Carevia Release Witness — Supervised

Inspiration

A translation can sound fluent while changing a unit, frequency or negation. Carevia Release Witness gives care-software maintainers a narrow way to inspect release evidence without allowing an AI to make clinical decisions. Everything demonstrated is synthetic. The AI never diagnoses, prescribes or changes doses.

What it does

A deterministic checker compares quantity, unit, form, route, frequency and negation against a frozen synthetic baseline in Portuguese, Spanish and English. Unknown wording and mismatches are rejected rather than interpreted by a language model.

A human Mention starts a supervised GitLab Duo Agent Platform flow on a synthetic text merge request. The flow checks the current MR, head, pipeline, required jobs and stamped JSON evidence, then creates a cited GO/NO-GO opinion. Its tools allow read-only GitLab API access and a same-MR note. They do not allow shell execution, file edits, commits, merging or deployment. Human release decisions remain separate.

How we built it

The starting point is a selected, sanitized MIT-licensed Carevia core, with 25 unchanged source files recorded in provenance. It is not the full private application. The competition repository adds TypeScript checks, real GitLab CI proofs, a commit-history boundary checker, synthetic shift replay, the Duo flow and a Google Cloud Run service exposing frozen synthetic report snapshots. The initial automation commit is dated 14 October 2026; commit metadata is not independent proof of when every line was written.

Evidence and challenges

The clinical unit-mismatch job is red. After canonical correction, pipeline 2957638700 has five successful jobs, with 59/59 manual cases and 50 expected rejections. A separate protected-path MR remains NO-GO after a revert and is closed without merging. This demonstrates detection with the jobs intact.

The replay job passes 11 invariants: two committed synthetic records, one lost acknowledgment, one duplicate acknowledgment without another write, and one version conflict for review. Current care is prioritized. The ledger is in memory; this does not demonstrate durable database storage or native offline readiness.

The real Duo opinion is in Portuguese. English configuration v1.2.0 was published afterward and has not been executed. The session may require sign-in, and anonymous visibility of the specific note remains unestablished. English explanatory captions preserve that distinction.

CI is controlled by the participant and can be changed by the participant. Pinned evaluators and public artifacts provide traceable evidence, not independent enforcement. GO does not approve clinical activation, merge or deployment.

The public service exposes build snapshots. Exact bodies were verified 15 October; that does not establish continuous availability. In that snapshot, /healthz returned a Google HTML 404. The existing image scan contains 31 occurrences across 25 distinct CVEs; the historical execution receipt records exit 1 and reachability remains UNKNOWN. The sanitized scan summary retains those findings.

What we learned

Traceability and independent enforcement are different properties. Synthetic negative cases, exact citations and visible limitations help a human assess an advisory outcome without treating it as clinical or release authority.