We built ParityBridge Release Assistant during the hackathon: a GitLab workflow that connects benchmark reports, test results, immutable container images, and verified deployments, with GitLab Duo assisting failure diagnosis.

Inspiration

Research releases need more than a passing build. A benchmark result should identify the source commit, the command that produced it, the measurement conditions, and the tests that passed.

ParityBridge Release Assistant adds that evidence to an existing MIT-licensed quantum error-correction research project, using GitLab CI/CD and GitLab Duo to support diagnosis and reviewed releases.

What it does

ParityBridge connects source code, test results, benchmark reports, container images, and a public deployment.

Its read-only report viewer displays:

The exact source commit and originating pipeline.
Test outcomes and measurement conditions.
Benchmark metrics and raw data.
Clear labels distinguishing measured, fixture, and historical results.

The release pipeline runs tests, performs a bounded CPU smoke benchmark, builds and checks a container image, and waits for human deployment approval. After deployment, verification checks that the live service serves the expected commit and a valid report.

How we built it

We built the viewer with Python and FastAPI. GitLab CI/CD captures JUnit test results and generates a report from the CPU smoke benchmark.

Rootless BuildKit packages the viewer and report into an image tagged with the full commit SHA. An approved manual job deploys that image to Render, followed by live verification.

We also demonstrated GitLab Duo Agent Platform’s Fix CI/CD Pipeline flow on a controlled report-field bug. The failing code read commit, while the report schema required commitSha.

Duo inspected the failure and proposed the correct field. We reviewed and applied the one-line suggestion through Git, preserving the original test assertion. The repaired hosted pipeline passed 89 tests with four skipped.

This is a Bring Your Own / Assisted entry: Duo supplies diagnosis and a proposed correction; application, verification, merges, and deployment remain separately reviewed.

Challenges we faced

Our first hosted pipeline exposed a YAML command being parsed as a mapping. We corrected the command and added a regression check for that error.

The first real Duo run then failed during startup because the slim execution image lacked curl. We corrected its environment through a reviewed merge and successfully retried the flow.

These failures reinforced the importance of testing the actual hosted workflow alongside local validation.

What we accomplished

The production pipeline passed every stage, including approved deployment and live verification. The public service serves the verified release commit with 88 passing tests, zero failures, four skips, and a four-row measured CPU smoke report.

The separate repair demonstration adds one test, accounting for its 89 passing tests. Its merge request remains unmerged so reviewers can inspect the controlled failure and correction.

What we learned

A useful agent-assisted release needs an inspectable chain: the original failure, proposed correction, preserved assertion, passing pipeline, immutable image, and matching live service.

Keeping those artifacts connected makes the release easier to review and reproduce.

What’s next

The custom Measurement Reviewer is defined but remains disabled; its separate review demonstration has not been executed.

Future work can evaluate that reviewer and extend validation. The current CPU smoke benchmark establishes release integrity; it does not establish a scientific decoding advantage.

Existing project and hackathon contribution

The original project already contained research implementations, benchmarks, scientific results, and tests.

Our hackathon contribution is the GitLab release workflow, report contract and viewer, container/deployment verification, and real Duo diagnosis demonstration. Existing scientific qualifications remain preserved.

Original MIT-licensed project: https://github.com/dakshverma-dev/ParityBridge