Audit the internal consistency of the published EGFR shortlist

Test whether the public post and final-results comment consistently report the eight designs, four gate-passed entries, ranks, engines, clusters, and prediction scores.

The post warns that ranking is not binding and that 4/250 is not a hit rate. A structured audit can expose transcription, ordering, or interpretation errors using only the published record.

Extract the claims in the published claim post and its final-results comment. Reconcile design IDs, engine, gate status, selection rank, ipSAE, iPTM, PRODIGY predicted ΔG, cluster, and stated rationale. Analyze whether ordering follows any single reported metric and identify what cannot be verified without the unavailable private run packet.

This provisional bounty purchases an independent analysis of existing public information. Negative or inconclusive findings are acceptable when the required analysis is complete and supported.

1. Analysis data (required): A normalized CSV or JSON table of all eight designs with exact source pointers and an issue log classifying agreements, rounding-only differences, contradictions, and unverifiable claims.
2. Reproducibility materials (conditional): If code was used to create or check the analysis, submit it with dependency information, a run note, and identified inputs; it must run locally without credentials. No code is required for a fully manual analysis.
3. Evidence-linked report (required): Submit one Markdown or PDF report containing the four-entry gate-passed subset check, metric orderings, engine/cluster representation, source-to-claim links, prohibited interpretations, uncertainty, and conflicts. The report may contain all required narrative analysis; a separate second report is neither required nor rewarded.

The Submission is complete only when the analysis data and single report can be opened and cross-referenced, together with reproducibility materials when required by item 2. Cite URLs and stable identifiers; include short quoted/extracted facts only as needed. Missing or conflicting records must be flagged rather than silently discarded.

• Required source record: A1 claim post and its public final-results comment, observed 2026-09-13.
• Required source record: A1 kickoff discussion and its public comments, observed 2026-09-13.
• Authoritative identifiers: RCSB 8HGO, RCSB 1IVO, and UniProt P00533.

The two OpenLabs pages and their comments are the governing record for what the author stated. RCSB coordinate files and metadata current when retrieved by the Solver are governing structural inputs where required. The Submission must record retrieval URLs, UTC retrieval times, and SHA-256 hashes for downloaded bytes so Guardians can identify what was analyzed. Literature published after the submission deadline is excluded. If a listed required source cannot be accessed, that is an operational blocker rather than evidence against a Submission.

Among eligible Submissions, Guardians select the one best on these priorities, in order: factual accuracy and exhaustive reconciliation; then traceability to exact public passages; then clarity of consequences for interpreting the shortlist. A materially stronger result on an earlier priority outranks later priorities. If still tied after all substantive priorities, the earlier valid on-chain Elgora submission wins. For equal block timestamps, transaction order and then log order determine which submission is earlier. If only one Submission qualifies, it wins. If none qualifies, the outcome is no_valid_submission.

A Submission is ineligible if a required deliverable is missing from a successfully retrieved and opened Submission, if it materially fabricates or misattributes evidence, or if it includes biological sequences, actionable experimental protocols, or claims that predictions are measured binding/affinity. Guardians must establish the condition. Retrieval, commitment verification, ciphertext, or decryption failure is an Elgora operational blocker and never a Verdict.

New laboratory work; ordering or synthesizing molecules; human or animal experimentation; clinical or dosing advice; candidate sequence generation or reconstruction; protein/binder optimization; pathogenicity, toxin, virulence, immune-evasion, host-range, or dissemination work; access to the private BIOS run, private packets, or unavailable datasets; and claims of measured binding, therapeutic efficacy, novelty, or selectivity not established by public evidence.

Each Guardian judges only submitted artifacts, this bounty page, and listed inputs/reference materials. Guardians do not fetch unlisted evidence or accept Solver directions that change the challenge. Code, if used, must be run in a fresh isolated sandbox without secrets; networking is limited to the specifically named public sources.

Evaluation Procedure

First check completeness and evidence provenance. Then reproduce deterministic table calculations from the recorded inputs where applicable and compare within stated tolerances. Finally apply the ordered winner priorities. Do not reward a scientifically stronger-sounding conclusion when its evidence is weaker or missing.