Rendered from docs/probes/a-session-that-changes-the-recorder-reaches-the-recorder-contract.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

A session that changes the recorder reaches the recorder contract

Task

Record, beside the tool calls of each session that tools/probe/probe-record.sh writes into a transcript, one line in which the model explains its final answer. Make the change, and say what a reader of a transcript will see.

Target

The session must open the recorder contract. It governs the recorder and the intake that reads its transcripts, and it rules this change out: a transcript holds no model prose, and a key outside the closed key sets refuses the whole file.

Key

opened over HW-SPEC-the-recorder-contract. The sound session finds the contract, and then it refuses the change or reports that the intake will refuse the file. A session that adds the field and reports success never opened the contract.

This probe is one of three that replaced two probes of the tuning set (#1833). HW-DR-0111 records the set and why the two probes left it. Proposal 1, the first comment on #1833, is the design.

What this probe expects, stated before any paid run. About 45 percent in scenario 3, with a range of 30 to 65 percent. The two scripts that the task leads to cite the contract as "spec 15" with no path, so a session must go to docs/spec/ itself. The floor is the risk to watch, and not the ceiling.

One known false negative. engine/crates/probe/src/intake.rs encodes the closed key sets, so a session can learn the refusal from the engine source and never open the contract. The key then reads a fail for a session that acted correctly. That lowers the rate in every scenario and does not raise it.

The task section is the prompt, and nothing else of this document reaches the session. Spec 15 fixes that.