model-complexity-routerlisted

# Model Complexity Router A task in, a tier out — with the four axis scores shown, and the error-cost rule applied exactly once. ## Verification gates (defined first; output is blocked until all pass) - **G1 — Scores shown:** every recommendation displays the four axis scores it derived from (scope, reasoning depth, error cost, context load — each 0–2) and their sum. A tier with no visible scoring fails. - **G2 — Floor applied exactly once:** error cost enters the total once, at scoring. The floor rule (error-cost 2 → never Haiku) applies only at the mapping step, only if the mapped tier came out Haiku. Any path where error cost moves the outcome twice — in the sum AND as a bump — fails the gate. - **G3 — Honest capability claim:** the output never claims to have switched the session's model. It recommends (user acts via /model) or delegates to a pinned subagent if one exists — and says which. ## The rubric | Axis | 0 | 1 | 2 | |---|---|---|---| | Scope | single file/artifact, cosmetic | one feature, few files | multi-file, architectural | | Reasoning depth | mechanical, pattern-match | known patterns, some judgment | novel tradeoffs, design decisions | | Error cost | trivially reversible | rework hours | compounds (prod, money, public) | | Context load | one prompt | a few files | large cross-file synthesis | Sum 0–8 → **0–2 Haiku · 3–5 Sonnet · 6–8 Opus.** Then the floor: if error cost scored 2 and the map says Haiku, raise to Sonnet — nothing else changes, and never