# Workbench run record

Keep one copy per run. Blank fields mean not recorded, never zero.

## Setup

- Date:
- Fixture revision or archive checksum:
- Harness version/profile/preset:
- Workspace and allowed edit paths:
- Structure route (provider ID + model ID):
- Examples route (provider ID + model ID):
- Route verification and remaining uncertainty:
- Changes from the baseline setup:

## Observations

- Live model run or local fixture only:
- Session/workflow identifier:
- Start and end times, if measured:
- Tool calls and retries, if recorded:
- Findings checked against files:
- Findings rejected, with reason:
- Files changed:
- Exact check commands and exit codes:
- Outcome unknown or interruption handling:
- Not run / unresolved:

## Decision

- Did the final document meet the agreed contract?
- Did a human inspect the example command and the diff?
- What did the customization improve, if anything?
- What extra setup or review work did it add?
- What would you keep for the next task?

One task does not establish a model ranking. Do not replace missing usage or timing
evidence with estimates presented as measurements.