Module probe

metrale_bench::benchmarks::ssm_poison

Module probe 

Source
Expand description

2026-09-26: The probe: a fixed conversation script that every round replays from scratch against the same server, and the anchors the reference round must meet.

Owner: bench, SSM poisoning gate. Invariants:

  • The script is constant text, with no run id or unique_prefix_tag, so every round’s turn 1 sends the same prompt prefix (probe_tests pins its SHA-256).

Constants§

LONG_PREFIX
2026-09-26: The document that opens turn 1. Every round sends it unchanged, so a replay’s turn-1 prefill can be restored from the prefix cache.
TURNS
2026-09-26: The user turns of the script, in order. Turn 1 is prefixed with LONG_PREFIX; each later turn is sent after the earlier turns and the model’s replies to them.