# Analysis chronology and prospective freeze protocol

## Retrospective chronology

| Date | Artifact or decision | Role | Was it untouched evaluation? |
|---|---|---|---|
| 2026-07-06 | `human_filtered_dmd_gene_consensus_signature.tsv` | Initial integrated disease-direction score | No; later re-audited for source sensitivity |
| 2026-07-06 | STATE readiness and candidate inference queues | Environment and exploratory triage preparation | No |
| 2026-07-10 13:11 +0800 | `hepg2_candidate_screen_queue_v3_classical_manifest.tsv` | Feasible candidate queue | No; created before the external fusion audit |
| 2026-07-10 | GPU closure, direct-head results and seen-delta positive control | Model-development closure | No; development evidence |
| 2026-07-12 18:35 +0800 | GSE293514 fusion summary and candidate safety flags | External muscle-context risk screen | No transcriptional endpoint; not used to create the earlier feasible queue |
| 2026-07-12 to 2026-07-16 | Five broad held-out-gene splits | Retrospective internal stability analysis | No; genes are reused across test splits |
| 2026-07-13 to 2026-07-14 | Official GEARS and scGPT frozen-split runs | Task-specific model comparison | Frozen within this benchmark, but not an untouched project lockbox |
| 2026-07-16 11:09 +0800 | Final candidate decision matrix | Integrates earlier queue with later risk/context flags | No; descriptive triage synthesis |
| 2026-07-16 | DMD source-sensitivity audit and tier multiverse | Robustness analysis | No; prompted by review |
| 2026-07-16 | Split-overlap, measurement-reliability and leave-HepG2-out DepMap audits | Major-revision sensitivity analyses | No; explicitly retrospective |

The timestamp order shows that GSE293514 did not generate the 10 July feasible queue. It was added later as a risk filter. This removes one specific circularity concern but does not turn the final 16 July decision matrix into an independent validation result.

## Prospective freeze protocol

Before collecting or opening a new confirmatory dataset, freeze and checksum:

1. dataset accession, eligibility rules, target identifiers and exclusions;
2. source-specific DMD effects and, if estimable, the random-effects model;
3. model source commits, checkpoint hashes, feature versions and prohibited features;
4. a single immutable lockbox or an external dataset that no model/tier decision has seen;
5. grouped split rules, including any gene-family/pathway grouping;
6. primary raw-direction endpoints, secondary RMSE endpoints and multiplicity family;
7. minimum per-target replicate/guide support and reliability handling;
8. coverage denominator and common/native coverage reporting;
9. candidate evidence/action rules without post-outcome threshold changes;
10. stopping rules, failure reporting and the claim ladder.

Any post-freeze deviation must be versioned, justified and labelled exploratory. A newly frozen subset of the already reused 2,160 genes must not be called a retrospective lockbox.