Real observations. Clear boundaries.

The question

Can a signed-in customer connect a real repository, request a hosted AI review, and return later to the same saved result? This is the first workflow AntiCode needed to prove before discussing broader outcomes.

The workflow

During September 19, 2026 validation, the AntiCode GitHub repository was connected and its pull requests synced. A customer initiated an AI review from the live dashboard. The completed run was stored against the current head, base, and target branch.

Reloading the page preserved the review and its findings. Separate checks confirmed that unauthenticated requests could not load the workspace. No synthetic test review was presented as a customer outcome.

What the review revealed

The first model pass produced speculative warnings and inaccurate line references. The service was tightened to require a concrete failure mechanism and exact source evidence. A stronger model was then evaluated on a controlled example and the real diff.

A line-ending issue was independently reproduced in the native scanner: CRLF input could hide executable code. The source fix was checked using actual Git diffs and the CLI. This is a development observation, not a claim that the installed app has been updated.

What this does not establish

A working review flow is not a benchmark for defect recall, false-positive rate, or time saved. The final review produced four additional findings for triage; valid source citations do not prove that every conclusion is correct.

Teams should evaluate usefulness on their own changes, record false positives, and compare against their existing review process. No external customer endorsement or percentage improvement is claimed.