[cschanhniem]

Change

Added batch-2 selection rationale report (G5) — a CLI that generates a synthetic candidate pool, runs the batch-2 selector with configurable weights, and produces a per-candidate rationale report classifying each selected candidate into exploit / explore / diversity / combined roles.

Why

Phase G aims to make active learning transparent. G5 closes the gap between the raw select-batch-2 output and a reviewer's understanding of why each candidate was chosen. Without this report, a reviewer sees selected candidates but not whether the selector is exploiting high-scoring regions, exploring model blind spots, or diversifying sequence space.

Evidence

• 19 new tests in tests/active_learning/test_batch_rationale.py
• JSON Schema conformance test validates output against schemas/batch_rationale_report.schema.json
• CLI smoke tests verify exit 0 and file output
• make agent-check passes
• Test count: 2955 → 2974 (19 new)

Safety impact

None. This is a synthetic-data-only report generator. All output is explicitly caveated as computational analysis requiring human review.

Proof level

Level 1 (code-path integrity). All data is synthetic. No biological claims.

Limitations

• Uses synthetic data with known labels — role classification reflects code behavior, not biological relevance
• Role threshold (>0.05 above second-place) is arbitrary; edge cases classified as 'combined'
• Single-role label does not capture multi-objective optimization tradeoffs

Human review needed

No