ACGP-1001: Unified Terminology & Definitions¶

Status: Draft
Last Updated: 2026-01-08
Spec ID: ACGP-1001
Normative Keywords: MUST, SHOULD, MAY (per RFC 2119)

Abstract¶

This document provides the authoritative terminology reference for ACGP v1.1. All specifications, documentation, and implementations MUST conform to these definitions.

Namespace Conventions¶

Overview¶

ACGP introduces multiple concepts that use the term "tier" or "level". To prevent ambiguity, all tier/level references MUST be qualified in documentation, APIs, and implementations.

Three Tier Namespaces [NORMATIVE]¶

Primary Terms (use these):

ACL Tier [0-5]: Agent capability and autonomy level
Eval Tier [0-3]: Evaluation strategy depth (governance contracts)
Conformance Level [Minimal/Standard/Complete]: Implementation completeness

Quick Reference Table¶

Concept	Short Form	Long Form	Set At	Example	Defined In
Agent Capability	ACL-2	ACL Tier 2	Design/deployment	"This is an ACL-2 agent"	This document
Evaluation Depth	Eval-1	Eval Tier 1	Runtime (per-request)	"Run Eval-1 checks"	ACGP-1010
Implementation	Standard	Standard Conformance	Deployment	"Standard implementation"	ACGP-1009

Usage in Code [NORMATIVE]¶

Implementations MUST use qualified names in APIs and data structures:

GOOD: Unambiguous

from acgp import ACLTier, EvalTier, ConformanceLevel

agent = Agent(acl_tier=ACLTier.STANDARD_OVERSIGHT)  # ACL-2
request = EvalRequest(eval_tier=EvalTier.FAST_SYNC)  # Eval-1
steward = Steward(conformance=ConformanceLevel.STANDARD)

BAD: Ambiguous

agent = Agent(tier=2)  # Which tier?! ACL? Eval? Conformance?

Usage in Message Formats [NORMATIVE]¶

Message field names MUST be self-documenting:

GOOD: Field names are qualified

{
  "agent_acl_tier": "ACL-2",
  "requested_eval_tier": "eval-1",
  "steward_conformance": "standard"
}

BAD: Requires context to understand

{
  "tier": 2,
  "level": "standard"
}

Usage in Documentation [NORMATIVE]¶

ALWAYS use qualified terms in prose:

GOOD: "An ACL-2 agent requests Eval-1 checks within a 300ms budget."

BAD: "A Tier 2 agent requests Tier 1 checks within a 300ms budget." (Which tier is which?!)

Exception: Context-Obvious Cases¶

When context is absolutely clear within a specific section, you MAY use shorthand only within that section:

## ACL Tiers

Within this section only, "tier" means "ACL Tier" unless otherwise noted.

- Tier 0: No autonomy
- Tier 1: Minimal autonomy
...

Even in these cases, using fully qualified terms is RECOMMENDED for searchability and clarity.

Important Distinctions¶

Understanding the difference between these three concepts is critical:

Conformance Level ≠ ACL Tier ≠ Eval Tier

Conformance Level (Minimal/Standard/Complete) = How much of the ACGP specification an implementation supports
Set at: Implementation/deployment time
Question answered: "How complete is this ACGP implementation?"
Example: "Our system has Standard conformance"
ACL Tier (0-5) = How much risk and autonomy an agent has
Set at: Agent design/deployment time based on ARS assessment
Question answered: "How much oversight does this agent need?"
Example: "This customer service agent is ACL-2"
Eval Tier (0-3) = How deep the governance evaluation goes for each request
Set at: Runtime, per-request based on governance contract
Question answered: "How thoroughly should we evaluate this action?"
Example: "Run Eval-1 checks for this low-risk action"

Real-world example: A Standard conformance implementation can govern agents at ACL-3 tier using Eval-1 checks for routine actions and Eval-2 checks for critical actions.

