docs(dead-ends): DE-57 — CLint epistemic-uncertainty gate fails (uncertainty-aware meta foreclosed)

[jam-sudo]

Summary

Pressure-tests the last diagnosis.md §9 CLint-floor candidate — calibrated-CLint-uncertainty → uncertainty-aware meta — against the §4 decorrelation gate. It is the one candidate that could structurally sidestep error-cancellation, because it is a per-drug weighting signal (down-weight the engine track where CLint is unreliable), not a point-estimate change. Verdict: FAIL → DE-57. Headline 2.743 untouched (gate pre-check, nothing integrated).

Method

Constructs a genuine per-drug CLint epistemic uncertainty two independent ways and correlates each with per-drug Cmax residuals (Spearman + bootstrap 95% CI, seed 20260422, N=107 from the canonical cache):

• U_boot — SD of log10(CLint) across a K=30 XGBoost bootstrap ensemble on the 1,213-compound hepatocyte training set (non-degenerate, range [0.069, 0.412]).
• U_ad — 1 − max Tanimoto(Morgan) to that training set (applicability-domain distance).

Gate: PASS iff any |ρ| ≥ 0.30 with CI excluding 0.

Result

signal	vs e_eng (engine err)	vs e_eng−e_ml (actionable)	vs r_meta
U_boot	−0.006 [−0.20, +0.18]	+0.119 [−0.07, +0.30]	−0.255 [−0.42, −0.07]
U_ad	+0.145 [−0.06, +0.34]	+0.121 [−0.07, +0.31]	+0.032 [−0.16, +0.23]

• Core null: CLint epistemic uncertainty does not predict engine error, though the engine is the CLint-fed track.
• Actionable null: both independent proxies give ρ≈+0.12 with CI∋0 on e_eng−e_ml — no reweighting signal for the meta.
• The one CI-significant term (ρ(U_boot, r_meta)=−0.255) is sub-threshold and wrong-signed (higher CLint uncertainty ⇒ lower meta error).
• Confound-clean: ρ(U_boot, obs-magnitude)=+0.002.

Why it fails (§10)

The binding wall is bioavailability-F blindness, shared across all four tracks and orthogonal to CLint. Cmax error is F-driven, so no CLint-derived signal can discriminate it — a perfect CLint-uncertainty router would have nothing to route on. Per §4 the gate precedes integration → the uncertainty-aware meta is not built. Distinct from DE-55 (which tested the propagated MC-output uncertainty, ρ=0.039; the shipped CLint CV is the fixed constant _CLINT_CV=1.0).

Changes

• docs/research/dead-ends.md — DE-57 entry
• docs/research/diagnosis.md — §9 open-follow-up marked closed
• scripts/clint_uncertainty_gate.py — reproducible gate (PYTHONPATH=src python scripts/clint_uncertainty_gate.py, deterministic, repo-data only)
• data/validation/clint_uncertainty_gate_2026-07-08.json — result artifact

Sanitization guard passes (711 tracked files); ruff clean.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f24d087440

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".

[chatgpt-codex-connector]

Use a monotone observed-magnitude confound

For this holdout cache, observed Cmax spans 0.00040072–32.9 mg/L, with values on both sides of 1 mg/L, so abs(log10(obs)) ranks both very small and very large concentrations as “large” and treats values near 1 as small. That is distance from 1 mg/L rather than observed magnitude, so the reported obs_mag correlation can look clean even if uncertainty is confounded with high-vs-low exposure; use a monotone quantity such as log10(obs) or raw obs ranks before claiming the DE-57 gate is magnitude-confound-clean.

Useful? React with 👍 / 👎.