Expand description
2026-09-26: Cross-contamination benchmark: concurrent requests must not
change each other’s output. score holds the leg design and the
classification, prompts the probe corpus, and driver the state machine
that runs the legs.
Owner: bench (contamination). Invariants: none beyond the types.
Re-exports§
pub use driver::DESCRIPTOR;pub use driver::METADATA;
Modules§
- driver
- 2026-09-26: The contamination state machine. Each
next()runs one step (endpoint probe, a leg, or scoring), so the pane updates as each leg lands and a cancellation is seen before the next step. - prompts
- 2026-09-26: The probe corpus: prompts built so that contamination is detectable.
- report
- 2026-09-26: How a contamination run is presented: the probe-by-leg table,
the summary tiles and the gate metrics, each a pure function of a
Score. - score
- 2026-09-26: Scoring for the contamination benchmark: classify every prompt at
every rung and in the post-check against its own solo reference. The module
reads no handle, network or clock;
score_tests.rstests it with in-memory outcomes. - transcript
- 2026-09-26: Re-export of
benchmarks::transcriptfor the contamination module’ssuper::transcriptimports. Thessm_poison,scheduler_equivalenceandkat_equalitydrivers use the same type, so every driver shares one equality contract.