Core Agent Terminology¶

Governance Steward¶

Primary Term: Governance Steward
Definition: A specialized runtime governance agent that observes Cognitive Traces from Operating Agents, calculates CTQ (Cognitive Trace Quality) scores, and issues interventions to ensure alignment with Reflection Blueprints and the Clarity Baseline.

Operating Agent¶

Primary Term: Operating Agent
Definition: The autonomous or semi-autonomous AI agent being monitored and governed by the Governance Steward. This is the agent executing actual tasks and using tools.

Steward Network¶

Primary Term: Steward Network
Definition: A coordinated network of multiple Governance Stewards working together to provide distributed governance across multi-agent systems.

Cognitive Trace¶

Primary Term: Cognitive Trace
Definition: A structured, comprehensive log of an agent's inputs, internal reasoning steps, tool calls, decisions, and outputs for a single execution step. This is the fundamental unit of observation for governance.

Risk Assessment Framework¶

ARS (Agent Risk Score)¶

Primary Term: Agent Risk Score (ARS)
Definition: A composite risk metric (0-15) that quantifies an agent's risk posture based on three dimensions:
Autonomy (0-5): Degree of self-initiation and tool control
Adaptability (0-5): Capacity to change goals, strategies, or write code
Continuity (0-5): Persistence, memory, and ability to spawn sub-agents
Formula: ARS = Autonomy + Adaptability + Continuity

ACL Tiers¶

Primary Term: ACL Tier
Range: ACL-0 through ACL-5
Definition: The agent's risk posture classification derived from its ARS, determining which security controls and governance mechanisms MUST be in place.

ARS to ACL Mapping¶

ARS Total	ACL Tier	Risk Level	Required Controls
0-2	ACL-0	Minimal	Basic logging, rate limits
3-4	ACL-1	Low	Standard policy rules, human override
5-7	ACL-2	Medium	Warnings, spend caps, approval for sensitive scopes
8-10	ACL-3	High	Approval gates, step-through, immutable audit
11-13	ACL-4	Very High	Crypto signing, isolation, multi-agent RBAC
14-15	ACL-5	Critical	Dual control, kill-switch, pre-commit review

Runtime Intervention System¶

Intervention Types¶

Primary Term: Intervention
Count: Six intervention types (five primary levels: OK, Nudge, Escalate, Block, Halt, plus orthogonal Flag that can be combined with any primary level)
Definition: Runtime decisions made by the Policy Engine based on CTQ scores and tripwires.

Six Standard Interventions¶

OK
Action: Allow (log only)
Semantics: Action is permitted to proceed with standard logging
Nudge
Action: Replace-or-allow-with-feedback
Semantics: Modify the action (e.g., redact PII, tighten prompt) and continue
Flag
Action: Allow + high-severity audit
Semantics: Permit action but increment trust debt; may degrade privileges later
Note: Orthogonal to other interventions (can be combined with OK, Nudge, Escalate, Block, or Halt)
Escalate
Action: Pause + require human approval
Semantics: Enter step-through mode pending Human-In-The-Loop decision
Block
Action: Block this specific action
Semantics: Prevent action but session continues
Halt
Action: Kill-switch activation
Semantics: Revoke keys, disable session/agent, cascade to sub-agents

Decision Thresholds by ACL Tier¶

Note: These thresholds use Risk Score (1.0 - CTQ), not CTQ directly.

ACL	OK	Nudge	Escalate	Block	Halt
ACL-0	<=0.40	0.40-0.55	0.55-0.70	>0.70 or tripwire	Critical tripwire only
ACL-1	<=0.30	0.30-0.45	0.45-0.60	>0.60 or tripwire	Critical tripwire
ACL-2	<=0.25	0.25-0.40	0.40-0.55	>0.55 or tripwire	Critical tripwire
ACL-3	<=0.20	0.20-0.35	0.35-0.50	>0.50 or tripwire	Critical/repeated tripwire
ACL-4	<=0.15	0.15-0.30	0.30-0.45	>0.45 or tripwire	Any critical tripwire
ACL-5	<=0.10	0.10-0.25	0.25-0.40	>0.40 or tripwire	Any severe tripwire

