How ARA works
ARA answers one question: what could go wrong with this AI use case, and what should the organisation do about it? It is decision support for risk, compliance and security teams, not legal advice.
The pipeline
| Step | Agent | Output |
|---|---|---|
| 1 | Scenario Intelligence | Scenario Brief: every attribute tagged user-confirmed, AI-inferred or not specified; missing information; refinement questions; alternative approaches. |
| 2 | Risk, Compliance and Security analysts (in parallel) | Risk register (likelihood, impact, inherent rating); regulatory findings with applicability and gaps; traditional and AI-specific threats. |
| 3 | Control Analyst | Specific controls mapped to risk IDs; residual rating per risk; overall inherent and residual. |
| 4 | Supervisor | Decision (Approve, Approve with conditions, Requires remediation, Escalate, Do not approve), conditions, 0-30 / 30-90 / 90+ day plan. |
| Optional | Red team | 8 adversarial tests judged against the proposed controls. |
Every agent returns typed JSON that is validated against a schema; malformed output is repaired or rejected, never shown as free text.
Evidence and honesty rules
- Facts not in the scenario are marked "not specified", never invented.
- When Secure-RAG is connected, the Compliance agent receives retrieved passages and the page lists them. Without it, findings are labelled "model knowledge: verify".
- Residual ratings are projections that assume the controls work; they are not evidence that they do.
- Regulator links point to landing pages maintained by the regulator.
Where ARA fits
| ARA | What can go wrong with this use case, and what to do |
|---|---|
| Secure-RAG | What the authoritative sources say, and whether a RAG application is secure |
| CALIBRE Governance Fabric | How much governance the system needs; register, approvals, runtime authority and evidence |
Each tool works on its own. Connected, the Fabric calls ARA's API for every new initiative, and ARA calls Secure-RAG for evidence.
Client deliverables (v3.1)
| Board pack and risk register | Word board pack (decision, risks, obligations, controls, remediation, framework mapping, simulations, envelope, sign-off) and an Excel risk register, with your branding. |
|---|---|
| Versions and sign-off | Every run is a numbered version sealed with SHA-256. A reviewer signs off a version, which locks the assessment until it is reopened with a reason. Differences between versions are shown. |
| Failure simulator | Model-drafted failure chains; you remove capabilities, set control status and accept risks, and exposure is recalculated by a fixed rule. |
| Operating envelope | Actions, tools, limits, data and stop conditions for an agent; versioned, hashed and pushed to the Fabric. |
| ADM transparency pack | Inventory of automated decisions, the APP 1.7 three-part test, APP 1.8 policy text and evidence file (Privacy Act, from 10 December 2026). |
| Vendor due diligence | Vendor documents checked against 14 AI vendor controls (the same set as Secure-RAG), with rating, questions and clauses. |
| Consistency check | Repeat runs of one scenario with an agreement score. |
| Framework mapping | Controls mapped at heading level to ISO/IEC 42001, NIST AI RMF, ISO/IEC 27001 and OWASP LLM 2025; suggestions to verify. |
Data handling
Scenarios and results are stored in this server's database so they appear in the library. When a language model provider is used, the scenario and intermediate analyses are sent to that provider. Do not enter confidential or personal information in a public demonstration.
Decision support only, not legal or regulatory advice. AI-generated ratings and citations must be reviewed by a qualified person before they are relied on.