claiming the aggregate result is “exactly” ~18%.
Useful? React with 👍 / 👎.
What
Closes the README "predates FLUX-1, not re-scored on the current engine" note on the 2026-06-01 prospective set with an actual measurement, and records the (negative) result across the research docs.
Result
Re-ran the 28-drug public-clone prospective set through the current engine (HEAD
f8a4663: FLUX-1 + CLF leak-free + UGT single-path + RBP honest-1.0) viascripts/score_prospective_candidates.py:Deep inside the N=28 CI [2.42, 4.37] → statistically indistinguishable, no improvement.
Why it matters (clean comparison + mechanism)
E-cap held Cmax too high (offsetting thefaunder-call); FLUX-1 correctly raisedE→ lowered Cmax → deepened the under-call (concentrated in high-extraction drugs, e.g. arimoclomol engine fold 4.96→16.10). This is the first actual full-set confirmation of the 2026-06-03 diagnosis §8 prediction / DE-43: the fixed-weight meta damps the engine's +0.25 to +0.08. Residual OOD deficit is absorption-side (fa) dispersion (DE-42 floor), not extraction.Changes
data/validation/prospective_N28_current_engine_2026-07-05.json(full per-drug + provenance + comparison block).dead-ends.md(negative result + telltale).experiment-log.md.diagnosis.md.Headline 2.743 untouched. The local prospective number is directional (sign + mechanism robust); a citable canonical number would need a CI-stack run.
Note
The public README's prospective-row note ("predates FLUX-1, not re-scored") is now stale; a maintainer edit to reference this re-score is provided separately.