Glossary¶
Key terms and definitions in the Agentic Cognitive Governance Protocol.
A¶
Governance Tier
Defines the default runtime enforcement strictness applied to an agent. Governance Tier values remain serialized as GT-0 through GT-5 in v1.0 alpha. Each tier has default risk-score thresholds (ACGP-3 §7.2). When both blueprint and Governance Tier thresholds exist, the stricter (lower) boundary applies.
Agent
An AI system that makes decisions and takes actions autonomously.
ARI Framework (Agent-Reflection-Intervention)
The core governance model of ACGP. ARI is the conceptual model describing the three-phase flow: agents create cognitive traces, governance stewards reflect on them, and appropriate interventions are issued. Not to be confused with ARS.
ARS (Agent Risk Score)
Numerical score (0–15) measuring agent autonomy, adaptability, and continuity across three dimensions (each 0–5). ARS is a deployment-time score used to determine the initial Governance Tier baseline for an agent. Steward implementations MAY maintain a stricter effective runtime posture later, but v1.0 alpha does not standardize lowering governance below the ARS-derived baseline. ARI is the model; ARS is the metric.
Audit Log
Immutable record of all governance events and decisions.
B¶
Blueprint
Configuration defining governance policies, evaluation criteria, and intervention rules for specific use cases. Expressed in YAML 1.2 or JSON. See ACGP-3 §2.
Block
Intervention level that prevents a specific action from executing. Wire value: "block".
C¶
Cognitive Trace
Structured record of an agent's reasoning and planned action, submitted for governance evaluation.
Cognitive Trace Quality (CTQ)
Score (0.0–1.0) measuring the quality of an agent's reasoning and planned actions across five standard metrics: reasoning quality, knowledge grounding, ethical alignment, tool safety, and context awareness.
Conformance Profile
For v1.0.0-alpha.2, ACGP has one active claimable profile: Standard. The Regulated Controls Badge is additive, Dev Mode is non-conformant, and Safety-Critical remains future-track / non-claimable in alpha.
E¶
Escalate
Intervention level requesting human approval before proceeding. Wire value: "escalate".
Evaluation
Process of assessing cognitive trace quality and determining appropriate intervention.
F¶
Flag
Orthogonal governance modifier that can be combined with any primary intervention level. On the wire, flagging is expressed as flags.flagged: true alongside the primary decision (e.g., "decision": "ok" + "flags": {"flagged": true}). "decision": "flag" is not a valid standalone wire value. When present, flagging adds the configured trust_debt.accumulation.flag weight on top of the primary decision's trust-debt accumulation. See ACGP-1 §3.4 and ACGP-3 §9.
G¶
Runtime Governance Contracts
Human-facing term for the optional per-action runtime negotiation capability carried through the reserved governance_contract extension field. The wire field and implementation type names remain unchanged for compatibility.
Governance Steward
Component responsible for evaluating cognitive traces and issuing interventions.
H¶
Halt
Highest intervention level, stopping the agent completely due to critical violations. Produced only by tripwires, never by threshold mapping. Wire value: "halt".
I¶
Intervention
Proportionate response to cognitive trace evaluation. Five primary levels (ok, nudge, escalate, block, halt) plus the orthogonal flag modifier. Decision values on the wire are always lowercase. SDKs MAY expose idiomatic enum names (e.g., InterventionType.OK) but MUST serialize as lowercase.
N¶
Nudge
Intervention level that suggests modification — the agent MUST acknowledge, and MAY retry or proceed. Wire value: "nudge".
O¶
OK
Intervention level indicating action can proceed without concerns. Wire value: "ok".
T¶
Tripwire
Hard limit that triggers immediate intervention when exceeded, regardless of reasoning quality. Condition expressions describe the violation pattern: if the condition evaluates to true, the tripwire fires.
Trust Debt
Accumulated governance debt (0.0+) tracking intervention history for an agent. Accumulates based on configurable per-decision weights (e.g., nudge: 0.5, block: 2.0, halt: 5.0) plus any configured orthogonal flag accumulation weight when flags.flagged is true. Decays over time using canonical decay_fraction semantics, where the configured fraction is the portion removed each period (e.g., 0.05 means 5% removed per period). When debt exceeds graduated thresholds (elevated_monitoring, restricted_mode, re_tiering_review), corresponding actions trigger and Governance Tier review may be required. See ACGP-3 §9.
Trust Score (Illustrative, Non-Normative)
Future-looking reliability model (0.0–1.0) shown only for teaching or aspirational design context. It is not part of the v1.0 conformance surface. Normative runtime behavior, threshold handling, and governance review flows use Trust Debt instead.
See ACGP-1: Core Concepts & Terminology for complete definitions