Research Assurance Case Engine

Problem

Research programmes depend on protocols, consent materials, data definitions, operating procedures, approval records, and reporting obligations that evolve in different systems. Individually correct artifacts can still produce contradictory requirements, unsupported claims, stale evidence, or approvals that cannot be reconstructed.

Solution

Research Assurance Case Engine turns versioned research documents and evidence into governed claims and replayable Research Assurance Packages.

• ProtocolSync detects cross-document conflicts, omissions, ambiguities, citation problems, and drift using exact source versions, fingerprints, and anchors.
• ProofFlow classifies claims as verified, partially supported, contradicted, unsupported, or not assessable.
• Protected actions remain blocked until an attributable human approves, conditions, or rejects the exact scope.
• DecisionReplay reconstructs the evidence, policy, authority, actions, and outcomes available at decision time and marks earlier conclusions stale when source evidence changes.
• Checksummed exports, idempotency, separation of duties, immutable evidence, and privacy-safe telemetry preserve an auditable boundary.

Agent architecture

Google ADK coordinates the bounded workflow. Gemini 3.5 Flash, accessed through the Google GenAI SDK, supplies structured non-authoritative semantic assessment through a fail-closed adapter. Deterministic code remains authoritative for source identity, citations, claim states, approval scope, transitions, retries, export checksums, replay, and staleness.

The exact frozen source 38b53a46c1b81f8815cf02b096eda10a43601f1b is locally verified offline: 281 tests passed, the benchmark passed 20/20 cases, the clean-clone gate passed 23/23 checks, and production-shaped local container smoke passed. An independent BoundedOps source-only evaluation returned PASSED_WITH_LIMITATIONS across 8 scenarios and 24 observations: 3 PASS, 5 intended HUMAN_REVIEW_REQUIRED, zero findings, zero remediation proposals, and zero target-reported side effects.

Separate bounded credentialed synthetic workstreams at pre-final commit 32bf130 exercised Gemini, Google ADK, Firestore, Cloud Storage, a non-public IAM-authenticated Cloud Run revision with no public invoker, rollback, and allowlisted Cloud Logging. Network ingress was all, not internal-only. These records do not establish an exact-final integrated hosted trajectory, production safety, or causal impact.

Data and safety

The demonstration uses synthetic protocols, consent text, CRF rows, data dictionaries, procedures, reports, and approval events. It contains no participant data, confidential institutional material, credentials, or private third-party repository content.

The engine improves evidence available to accountable people. It does not establish scientific validity, ethical acceptability, legal or regulatory compliance, clinical appropriateness, financial authority, programme effectiveness, or causal impact.

What we learned

Useful agentic autonomy does not require transferring authority to a model. Semantic assistance can propose bounded meaning while explicit machine contracts preserve provenance, approvals, retries, immutable evidence, and safe stops.

Reproducibility

The private repository includes step-by-step local setup, locked dependencies, synthetic demos, complete tests, container smoke, clean-clone validation, architecture source, an independent-evaluation contract, and a judge guide. Read access has been invited for the official Devpost testing account.

Built With

• docker
• fastapi
• gemini-3.5-flash
• google-agent-development-kit
• google-cloud
• google-cloud-firestore
• google-cloud-logging
• google-cloud-run
• google-genai-sdk
• opentelemetry
• python