[cschanhniem]

Summary

1. Helpful CSV column validation (loaders.py)

Previously, a CSV with seq instead of sequence would crash with a bare KeyError deep in the pipeline — no file name, no row number, no hint of what was expected. Now it raises ValueError with the actual columns found:

ValueError: CSV row 1 is missing required 'sequence' column.
Found columns: ['id', 'seq', 'source'].
Check that the CSV header spells it exactly as 'sequence'.

2. Scorer disagreement hard gate (pipeline.py + configs)

The pipeline already computed disagreement = |activity - boman_activity| per candidate, but never used it as a filter. A candidate with activity=0.9 and boman_activity=0.2 could still be nominated — contradictory signals with no resolution.

New eligibility condition:

and item.scores.get("disagreement", 0.0) <= max_disagreement

Config values added:

• pipeline.yaml: max_disagreement: 0.40 (lenient — allows some model spread)
• phase3.yaml: max_disagreement: 0.30 (strict — final synthesis batch needs consensus)

Default (when key absent): 1.0 = gate is off, existing callers unaffected.

Science rationale: two independent scoring approaches (physico-chemical features vs Boman interaction potentials) both predicting activity is better evidence than one score alone. High disagreement means we can't tell which model is right — don't spend synthesis money on it.

Test plan

• test_missing_sequence_column_raises_value_error — correct error type and message
• test_missing_sequence_column_error_names_found_columns — error lists actual columns found
• test_disagreement_gate_blocks_high_uncertainty_candidate — max_disagreement=0.0 rejects every real sequence (all have disagreement > 0)
• test_disagreement_gate_config_values_documented — pins pipeline.yaml=0.40, phase3.yaml=0.30, confirms phase3 is stricter
• Full suite: 781 passed
• ruff check src tests — clean

🤖 Generated with Claude Code