RIFT Verifier

Faulty baseline: two requests for Profile B return cached Profile A. Exact saved native trace.
Public checker: browser verifies recorded native evidence. Healthy 4/4, faulty 2/4, repaired 4/4.
Repaired candidate: the same Profile B requests return Profile B. Recorded trace, newly checked in the browser.

A repair you can inspect

RIFT connects GitLab Duo repair work with a concrete question: does the committed change pass the same experiment that exposed the bug?

We built a supervised workflow around an MIT-licensed AffinityQA component. An operator starts and reviews bounded runs. Two real GitLab Duo Agent Platform sessions handle the Incident Proposal and Minimal Remediator steps. For this controlled case, the repair contract supplies the exact old and new line, and the remediator applies that one-line edit.

Same trial, measured result

The incident is easy to see: a cache key omits the requested profile, so Profile B receives Profile A's cached ranking. The catalog and application engine are synthetic, making the experiment reproducible.

A separate native GitLab CI verifier runs healthy, faulty and repaired versions in separate containers against one sealed four-case trial. The measured result is healthy 4/4, faulty 2/4, repaired 4/4.

Explore the proof yourself

Open the public dashboard and press Check saved proof. Your browser verifies all 12 retained files and recomputes the profile-isolation checks from the saved bytes of that real native run.

Switch between the three trace views to follow the requested, transmitted, tool and returned profiles. The faulty view exposes the crossover; the repaired view preserves each profile. Expand the fingerprints and JSON to inspect the underlying records. The button checks saved execution evidence; repair execution belongs to the supervised GitLab workflow.

Built to reproduce

RIFT combines GitLab Duo, GitLab CI/CD, Python, Docker, Git object checks, JSON Schema and a JavaScript proof checker. The MIT repository and reproduction instructions make the same receipt checks runnable with Python.

Explore the real Incident Proposal and Minimal Remediator workflows, plus the successful public verification and website pipeline 2993692196.