Tripwires¶

Primary Term: Tripwire
Definition: Critical conditions that trigger immediate intervention regardless of CTQ score
Categories: Standard, Critical, Severe
Examples:
Secrets exfiltration
Irreversible production changes
Regulated data movement
Fraud signals
Agent collusion attempts

Governance Contracts (ACGP-1010)¶

Risk Levels¶

Primary Terms: low_risk, elevated_risk, critical_risk
Definition: Per-action classification of consequences, independent of ACL Tier
Values:
low_risk: Easily reversible or no side effects (default)
elevated_risk: Has consequences but not catastrophic
critical_risk: Difficult/impossible to reverse or high-value target
See: ACGP-1010 Section 2

Evaluation Tiers¶

Primary Term: Eval Tier or Evaluation Tier
Range: Eval-0 through Eval-3
Definition: Governance evaluation strategy depth, independent of ACL Tier
Values:
Eval-0: Must-pass synchronous checks (<100ms, no external deps)
Eval-1: Fast synchronous policy checks (<300ms target)
Eval-2: Deep asynchronous analysis (LLM-based, seconds to minutes)
Eval-3: Human-in-the-loop approval (human time)
Note: NOT the same as ACL Tiers or Conformance Levels
See: ACGP-1010 Section 3

Performance Budget¶

Definition: Per-request latency contract specifying how long the agent will wait for governance decision
Components: latency_budget_ms, fallback_on_timeout, tier_budgets
Semantics: Timeout with explicit fallback behavior (not an SLA)
See: ACGP-1010 Section 4

Fallback Behaviors¶

Primary Terms: deny, allow_and_log, cached_decision, escalate
Definition: Agent behavior when governance times out or is unavailable
Values:
deny: Block the action (safety > availability)
allow_and_log: Allow with heavy logging (availability > safety)
cached_decision: Use cached policy if available, else deny
escalate: Move to higher eval tier or human approval
See: ACGP-1010 Section 4.4

Governance Status¶

Primary Terms: OK, GOVERNANCE_TIMEOUT, PARTIAL_EVAL, UNAVAILABLE
Definition: Steward response indicating governance evaluation result
Values:
OK: Evaluation completed within budget
GOVERNANCE_TIMEOUT: Budget exceeded, returning partial/no result
PARTIAL_EVAL: Completed some tiers, others incomplete
UNAVAILABLE: Steward cannot evaluate at all
See: ACGP-1003 Messages

Quality Metrics¶

CTQ (Cognitive Trace Quality)¶

Primary Term: Cognitive Trace Quality (CTQ)
Definition: A normalized score (0-1) representing the quality, safety, and compliance of a Cognitive Trace against defined blueprints and baselines.
Calculation: Weighted average of individual metric scores

Risk Score¶

Definition: The inverse of CTQ used for intervention decisions
Formula: Risk_Score = 1.0 - CTQ_final
Range: 0.0 to 1.0 (0 = perfect quality, 1 = complete failure)

CTQ Vector¶

Definition: The array of individual quality metric scores before aggregation
Five Standard Metrics: Reasoning Quality, Knowledge Grounding, Ethical Alignment, Tool Safety, Context Awareness
Example: [reasoning: 0.90, grounding: 0.88, ethical: 0.95, tool_safety: 0.92, context: 0.89]

CTQ Final¶

Definition: The weighted aggregate score derived from the CTQ Vector
Range: 0.0 to 1.0
Note: Higher is better (opposite of Risk Score)

Trust System¶

Trust Debt¶

Definition: Accumulated risk score from flagged behaviors
Accumulation: Based on flag severity:
Low severity: +0.1 trust debt
Medium severity: +0.3 trust debt
High severity: +0.5 trust debt
Decay: Trust debt decays by 5% every 24 hours (multiplied by 0.95)
Formula: debt_current = debt_initial × (0.95^days_elapsed)
Example: After 7 days, a 0.5 debt becomes 0.5 × (0.95^7) ≈ 0.348
Effect: Triggers re-tiering when exceeding thresholds

