How Compass suggestions perform under human review
Compass is assistive. Every suggestion inside RiskAlign is shown to a named risk owner who ratifies, overrides or dismisses it — and every decision is written to a tamper-evident audit trail. This page documents the method. Aggregate ratification telemetry (n =…) will be published from /api/public/compass-methodology once the first paid pilot has decided enough suggestions for the numbers to mean anything.
One operating chain across RegAlign and RiskAlign
Compass uses the same risk taxonomy and audit chain as RegAlign. A finding raised in RegAlign maps to a KRI breach in RiskAlign without re-classification; a control evidenced in RiskAlign satisfies the same obligation in RegAlign. The hash-chain is continuous — auditors verify both products from a single chain head.
See RegAlign + RiskAlign for the bridge contract.
Where Compass is used
- Risk-register triage
Suggests the likely category, owner and linked obligations for a newly logged risk or KRI breach.
Human in the loop: Risk owner accepts, edits or rejects. The suggestion and the decision are both written to the audit trail.
- Scenario assist
Drafts plausible loss-scenario inputs (frequency, severity, recovery time) from the residual posture.
Human in the loop: Scenario owner signs off before any draft scenario can move into the board pack.
- Board-pack narrative
Composes the first draft of the board commentary from the current register, appetite and KRI state.
Human in the loop: CRO ratifies. Draft → published is a deliberate, logged action — never automatic.
Where Compass is never used
- Signing the board pack or any external assurance statement.
- Changing risk appetite thresholds.
- Adjusting residual scores without a named human override.
- Adverse decisions about a real customer or employee.
Governance controls
- Model and version logged on every call (agent_runs table: model_version, prompt_id, input/output hash, token counts, source).
- Deterministic fallback if the AI gateway is unavailable — features degrade, they don't fail open.
- Anti-prompt-injection: user-supplied text is sanitised and wrapped in untrusted delimiters before any model call.
- No training on your data — calls route through the Lovable AI Gateway under no-training contractual terms.
- Per-tenant AI cost cap so spend can never exceed the published budget.
- Tenant administrator can disable Compass entirely from Settings → AI. All workflows still work; suggestions just don't appear.
See also: AI Use Disclosure, Trust Centre, Known Limitations.