Failure
The clean-room repository began at 1 failed and 6 passed. A three-day billing period returned 2 instead of 3.
Clean-room customer scenario · observed run
A one-line billing fix, with the proof attached.
A realistic date-boundary defect moved from a red test to a focused fix and a green suite inside the real Verdict Desktop application. No synthetic UI, no rewritten test output, and no hidden quality claim.
Failure → investigation → proof
The sequence below is the actual scenario contract and observed outcome, not a claim that every repository repair will follow the same path.
Investigation
Verdict Desktop opened the bound repository, inspected the billing implementation and tests, and ran the failing test.
Root cause
Python date subtraction excludes the end boundary. The public contract required both billing dates to count.
Fix
One behavior line changed: (end - start).days became (end - start).days + 1.
Regression check
The focused billing test and the full tracked suite ran again, including the existing reversed-range error test.
Green run
All 7 tests passed. An independent post-run check repeated the same 7/7 result with a clean one-line source diff.
The run in Verdict Desktop
This is a direct capture from the packaged Desktop app. The sidebar and details rail were hidden with the app's own controls for focus; the run output, model path, harness, tests, and receipt were not altered.
Proof detail
The same live terminal state, captured directly from the Desktop Verdict Code widget for legibility.
Resource receipt — scoped to the final model response
Tokens
6,998
6,676 input · 322 output
Model-call time
5.0 s
Not whole-task wall time
Throughput
64.1 tok/s
Reported by the model path
Energy
0.032 Wh
Profile-estimated, not metered
These values are the visible terminal receipt for the final answer, not an aggregate for discovery, tool calls, approvals, and tests. LAN execution recorded no provider-dollar charge; hardware and electricity are still real costs.
What this proves — and what it does not
Observed here
Desktop bound the repository; Verdict Code found the boundary error; one source line changed; the tracked suite passed; and the final response carried a model-resource receipt.
Not established
This single clean-room task does not establish a universal repair rate, a Smart Harness rescue advantage, a PAS advantage, a whole-task energy total, or production behavior on every surface.
Fix the bug. See the proof. Keep control.
Bring the supported model and harness that fit the job. Verdict keeps the executed path, change, checks, and receipt visible.
Download Verdict