[cschanhniem]

Implements F8 from NEXT_100_PR_MAP: benchmark for rejected candidates resembling known failure modes.

What

• src/openamp_foundry/evidence/failure_mode_similarity_report.py: FMS- prefix schema
• FailureModeSimilarityReport dataclass: 14 fields
• VALID_FAILURE_MODE_CATEGORIES (11 values): hemolysis_risk, low_selectivity, toxicity_flag, dual_use_concern, known_resistance_motif, charge_shortcut, length_outside_range, low_novelty, sequence_degeneracy, poor_charge_distribution, other
• VALID_SIMILARITY_ASSESSMENT_STATUSES (5 values): low_similarity, moderate_similarity, high_similarity, pattern_repeated, insufficient_data
• HIGH_SIMILARITY_THRESHOLD = 0.60, PATTERN_REPEATED_THRESHOLD = 0.80
• 11 validation rules: prefix, failure mode category, similarity status, benchmark status, fraction range, fraction consistency check, min candidates gate, dry_lab_only, pattern_repeated_flag consistency, non-blank fields, n_with_failure_mode <= n_rejected
• validate_failure_mode_similarity_report(), format_failure_mode_similarity_report()
• tests/evidence/test_failure_mode_similarity_report.py: 63 tests, 5 classes
• BASELINE: 8284 -> 8347

Why

A pipeline that repeatedly nominates candidates with the same failure-mode signature (hemolysis risk, dual-use concern, low novelty) is not learning from history. FMS- records surface this pattern machine-readably so calibration can correct it. The pattern_repeated_flag and calibration_action_recommended fields give calibration reviewers a machine-checkable gate.

Disconfirming pass

• dry_lab_only enforced on all reports
• fraction consistency check prevents manual override of statistics
• pattern_repeated_flag consistency with PATTERN_REPEATED_THRESHOLD = 0.80 prevents under-reporting
• min candidates gate prevents over-claiming from micro-batches