Module decode_floor 

2026-09-26: Decode Floor Gate: the pinned single-user decode throughput floor. Every generation knob is pinned so that two runs are comparable, and the one metric, the median server decode rate over RUNS runs, is judged against the served variant’s BENCH.toml server_decode_tok_s bound under --pull-request-gate. decode-floor is listed in gate::coverage::REQUIRED.

The pins (the benchmark’s definition, not parameters):

• Prompt: MINHEAP_PROMPT. A different prompt is a different benchmark.
• Request: temperature 0, seed 0, max_tokens = MAX_TOKENS, and reasoning_effort: "none" in the body, which the server maps to thinking off, so the gate serve needs no thinking flag. Thinking off matters: speculative dispatch is off inside <think> unless METRALE_DFLASH_SPEC_THINK=1 (spec_dispatch_eligible), so a thinking-on run would measure serial decode.
• Runs: RUNS, with no warmup parameter. The metric is the median usage."response_token/s".

Vacuity pins (score::evaluate): the run is inconclusive, rendered as a failing verdict, unless every run emits at least MIN_OUTPUT_TOKENS, reports a finite positive server decode rate and a nonzero usage.completion_tokens_details.accepted_prediction_tokens, and the mean of the per-run accept lengths (stats::accept_len: emitted tokens per decode step, completion / (completion - accepted)) is at least MIN_ACCEPT_LEN.

Owner: bench, decode_floor. Invariants: the run verdict is a pass only for Evaluation::Measured with min_tok_s > 0 and a median at or above it (verdict_for).

Structs

DecodeFloor

Constants

DESCRIPTOR

METADATA