out/_probe/

The recall probe's questions, answers and scores, one set per leg.

What it holds

NN-L.questions.md (five questions, "Answer from memory, without opening any file"), NN-L.answers.md (the reply verbatim) with NN-L.answers.md.session (the fork's id), NN-L.json (answers, what was expected, misses, score out of 1), and notes.jsonl, one line per leg with a miss.

Who writes it, and when

close.sh, after the leg closes: probe.py --ask, then protocol_guard.py probe runs one headless turn in a fork of the leg's session with Bash, Read, Edit, Write, Glob, Grep, NotebookEdit, Agent and Task disallowed, then probe.py --score. The leg's own transcript is untouched.

Why it exists

STANDING §2: the probe "measures whether the goal survived the leg"; a miss the head or the diary answers is a note, and a miss neither answers is a defect. It is closed to the analyst until E1, by the permission rules and the sandbox, so it cannot learn the questions. E1 writes probes-E1.md from it.

Where

out/_probe/.

Checked against: tools/probe.py, harness/close.sh, harness/protocol_guard.py, STANDING-INSTRUCTIONS.md §2, stages/STAGE-E1-formalizing-audit.md.