agenticaf-community/aaf-epistemic-gates

AAF Epistemic Gates

An epistemic gate is the point where an AI output stops being a possibility and becomes something the system (or a human) treats as knowledge, decision, or action. This skill helps place and scale those gates.

Core rule

Epistemic gates must scale with risk.

• Low-stakes use can tolerate lighter validation gates.
• High-stakes use requires strong, unavoidable gates and explicit accountability.

Three phases (separate them)

1. Generation

The model produces candidates: drafts, options, plans, hypotheses, tool arguments. At this stage treat outputs as candidates, not truth.

2. Validation

Truth and constraints are reintroduced through mechanisms the model does not control:

• Deterministic checks (schemas, assertions, policies)
• Grounding (databases, verified APIs, retrieval with attribution)
• Human review for high-risk classes

3. Authority

A defined actor (human, policy engine, or supervisor under strict rules) converts validated output into action. If authority is implicitly granted to the model (“because the AI said so”), the epistemic boundary has collapsed.

Where to place gates

• Before any write or irreversible action: Validate arguments and policy; require approval when risk is high.
• Before treating model output as “done”: Check against Definition of Done and collect evidence (state, checksums, logs).
• At handoffs: When one agent or step passes output to another, treat the handoff as a gate (validation + scope).
• At external boundaries: Before data or commands leave the system (APIs, tools, user-facing decisions).

Design questions

• Where are the boundaries between generation → validation → authority in this flow?
• Which gates are deterministic (schema, policy engine) vs human approval vs automated-but-auditable?
• Who or what is authorized to convert validated output into action at each risk level?
• What evidence is required before authority is conferred?

Escalation as a gate

When the agent cannot verify, or action is high-risk or irreversible, the system should stop and escalate—not proceed. Escalation should be a first-class interrupt (e.g. APPROVAL_REQUIRED) with a structured packet: proposed action, risk class, policy binding, evidence so far, resume semantics. This makes approvals auditable and resumption deterministic.

Outcome governance link

“Done” cannot be a narrative claim. Every task class should have a Definition of Done, acceptance checks, and evidence artifacts. Validation gates should be deterministic where possible; verification evidence should be surfaced to operations.

Glossary (AAF)

• Policy gate: Permission/authorization decision (e.g. “allowed to call this tool?”).
• Validation gate: Correctness check against truth/constraints (schema, tests, evidence).
• Epistemic gate: The boundary where output transitions from candidate → validated → authoritative decision/action.

Additional resources

• What is an agent (epistemic gates): docs/03-what-is-an-agent.md (§2.5)
• Autonomy & outcome governance: docs/13-autonomy-governance.md
• Executive summary: docs/01-executive-summary.md
• Whitepaper: https://agenticaf.io/