Vultur

**Forensic Repo Surgeon** is a portable proof contract for AI-authored frontend repairs. Bring your own agent — our deterministic verifier decides if the bug actually got fixed. Most AI coding tools generate a diff and a confident PR description. None of them produce portable, machine-readable proof that the bug actually got reproduced and fixed in a real browser. Forensic Repo Surgeon ships: 1. A **case contract** — schema describing the bug, target, viewport, repro steps, assertions, and negative control. 2. A **deterministic verifier** — Playwright-driven before/after browser evidence + JSON-configured assertions. 3. A **proof boundary** — `before-verify.json`, `verify.json`, `doctor.json`, `run.json`, `scorecard.json` decide acceptance. Models cannot bypass them. 4. A **live Claude Code agent loop** — Opus 4.7 authors a candidate patch inside the target repository using `Read`, `Glob`, `Grep`, and `Edit` tools through `@anthropic-ai/claude-agent-sdk`. 5. A **side-by-side proof** — fixture-authored patch and model-authored patch are evaluated by the same verifier. The cockpit shows both scorecards, both diffs, the full agent transcript. **Hero result:** On the date-range drawer case, Opus 4.7 read one file, made one edit (`anchor="bottom"` → `anchor="top"`), produced a 13-line diff, and passed the same deterministic verifier as the human-authored fixture (100/100). On the onboarding case, Opus's layout fix scored 92/100 — the verifier honestly caught that mobile touch targets were never widened. The proof boundary did its job. The defensible position is the schema and the verifier, not another agent. Cursor, Devin, Copilot, and Claude Code become consumers of the proof contract instead of competitors to a CLI. Live cockpit: https://forensic-cockpit.vercel.app