Module probes

Module probes 

Source
Expand description

2026-09-26: The serve matrix’s per-round probes: each turns its requests into a Signal (or a tok/s figure); the bars that read them are in score.rs.

Owner: bench, serve matrix. Invariants:

  • Every probe request is greedy (temperature 0) and carries no sampling penalty (base_body, and coherence::probe for the coherence leg). A repetition penalty would penalise the repeated \n that code_shape requires.

Structs§

Coherence
2026-09-26: What one round’s coherence leg established.

Constants§

NEEDLE
2026-09-26: Needle for the long-context probe; a test asserts the filler never contains it.

Functions§

code_shape
2026-09-26: The structural test on the reply text.
codegen_probe
2026-09-26: Structural codegen: does the reply define fib with an indented body (code_shape)? The reply is not executed.
coherence_probe
2026-09-26: Run crate::coherence::probe (the coherence::CHECKS questions and the model-list check) and map it to the round’s coherence count and identity signal. The string is the probe’s concern, if any.
long_context_probe
2026-09-26: Needle-in-a-haystack with a tokens-token filler.
tool_call_probe
2026-09-26: Did the model call get_weather for Paris? A 4xx refusal of the request is NotApplicable (is_client_rejection); any other error, or a reply without that call, is a Fail.
tps_probe
2026-09-26: Decode tok/s from the client’s TPOT (ChatOutcome::tpot_ms) on a budget-token request. None when TPOT is undefined (no delta arrived, or fewer than two completion tokens); a failed request reports 0 with its error.