ClaudeandClaude Opus 5.5
db2fc3fcfe
The writing rules were only a slice of the agent's input; the rest pulled answers toward a report. Product approved six cuts (2026-11-04): 1. Chat tool results no longer carry the router's English "Shared mandatory baseline" (degrees, nakshatra, MEVG, case calibration, A/B/C labels for every reading). 2. Checklist wording the model read aloud: house-counting is for judging only; "双体星座" becomes the four sign names. 3. The bound skill excerpt keeps only the truth-boundary paragraph (KP / Muhurta table, Ashtottari, P0/P1 list and three-layer method dropped; marker kept). 4. The reference_transparency block goes: the card never carries it. 5. English rules that restated VISIBLE VOICE go; "keep conclusions conditional" becomes "never call it verified, but say plainly what the chart shows". 6. The rectification offer rule is the offer tool's description instead of riding in every user turn. Parents backtest: unglossed terms 16 -> 11 (std) and 25 -> 11 (verbatim question); no absent parent written as present or involved. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbBXoyeZrYn6ryNcSe5Y2Y