Retrieval agent
chooses one bounded action
One bounded retrieval agent turns a biomedical question into inspectable PrimeKG evidence. Deterministic software replays every citation before a claim reaches the screen.
Which drugs are indicated for acromegaly?
5 cited rows · reproducedLanguage-model prose remains a candidate. Guarded software controls graph access; an independent verifier decides which cited claims can be displayed.
chooses one bounded action
inspect, resolve and query
returns bounded evidence
re-executes every citation
preserves every outcome
Record the biomedical question and immutable run identity.
One retrieval agent uses bounded tools. The resolver, graph, verifier and artifact store are independent subsystems—not additional agents.
Verification integrity and task correctness answer different questions. The miss stays beside the successes.
Provenance and reproducibility—not clinical truth.
The ranked-exclusion output-shape failure remains retained.
No statistical-significance claim is made.
Official levels define readiness. These project-specific gates state the evidence needed to move forward.
Frozen graph, retained runs, guarded tools and deterministic citation verification.
Independent partner, representative biomedical tasks, predefined benchmark and domain review.
Integrate into a partner research workflow; validate usability, robustness, security, latency and cost.
Named operators, monitoring, escalation, incident response, training, rollback and acceptance criteria.
Frozen intended purpose, lifecycle quality, risk, validation, cybersecurity and governance evidence.
Sustained authorised use with longitudinal performance, drift, incidents and corrective action.
The live research demonstrator exposes questions, bounded execution, citations, graph provenance, failures and replay.
Open the live demo ↗