Glossary

Agent standing

The first pipeline stage. Confirms the agent is registered to the partner and assesses its standing — trust score and exposure ceiling. Agent IDs are self-declared then ownership-checked, not cryptographically authenticated.

Governance Pipeline

The six-stage chain that every payment sent to Revvyn flows through: Agent standing → Orchestrator → Governance → Control → Execution → Audit. Each agent can CONTINUE, REJECT, ESCALATE, or HALT.

Trust Score

A 0-100 score assigned to each registered agent, starting at a baseline. Today it only decreases, on violations — there is no automatic upward or earn path. Below 40 triggers escalation; below 20 triggers rejection.

Hard Payment Cap

The maximum value of a single payment an agent may make. A payment above this amount is rejected outright. This is a per-payment amount cap — not a cumulative, daily, or lifetime budget (Revvyn does not accumulate an agent's spend). Configured per agent in the agent-standing registry; the softer escalate threshold is max_single_payment.

Policy Vertical

A category of business that determines which governance rules apply. Verticals: ai_commerce, treasury, marketplace, enterprise, consumer.

Rail Selection

The Control agent scores all available payment rails (SEPA, SWIFT, ACH, etc.) on cost, speed, compliance, and vertical fit, then selects the optimal path.

Agent Memory

Agents write decisions and events to a per-agent memory store and recall them on later events. Recalled cross-agent concern memory can raise a payment to human review (via the aggregate-risk gate); memory never rewards an agent or raises its trust.

Consensus Signal

A collaborative decision between agents via the Message Bus. When Agent standing and Control agree, Governance can fast-track or escalate. Recorded in the audit trail.

HITL (Human-In-The-Loop)

When a payment is escalated, it requires human review. An operator approves or rejects via the dashboard or API. The pipeline resumes after the decision.

Kill Switch

Emergency mechanism that halts all payment processing across all agents immediately. Persisted to database. Activated via API or dashboard.

Append-Only, Hash-Chained Audit Trail

Every agent decision is recorded in audit_entries with a PostgreSQL trigger preventing UPDATE/DELETE. Each record is stamped with a 7-year retention target; automated deletion after 7 years is not yet implemented. Hash-chain verification detects any edit. Not yet externally anchored, so it does not yet protect against changes by Revvyn's own database administrators.

Anomaly Detection

The per-agent baseline engine detects deviations in each agent's decision behaviour — approval/block/escalate-rate spikes, decision-volume spikes, latency and confidence shifts. Transaction typologies such as geographic and amount patterns are handled by a separate fraud-detector module, not this per-agent baseline.

Circuit Breaker

If a rail fails 5 times in 60 seconds, the circuit opens and Control routes to the next best rail. Prevents cascading failures.

Idempotency Key

A client-chosen value sent in the X-Idempotency-Key header. A retry with the same key within 30 minutes, after the first request has completed, returns the stored response. The request body is not compared, and keys are scoped per account and environment.

Regulatory Framework

A compliance regime that applies based on jurisdiction: FATF (cross-border), PSD2 (EU), MiCA (EU crypto), BSA (US), GDPR (EU data), FCA (UK).