For judges
Every claim, with its evidence
Each row links a claim to something you can open without signing in. 23 of 24 are live today; rows marked "in progress" will link to evidence before submission.
Repository: gitlab.com/gitlab-ai-hackathon/transcend-october-2026/4656288/virtual-trial-oref0. Built on OpenAPS oref0 (MIT).
Technological implementationTechnological implementationWhat it is. How thoroughly and skillfully the project uses GitLab to automate the post-code DevSecOps lifecycle, with working, non-trivial code and pipeline config.Why it matters. The rows below link to the code, pipelines, and job logs that show it.
| Claim | Evidence | Status Live or in progressWhat it is. Live means you can open the evidence now without signing in. In progress means it is still being built.Why it matters. We never claim something is done before you can check it. |
|---|---|---|
| oref0 runs as a deployed dosing engine, built from this repository's lib/ | Live engine health and version | Live |
| Every engine release is a canary: no traffic until 151 decisions pass the safety envelope, checked again after promotion | Deploy job log (search for PASSED) | Live |
| GitLab CI deploys to Google Cloud with no stored keys | Workload Identity Federation (infra/wif.tf) | Live |
| All Google Cloud infrastructure is code | Terraform in infra/ | Live |
| Security scans on every merge request and on master | Pipeline: SAST, secret detection, dependency scanning | Live |
| oref0's own tests run on every change; any new failure blocks | Test report: 234 passing, 9 known upstream failures, 0 new | Live |
| Uptime checks and alerts on everything judges open | Cloud Monitoring config (infra/monitoring.tf) | Live |
| Closed-loop trial on every merge request: base and head oref0 dose the same virtual patients | Trial and gate jobs in the master pipeline | Live |
| Ten virtual patients live on the deployed engine around the clock | Live clinic page | Live |
| Two GitLab Duo flows plan trials and explain verdicts, started by a mention, published in the AI Catalog | Flow definitions (.gitlab/duo/flows/) | Live |
| Agents cannot change dosing code, the trial, or the gate limits, enforced by the pipeline | Agent boundary check (ci/agent_boundary.py) | Live |
| Eight scenarios, five of them faults modeled on real bugs and recalls | trial/scenarios.py | Live |
| The gate blocks small but consistent harm with a paired exact sign test, in both directions | gate/thresholds.yaml | Live |
DesignDesignWhat it is. Whether the project delivers a complete, coherent end-to-end workflow, not just a technical proof of concept.Why it matters. The rows below show the whole path, including what happens when things go wrong.
| Claim | Evidence | Status Live or in progressWhat it is. Live means you can open the evidence now without signing in. In progress means it is still being built.Why it matters. We never claim something is done before you can check it. |
|---|---|---|
| A risky engine release is stopped before anyone is served | Seeded bug (safety cap 3 to 6) blocked: 25 of 151 decisions above the limit | Live |
| One click rolls the engine back to the previous version | rollback-engine job (infra/ci/deploy.gitlab-ci.yml) | Live |
| One merge request tells the whole story: plan, trial, verdict, explanation | Demo MR !10: the plan agent's note and label, the 6-shard trial, the gate's BLOCKED summary, and the explain agent's note | Live |
Potential impactPotential impactWhat it is. Whether the project makes a credible, specific case for solving a real problem for a real audience, and actually addresses it.Why it matters. The rows below link to the recalls, the broken upstream tests, and the benchmark.
| Claim | Evidence | Status Live or in progressWhat it is. Live means you can open the evidence now without signing in. In progress means it is still being built.Why it matters. We never claim something is done before you can check it. |
|---|---|---|
| Software changes to diabetes devices have hurt people | FDA recall notices and studies | Live |
| The existing tests on this real dosing code were already failing and ignored | Known upstream failures, each with its reason | Live |
| A seeded dosing bug (max insulin on board x4) is blocked by the trial in a real pipeline | Gate job log: BLOCKED, low blood sugar risk +30% in 222 of 300 matched runs | Live |
| Virtual Trial catches harmful dosing changes that today's checks miss, without flagging harmless ones | Benchmark: 14 seeded changes, four checks | Live |
InnovationInnovationWhat it is. How creative and novel the concept is, and how inventively it applies agentic automation compared to existing work.Why it matters. The rows below show what is new and name the prior art honestly.
| Claim | Evidence | Status Live or in progressWhat it is. Live means you can open the evidence now without signing in. In progress means it is still being built.Why it matters. We never claim something is done before you can check it. |
|---|---|---|
| A clinical trial as a pipeline stage | How it works | Live |
| Agents may test but never touch dosing code | Safety page | Live |
PresentationPresentationWhat it is. Whether the video and materials clearly show the automation running end to end, and who it is for and why it matters.Why it matters. The site, the try-it page, and the live clinic let you check it yourself.
| Claim | Evidence | Status Live or in progressWhat it is. Live means you can open the evidence now without signing in. In progress means it is still being built.Why it matters. We never claim something is done before you can check it. |
|---|---|---|
| Anyone can ask the deployed engine for a decision | Try it page | Live |
| Demo video under 3 minutes | YouTube | In progress |