Verification Loops: Never Trust an Unrun Edit

PROJECT CONTEXT:
My agent edits code, declares success, and the code is broken. The harness-engineering literature treats this as THE core lesson: model claims are hypotheses, and the harness must verify every edit with real execution before accepting it.

GOAL:
Build a layered verification pipeline into the agent loop: syntax checks on write, targeted tests after change, and evidence-based completion.

REQUIREMENTS:
- Automatic syntax/lint verification on every file write, with revert on failure
- A verify step after each logical change: run the relevant tests or a behavioral check
- Completion gating: 'done' requires machine evidence, not model assertion
- Verification results fed back to the model verbatim
- An evidence record per task: what was checked and what it showed

CONSTRAINTS:
- Verification runs in the harness, not at the model's discretion
- A failed verification can never be papered over by re-describing the change

DEFINITION OF DONE:
- A syntactically broken edit is auto-reverted with the error shown to the model
- A behavior-breaking edit is caught by the post-change test run
- Task completion output includes the evidence record
- The agent cannot mark work complete while verification is red

COMMON FAILURES TO AVOID:
- Trusting 'I have fixed the issue' without a single command being run
- Lint-only verification that misses behavioral breakage
- Verification output hidden from the model, so it cannot self-correct
- Tests skipped 'for speed' exactly on the changes that needed them