Module prompts

Expand description

2026-09-26: The probe corpus: prompts built so that contamination is detectable.

Each prompt carries a canary and asks the model to echo it, so the canary is in both the prompt tokens and the generated tokens. A foreign canary in a reply is leakage on its own evidence (super::transcript::Transcript::carries_foreign_canary).
The canaries share no --separated component and neither contains the other (tested below).
The prompts are identical up to the canary: the shared preamble is a prefix-cache collision surface, and the text right after it is the canary sentence, so resuming the wrong continuation at that boundary emits the foreign canary.
The topics differ (lighthouse, water mill), so a leak that mangles the canary still diverges from the solo reference (Class::Diverged).
The driver sends temperature 0, and the prompts ask for a fixed shape (code, three sentences, code).

Owner: bench (contamination). Invariants: every canary occurs in its own prompt’s tail and in no other prompt (tested below).

Structs§

Probe: 2026-09-26: A prompt and the canary that must not appear in any other probe’s reply.

canaries: 2026-09-26: The canaries in probe order: super::score::score reads the canary at a probe’s index as that probe’s own.