- Backtest harness writes, per run, the exact model request bodies (no
headers or key), the computed card and a manifest: system prompt
sections with char counts, tool calls and results, and a truncation
check of the transmitted card (parses in full, equals the computed card,
strings at the 800 cap, arrays at projection caps or shorter than the
engine output, card gaps). --no-manifest turns it off.
- consult-answer-factcheck.py extracts checkable claims (regex, plus
optional model extraction into JSON only) and verifies every claim
deterministically against the card: matches / contradicts / not_on_card
/ status_mislabeled; per-layer transmitted / named / cited-correctly /
audit status with "Used but never cited" flags; date precision,
conclusion position and grade-rule checks; --report confirms answers
written from a skill report.
- Professional voice (structure only): a short 结论 block first; point and
lord facts only from the card field, one value per point; status words
copied (the Yoga line no longer says 「只写 yogas.hits 里成立的」); a dasha
level the card does not give is 未交付; dates are 名义计算边界 unless the
birth time is rectified; each grade cites its rule and missing items.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
ClaudeandClaude Opus 5.5
d890ecb7be