Trust Debt Thresholds by ACL¶

ACL Tier	Warning Level	Re-tier Trigger	Suspension Trigger
ACL-0	2.0	5.0	10.0
ACL-1	1.5	4.0	8.0
ACL-2	1.0	3.0	6.0
ACL-3	0.75	2.0	4.0
ACL-4	0.5	1.5	3.0
ACL-5	0.25	1.0	2.0

Dynamic Re-tiering¶

Definition: The automatic process of recalculating ARS and adjusting ACL tier when agent capabilities, goals, or trust scores change significantly
Triggers: Trust debt threshold, capability changes, repeated violations
Direction: Can increase (escalate) or decrease (de-escalate) ACL tier

Governance Artifacts¶

Reflection Blueprints¶

Primary Term: Reflection Blueprints
Definition: YAML/JSON configuration files that specify expected agent behavior, quality metrics, thresholds, checks, and evidence requirements for specific domains and goals.

Clarity Baseline¶

Primary Term: Clarity Baseline
Definition: A mandatory, cross-domain blueprint that evaluates reasoning clarity, logical consistency, knowledge grounding, and absence of cognitive biases. Acts as the universal "cognitive seatbelt" for all agents.

Certified Source Registry¶

Primary Term: Certified Source Registry
Definition: A curated, cryptographically signed catalog of authoritative data sources with metadata including authority level, jurisdiction, recency, and relevance scores.

System Components¶

ReflectionDB¶

Primary Term: ReflectionDB
Definition: An append-only, immutable event store that maintains a complete audit trail of all governance events, evaluations, and interventions.

Policy Engine¶

Primary Term: Policy Engine
Definition: The runtime component that evaluates CTQ scores against ACL-specific thresholds and issues intervention decisions.

Trusted Monitor (TM)¶

Primary Term: Trusted Monitor
Definition: An independent observer model/process that validates agent behavior, detects anomalies, and can trigger interventions.

Human-In-The-Loop¶

Primary Term: Human-In-The-Loop
Abbreviated: HITL (only in diagrams/tables where space is limited)
Definition: The escalation pathway that routes critical decisions to human operators with appropriate context and evidence for manual review.

Protocol Components¶

MPA (Multi-Party Authorization)¶

Primary Term: Multi-Party Authorization (MPA)
Definition: A security mechanism requiring multiple authorized approvers for sensitive changes to blueprints, registry entries, or critical configurations.

ATS (Agent Telemetry Standard)¶

Primary Term: Agent Telemetry Standard (ATS)
Definition: The standardized event format for semantic telemetry including traces, decisions, and audit records.

Operational Terms¶

Mission Envelope¶

Definition: The defined boundaries of allowed tools, resources, data domains, and budgets for a specific agent.

Message Retry Policy¶

Definition: Standard behavior for handling failed messages
Specification:

retry_policy:
  max_attempts: 3
  backoff: exponential
  timeout_ms: 500
  failure_action: escalate

Emergency Override¶

Definition: Authorized procedure to immediately halt all agent operations
Requirements: Dual authorization for ACL-3+, single authorization for ACL-0-2

Open Source vs Proprietary Boundary¶

Open Source Components¶

Protocol specifications
Message formats (TRACE, EVAL, INTERVENTION, SYNC)
ATS event schemas
CTQ weighted average formula
Metric weights (MUST be disclosed)
Policy Engine decision mapping
Validator and conformance suite
Reference implementations (basic)

Proprietary Components¶

Individual metric scoring algorithms (MAY be proprietary)
Feature engineering methods
Anomaly detection models
Threshold auto-tuning logic
Fusion algorithms for multi-metric evaluation

This document serves as the authoritative terminology reference for ACGP v1.1. All specifications, documentation, and implementations MUST conform to these definitions.