Skip to content

Data Format

m4bwav edited this page Nov 21, 2026 · 1 revision

Files

Both files sit in TRUSTWRIGHT_HOME (default ~/.trustwright, created on first use). Each line is one JSON object, written as UTF-8 with LF line endings and non-ASCII characters kept as they are. trust.py only appends; it never edits or deletes a line. To correct an entry, edit the file by hand.

File Written by Read by
behaviour.jsonl log show, diff
stops.jsonl stop show

behaviour.jsonl

Fields in order: date, product, version, model, kind, observation, evidence. Optional fields left out on the command line are stored as empty strings.

{"date": "2026-11-03", "product": "claude-code", "version": "2.1.3", "model": "", "kind": "instruction", "observation": "Ran npm test instead of the test command in AGENTS.md", "evidence": ""}

stops.jsonl

Fields in order: date, harness, version, model, task, source, category, message, workaround.

{"date": "2026-11-14", "harness": "claude-code", "version": "2.1.5", "model": "example-model", "task": "Explain how a known CVE in my own server works", "source": "model-refusal", "category": "security-research", "message": "I can't help with that.", "workaround": "Said it is my server and asked for the patch"}

How versions compare

show sorts and diff filters by version. A version is split on . and -; number parts compare as numbers, so 2.1.10 comes after 2.1.9. Since 0.1.1, a text part sorts after any number part, so unknown comes after every numbered version and 2.1.0-beta after 2.1.0. In 0.1.0 a text part raised a TypeError as soon as it had to be compared with a number.

That ordering has one consequence for prereleases: 2.1.0-beta sorts after 2.1.0, the opposite of semantic versioning. Use plain numbered versions where you can.

date is a string; entries on the same product and version sort by it, which works for YYYY-MM-DD.

Stop sources

Source What it is
provider-classifier a hard stop from a safety classifier around the model; in Claude Code an API error naming the Usage Policy, with no model text in the turn
model-refusal the model itself declines or answers only part
harness-permission the harness blocks a tool call (a permission rule, the sandbox, an auto-mode classifier); not a model judgement
other hedged partial answers, an agent quietly skipping a tool action

Record the source every time: a large part of a measured over-refusal rate can come from the classifier rather than the model.

Privacy

Write --task and --message without secrets, tokens or private names; paraphrase the sensitive part. The files are plain text and anyone who can read your home folder can read them.

Clone this wiki locally