Glossary
The terms below appear throughout the docs. They are grouped by the lifecycle act they belong to.
Capture
Decision — A single choice an AI system makes that you want to govern: a classification, a route, an approval. In Briefcase it is recorded as a DecisionSnapshot.
DecisionSnapshot — The structured, persistent record of one decision: its inputs, outputs, model parameters, execution context, and timing. You build it, store it, reload it, and replay it.
@capture — A decorator that records a lightweight dict for every call to a function and hands it to an exporter. The quick path for live observability; it does not persist a DecisionSnapshot on its own.
Input / Output — Typed wrappers (Input(name, value, data_type), Output(...)) around a single named value. An Output can carry a confidence score.
ModelParameters — The model configuration captured at call time: model name, provider, and per-parameter settings. Captured so you can tell when a model change caused output drift.
ExecutionContext — The runtime environment a decision ran in: runtime version, resolved dependencies, random seed, and environment variables. Captured so a replay can run in a comparable environment.
Exporter — A sink that receives captured decision records — console, JSONL file, in-memory, or a custom backend. Wired up with observe().
Fingerprint — A stable hash of a decision’s content (DecisionSnapshot.fingerprint()), used to compare and group decisions.
Control
Guardrail — A control that decides whether an agent may perform an action on a resource, returning an EvalResult. Deny-by-default and side-effect-free. See Guardrails.
EvalRequest / EvalResult / Effect — A guardrail evaluates an EvalRequest (agent, action, resource, context) and returns an EvalResult whose Effect is ALLOW or DENY.
Fail closed — The principle that a control which errors must deny, never allow. An error should never grant access.
Routing decision — The choice of where a decision goes (e.g. which queue or handler), produced by a router. See Routing.
Policy / PolicyVersion / PolicyRule — A policy is a named set of rules. A PolicyVersion is an immutable, published snapshot of those rules; a PolicyRule maps a condition to a choice with a rationale. See Versioned Routing Policy.
Store & Query
Storage backend — A durable store for decisions and snapshots (e.g. SqliteBackend). Queryable via SnapshotQuery. See Storage Adapters.
Valid time — When a fact was true in the real world.
Transaction time — When the system learned about that fact and recorded it.
Bitemporal record — An append-only record carrying both valid time and transaction time, so corrections are new records rather than edits and any past state can be reconstructed. See Bitemporal Storage.
As-of view — A reconstruction of a store as it stood at a chosen transaction (or valid) time — what was known then, without look-ahead.
Snapshot policy — Controls how often an external source is snapshotted (EVERY_CALL, ON_CHANGE, HOURLY, DAILY, WEEKLY).
Manifest — In RAG versioning, the atomic record of which documents and embedding model produced an index, used to detect staleness.
Replay & Verify
Replay — Re-executing a stored decision and comparing the result to the original. Modes are strict (exact) and tolerant (allows minor differences). See Deterministic Replay.
Drift — How much a model’s outputs vary across repeated runs of the same decision. Measured by a consistency score and related metrics. See Drift Detection.
Audit bundle / ExaminerBundle — A self-contained artifact that joins a decision, its bitemporal evidence, and the policy version in effect, sealed with a SHA-256 content hash. See Audit Bundles.
Content hash — The SHA-256 hash that seals a bundle. verify() recomputes it and raises if a single byte changed.