Skip to content
Gilang AdamRequest a quote
← All work

Case study / Forensic Repo Surgeon

A verifier, not a model, decides whether the fix works.

Portable proof contract for AI-authored frontend repairs. Bring an agent and let a deterministic verifier decide if the bug is fixed.

Source for Forensic Repo Surgeon ↗

Source available. The public cockpit is withheld here while recorded evidence is sanitized for local filesystem paths.

TypeScript

The problem

An AI-authored patch is not proof that a bug was reproduced or fixed.

The approach

  1. Define a case contract with a viewport, reproduction steps, assertions, and a negative control.
  2. Run the same deterministic Playwright verifier against fixture and model patches. Export machine-readable scorecards and hashed artifacts.

What the evidence shows

The public README reports a passing date-range repair and a separate onboarding repair that failed the verifier because mobile touch targets were still too small. These are controlled demonstration results, not a production reliability estimate.

Known limits

Two published cases on one target are a narrow benchmark. The repository explicitly says production confidence has not passed. The public cockpit's path-redaction issue must be resolved before it is promoted as a clean evidence package.

Inspect the evidence

Have a related problem?

Request a scoped quote ↗