A captured workflow · 25 October 2026
A small change.
A better question.
The tests pass. But would they notice a mistake at the boundary?
These are recorded local service runs on a small Rust fixture, not a public code-execution service or a general performance benchmark.
1. Make the function easier to reason about
Poopbags compared the original surcharge function with a refactor. Reported complexity fell from 5 to 1; CRAP fell from 30 to 2. Coverage was assumed 0%, not measured. The helper functions remained visible in the scan.
This shows a reduction in the analyzer's complexity estimate. It does not establish that the tests became stronger.
2. Ask what the tests would miss
codeMUTATOR changed disposable copies of the Rust source and ran the tests. Both conditions used the same nine selected sites and operators. Production source stayed unchanged between the weak and strengthened test conditions.
| Test condition | Killed | Survived | Invalid / errors |
|---|---|---|---|
| Original assertions | 8 | 1 | 0 / 0 |
| Stronger boundary assertions | 9 | 0 | 0 / 0 |
The boundary change that escaped the original assertions was caught after the assertions were strengthened. A survivor needs investigation: equivalent changes can survive too. Compiler rejections and infrastructure failures are excluded from the scored denominator.
3. Keep the limitations with the result
The Rust analyzer uses lexical rules. The disclosed fixture revision parenthesized the boundary operand to match its definite-left-hand-side scanner rule. An earlier scoped run failed because the disposable workspace omitted Cargo context; that failure was retained, and the manifest-copy repair was verified before this comparison.
The advisory Poopbags and asynchronous codeMUTATOR Claude integrations have real-service receipts. Further freshness, partial-result, timeout and cancellation checks are still being repaired. An accepted ordinary Claude CLI workflow is not claimed here.
4. Remember why the judgment changed
A separate Licorice demonstration used a disposable identity to record an observation, explicitly adopt a position, revise it with contrary evidence, and recover it in a fresh process and after restore. Its recalled packet used 455 of the allowed 2,000 tokens.
The service comparison and memory demonstration are separate captures. Their joined everyday-harness demonstration is the next step.