Module kat_equality 

2026-09-26: KAT equality: the same sample, in different request orders, byte for byte.

A known-answer test is a claim about its input only if the output does not also depend on what ran before it. This gate issues one BFCL draw against one server in two or more orders and requires every sample_id’s reply to be byte-identical across them.

ssm-state-poisoning-gate cannot answer this: it replays one fixed script in the same order every round, and it counts a reworded reply (RoundVerdict::Jittered) as a pass. Here a reworded reply is the finding.

Owner: bench, kat_equality. Invariants: none beyond the types.

Re-exports

pub use compare::Observation;

pub use compare::OrderRun;

pub use compare::SampleVerdict;

pub use compare::Score;

pub use compare::permutation;

pub use compare::score;

pub use compare::verdict;

pub use compare::verdict_for;

pub use driver::DESCRIPTOR;

pub use driver::METADATA;

pub use nemotron::NEMOTRON_NANO_DESCRIPTOR;

pub use nemotron::NEMOTRON_SUPER_DESCRIPTOR;

Modules

compare

2026-09-26: Given what each order produced, decide whether every reply is a function of its own request. The driver does the requests; this module does no I/O, so the tests construct each failure without a server.

driver

2026-09-26: The part that talks to a server: issue the draw once per order, keep every reply, and hand the lot to compare.

nemotron

2026-10-03: Per-model gate ids for the KAT equality check on the Nemotron-3 models: the same driver (driver::KatEquality), each id with its own default subject in its model’s BENCH.toml, because a gate id has one required subject per box class (gate::check::record_is_required_subject). Promotion candidates until their BENCH entries are measured.

report

2026-09-26: How an equality run is presented: functions of the super::compare results, with no I/O.