pub const DEFAULT_EFFORT_CAPPED_AT_CEILING: bool = false;Expand description
Default [behavior].effort_capped_at_ceiling — whether qualitative
reasoning_effort levels are clamped at the model’s effective ceiling E
(max_thinking_budget / --max-thinking-budget).
false preserves the historical ladder shape: high = 2E and xhigh = 4E
EXCEED the ceiling, exactly as the pre-symbolic absolutes did (512/1024
over the built-in 256). Parity at defaults is pinned by
effort_ladder_at_default_ceiling_matches_the_historical_absolutes.
true is for models with MEASURED non-monotonic degradation above their
ceiling — where a bigger thinking budget scores WORSE, so a client’s
boilerplate reasoning_effort: high must not double a deliberately small
E (e.g. Qwen3.5-397B NVFP4, 2026-05-07 sweep: budget 256 is worse than
128). The clamp binds ONLY the server-policy effort ladder; an explicit
client token budget (thinking_token_budget etc.) is never touched by it.