Skip to content

Evidence

An agent’s own report that it ran the check proves little, so the book reads what the tools returned. The extension reads five kinds of evidence and asks the agent only for what none of them can show.

What counts as evidence, and how it is read
  • Production edits: a write or edit in any git repo, outside its scratch/ and docs/specs/ and the files git ignores. A linked child’s edits and a successful git merge, cherry-pick, or pull count too.
  • The exit code of the bash call that ran the anchor command. An anchor that passed, failed, and passed again records its own friction entry, already reflected as resolved: the rerun command and the failed run’s key line, at most 120 characters and never a stack trace. The agent is told nothing.
  • Reviewer Blocking: P0=<n> P1=<m> counts. They come only from the reviewer and adversarial-reviewer roles, are read in a subagent session too for its own reviewer, and are counted from the reviewer’s dispatch, so a change made while it read is not covered.
  • Summaries of docs/specs/*/ artifacts, read as they are edited.
  • A returned subagent that Jev judges to have been a fresh-context read of the spec.