Negative results and audit findings

1. The generator reward was effectively QED alone#

The docking component of the reinforcement-learning reward never supplied data (stage1_docking_score empty in 100% of 63,441 rows). The whole campaign therefore generated drug-like compounds rather than DHFR-targeted ones. Fixing this requires a new RL campaign on a GPU cluster.

2. The composite score is worse than its own best component#

On the whole-cell axis, the aggregate correlates with activity more weakly than the affinity component alone. This is one of the main findings of the work: multi-criteria aggregates can mask deficiencies in their components. "Fixing" it by replacing the aggregate with affinity alone would remove the argument. Redesigning the weights belongs to the next iteration, after an untouched test set has been collected.

3. The permeability and efflux layer remains an unconfirmed hypothesis#

On the only available whole-cell benchmark, the conceptual differentiator of the system does not differ significantly from a random predictor. The permeability and efflux mappings have never been calibrated against real intracellular accumulation data, because such data does not exist. In score v2 the layer is reported as an annotation and does not enter the aggregate.

4. Stage 2b was meant to be a gate and was an annotation#

A configured retention threshold of P ≥ 0.95 existed, but the filtered result was overwritten in code by the full table. The finding is closed: a counterfactual recomputation and calibration showed that no retention threshold improves agreement between the ranking and experiment (ROC AUC 0.44 on the enzymatic axis, in the opposite direction), so the intended end state of annotation plus input to cellTE is the correct one.

5. The "resistant" phenotype label is a repainted threshold on K_D#

Three numbers from the config reproduce 100% of the stage 6B labels. The class carries no phenotypic information independent of affinity, and it is a property of the chosen concentration rather than of the compound: between the two computed scenarios, 99.5% of compounds change class. That is why the systems biology block does not enter the objective function.

6. A withdrawn xenobiotic metabolism module#

BioTransformer3 and PROXIMAL2 could not generate a sensible metabolic route for E. coli, on four converging lines of evidence. The module was deliberately withdrawn from production rather than masked. A correctly closed negative experiment.

7-9. Silent aggregation defects caught by audit rules#

• A phantom docking block. One fifth of the score weight was empty in every row; an empty component reduces to a neutral element without raising an error. Removed from the config, and the result recomputed bit-identically.
• The systems score was a copy of affinity. Rho = 1.000000. Removed, ranking recomputed, top-10 and top-50 unchanged.
• Misleading stage 5 column names. "Residence time" was a constant simulation length and "koff_proxy" a rescaling of the stability indicator. Both removed from the output.

10-11. Gates computed but not applied#

The MD quality verdict is reported, but at 0.1 ns with a single replica it would have no statistical power, so it blocks nothing. PoseBusters exists only on the benchmark path, so Boltz-2 poses were never filtered for physical plausibility before dynamics. The diagnostics are in place; a live production gate is a priority for the next campaign.

12-13. Stale retrosynthesis ranking and gaps in the funnel#

All retrosynthesis batches refer to a ranking from before the aggregation fix; the staleness caveat is stated in the text. Two unexplained numerical differences between the funnel artifact, the lineage register and the main stage 7 output (345 and 337 records) remain open, reported as such rather than guessed at.