Production fixes:
- Cross-business join safety: all queries join on (review_id, business_id)
- Timestamp normalization: iso_z() for all output timestamps
- Score formula alignment: matches PERIOD_SCORES_QUERY for consistency
- Invariant check: fails if scores.overall != comparisons.current
- primary_run_id: uses max(created_at) in time_window mode
- Language normalization: auto/auto-detect -> unknown
- Review language: majority voting over spans per review
Executive summary guardrails:
- Weakness priority: negative driver > qualifying dip > none
- Dip qualification: within 90 days AND review_count >= 3
- Most recent dip selection when multiple qualify
- No contradiction: "dip" cannot pair with "no major issues"
- Action grounding: must tie to cited weakness or top positive driver
CLI options:
- --no-summary: disable executive summary
- --require-summary: exit code 2 if LLM fails
- --summary-model: configurable model (default gpt-4o-mini)
Includes unit test suite (16 tests) for narrative guardrails.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
5324542e72