Research / Implementation programme
Research Agent Lab
This interactive prototype performs local, deterministic comparison of two illustrative system architectures. It demonstrates the evidence schema, comparison logic and rule-based issue detection. It does not fetch live sources, operate autonomous AI agents, test suppliers or certify legal compliance. Everything shown is simulated educational test data, not findings about named products.
Comparison results
Sample system data and comparison rules
Each system has a stable ID, explicit component list, dependency edges, claimed controls, demonstration-only source record and authorisation level. Jaccard similarity is intersection over union for component sets, so 100% component overlap does not establish functional or ethical equivalence. Rules flag lack of approval, unknown logging and absence of versioned citations as review questions, not verified failures.
Sequential research-agent specification — 12 stages
| Stage | Role | Output / release gate |
|---|---|---|
| 1 | Set purpose, jurisdiction and authorisation | Written scope and approved information sources |
| 2 | Discover system candidates | Deduplicated inventory with source URLs |
| 3 | Acquire authorised documents | Versioned originals, timestamps and hashes |
| 4 | Extract architecture claims | Components, edges, data flows and confidence |
| 5 | Normalise concepts | Controlled terminology and synonyms |
| 6 | Verify each source and assertion | Claim-to-evidence coverage and unresolved assertions |
| 7 | Compare architectures | Shared, distinct and structurally differing features |
| 8 | Map tools and dependencies | Licences, providers, permissions, versions and known uncertainties |
| 9 | Assess assurance requirements | Applicable duties, risks and evidence gaps; no automated certification |
| 10 | Adversarially review findings | Contradictions, false positives and counterexamples |
| 11 | Human review and sign-off | Accountable decision with corrections recorded |
| 12 | Publish approved outputs | Versioned report and audit trail; no automatic release |
Engineering gates before real agents
Document parser fidelity; source permission and robots/site restrictions; scoped tool tokens; sandboxed read-only ingestion; privacy and retention; prompt-injection resilience; citation verification; withheld verdicts when evidence is inadequate; benchmark against human-labelled cases; approval before externally visible publication; monitoring, rollback and incident response.
Agent security: OWASP · Evidence provenance: W3C · Risk management: NIST · Systems Atlas