| name | metacog-bench |
| description | Run a minimal metacognitive evaluation pass over the current task or repository change. Use to assess quality, calibration, and overhead. |
| user-invocable | true |
| disable-model-invocation | true |
| allowed-tools | Read, Grep, Glob, Bash |
Minimal MetaCog Bench
Evaluate the current work on six axes:
| Axis | Question | Score 1-5 |
|---|
| Correctness evidence | Is the result backed by tests, checks, or verifiable facts? | |
| Calibration quality | Does the stated confidence match the actual evidence strength? | |
| Tool efficiency | Were tools used only when necessary? Any redundant calls? | |
| Context overhead | How much context was consumed vs. the value delivered? | |
| Workflow friction | Were there unnecessary round-trips, retries, or dead ends? | |
| Residual risk | What is the likelihood of an undetected issue? | |
Return exactly
Scores: correctness=X, calibration=X, tools=X, context=X, friction=X, risk=X
Strongest gain: [what Aletheia workflows helped most]
Strongest cost: [what added overhead without proportional value]
Improvement: [one concrete instrumentation or workflow change]
Rules:
- Be honest. A 5/5 on correctness with no tests run is wrong.
- "Strongest cost" must be actionable (not "it took time").
- The improvement must be specific enough to implement in the next session.
