Glossary
The vocabulary of governed AI on the desk.
Fourteen terms that show up across the platform pages, articles, and audit rows — defined without the marketing gloss.
- Policy gate
- A single pure function, decide(action, context) → Decision, that every governed action passes through. It performs no I/O and holds no state, so the same inputs always produce the same decision and any historical decision can be replayed from its audit row.
- Policy gate →
- Tier routing
- Classifying an action by how reversible it is rather than by category. T0 is read-only, T1 is reversible writes, T2 requires a single human approval above materiality, and T3 is dual-HITL for irreversible actions.
- Policy gate →
- HITL approval envelope
- A server-generated, HMAC-signed authorization scoped to one exact action — ISIN, notional, and side — with a 90-second expiry. The agent requests it but never sees the signing key, so it cannot forge or re-use one.
- HITL envelopes →
- scope_hash
- A sha256 over (action || isin || notional || side) embedded in an approval envelope. It cryptographically binds the approval to a single action, so re-using the envelope for a different order fails verification.
- HITL envelopes →
- Materiality ledger
- A session-aggregate record of cumulative notional. It closes the split-trade bypass: HITL is triggered when the running total crosses the materiality threshold even if no individual slice does.
- HITL envelopes →
- Append-only audit chain
- A Postgres audit table where BEFORE UPDATE/DELETE triggers reject mutations at the database layer, so immutability survives the application being compromised or someone holding psql access.
- Audit chain →
- Hash chain (sha256 row chaining)
- Each audit row stores sha256(prev_hash || payload || seq). Inserting a row between two existing rows breaks the chain at the join point, so tamper-evidence is verifiable in a single recursive SQL query.
- Audit chain →
- framework_ref
- A NOT NULL column on every audit row that points to the FINOS AIGF v2.0 risk and mitigation the decision satisfies. A CHECK constraint validates it against the live risk taxonomy, so a mis-mapped call fails the write.
- AIGF coverage →
- FINOS AIGF v2.0
- The FINOS AI Governance Framework, version 2.0 — a taxonomy of 23 AI risks and 23 mitigations for agentic AI in financial services. BondFoundry is a working reference implementation of it.
- FINOS mapping →
- AIGF coverage gate
- A CI check that runs the AIGF coverage report on every PR. The build fails if coverage drops below 85% or any AIGF v2.0 risk has zero passing eval cases — governance enforced as continuous integration.
- AIGF coverage →
- Four-dimension eval
- The eval harness measured across accuracy (pricing and risk vs. QuantLib), policy (adversarial tier/HITL bypass attempts), robustness (prompt injection and jailbreak), and latency (per-tier SLO commitments).
- AIGF coverage →
- Segregation of duties (SoD)
- The requirement that the approver of an action differs from the caller. BondFoundry enforces SoD at the API layer; a self-approval attempt returns 403 and is itself written to the audit log.
- HITL envelopes →
- Verbatim rule citation
- The exact policy text returned on a blocked or HITL-routed decision — not a paraphrase or model-generated explanation. The CI suite asserts a character-for-character match against the policy bundle.
- Policy gate →
- MCP boundary
- The tool-call interface between the agent loop and the model. Because the model talks to the system through tool calls, swapping the frontier model is a configuration change rather than a re-architecture.
- For engineers →
See the terms in action