[timodonnell]

The released proofreader can identify better next contacts, but its guided rollouts can become shorter, lose coverage, or exploit scoring errors. This experiment compares 15 guidance/search policies, threshold mixtures, fixed-length controls, and ordinary rollout aggregation using complete contact-map metrics and measured GPU costs.

All phases are complete: 1,920 guided trajectories on 32 pilot proteins; ordinary-pool validation on 1,375 structures on H100 and GB200; and one frozen reserved comparison on 1,473 structures in 295 related groups. No guided method or tested threshold mixture improves the measured ordinary-pool quality/cost frontier. The four-rollout 10/25/50% mixture loses 0.02587 R-precision against top-50% alone (95% paired interval −0.03373 to −0.01777).

The selected practical change is 128 ordinary rollouts with votes weighted by squared predicted rollout F1. On the reserved comparison, R-precision improves from 0.54501 to 0.55021: +0.00520 (95% paired group interval 0.00299–0.00767); long-range improves by 0.00419 (0.00199–0.00657). Mean warm GB200 cost is 5.477 versus 3.620 seconds per map. The plan was committed before new test generation; only these two configurations were evaluated, after every input completed. The cohort was previously used for the proofreader release, and generator pretraining decontamination is not established.

Includes a detailed report, 47-page PDF, 27 figures with generating commands, per-protein metrics/timings, immutable execution source bundles, and recovery/compute ledgers. Public artifacts: https://huggingface.co/buckets/open-athena/MarinFold/tree/data/exp368/reports. Report: https://github.com/Open-Athena/MarinFold/blob/codex/exp368-proofreader-guidance/experiments/exp368_evals_proofreader_guided_contact_maps/README.md.

Validation: 11 CPU correctness tests passed; real-model H100 cache and GB200 full-forward replay checks passed at their original tolerances. Failed BF16/GB200 cached-search paths were excluded. All 188,544 reserved trajectories retain identical raw/scored contact and token identities; all 48,030,172 reader contact probabilities are finite and in range; source/plan pins and paired-table identities were verified. All owned GPU jobs are terminal, and native bootstrap resources were removed. Warm compute totals exclude loading, staging and preempted work; operational incidents are disclosed in the execution appendix.

Stacked on #358. Experiment tracking: #368. No experiment issue was closed and no shared cluster configuration was changed.