Glossary

Precise definitions of ΔOS terminology

Last reviewed: 2025-01-25

Glossary

Precise definitions of terms used in ΔOS documentation. When terminology is unclear, refer here.

ℹ️
Language Precision

We use specific terms with specific meanings. "Guarantee" means something different from "goal." "Judgment" means something different from "decision." Read carefully.

A

Action

The operation an agent wants to perform. Actions are named using a domain.operation convention (e.g., payment.initiate, database.write). An action only executes after its Intent receives an Allow judgment.

Agent

An autonomous system that initiates actions. Agents may be AI models, automated scripts, or any software that acts with some degree of independence. ΔOS governs agent actions—it doesn't build or run agents.

Allow

A Judgment type indicating the action may proceed. The most common judgment outcome.

Audit Trail

The immutable record of all Intents and Judgments. Hash-chained for integrity verification. Cannot be modified or deleted.

Authority Boundary

The defined scope of what a component (human, system, or LIM) can decide. Enforced by system design, not policy.

B

Block

A Judgment type indicating the action must not proceed. Includes a reason explaining why.

C

Capability

A specific function that ΔOS provides. Capabilities may be affected differently during partial failures.

Context

Additional information provided with an Intent to inform policy evaluation. Part of the evidence bundle.

D

Deterministic

Producing the same output for the same inputs, every time. LIM evaluation is deterministic—replay any Intent with its evidence and get the same Judgment.

E

Escalate

A Judgment type indicating human decision is required. The action cannot proceed until a human approves or rejects.

Escalation

The workflow triggered when an Intent receives an Escalate judgment. Includes routing to the appropriate team, SLA tracking, and decision recording.

Evidence

Context gathered to inform policy evaluation. Includes automatic evidence (agent trust score, system health) and custom evidence (provided by the submitter).

Evidence-Only Mode

Operating mode where Intents are recorded but not blocked. Used for observation and onboarding.

F

Fail Closed

Behavior where failures result in blocking (not allowing) actions. ΔOS fails closed by default.

G

Governance

The process of applying policy to agent actions. Autonomy Governance specifically applies this before execution.

Guarantee

A property enforced by system design that cannot be violated under normal operation. Stronger than a goal or SLO.

H

Human Authority

The capability of humans to define policies, override judgments, and maintain ultimate control. Always preserved.

I

Intent

The fundamental unit of governance. Represents an agent's request to take a specific action. Contains the action, resource, parameters, and context.

Immutable

Cannot be changed after creation. Intents and Judgments are immutable.

J

Judgment

The result of evaluating an Intent against policies. Always one of: Allow, Block, or Escalate.

K

Kill Switch

Emergency mechanism to immediately block all agent actions. Always available, cannot be disabled by the system or LIMs.

L

LIM (Ledger-Integrated Module)

A policy evaluation module that takes an Intent as input and produces a Judgment recommendation as output. LIMs are deterministic and composable.

Ledger

The append-only audit trail where all Intents and Judgments are recorded.

O

Override

A human action that changes a Judgment. Allows overriding blocks or blocking allowed actions. All overrides are recorded.

P

Policy

A rule that defines how Intents should be judged. Expressed through LIM configuration.

Pre-Execution Evaluation

The practice of evaluating policy before an action executes, not after. The core principle of ΔOS.

R

Replay

Re-evaluating a historical Intent with its original evidence to verify the Judgment. Used for auditing and debugging.

Resource

The target of an action. Identified by a path-like string (e.g., users/123, production/api).

S

SLA (Service Level Agreement)

A commitment with contractual implications. See also SLO.

SLO (Service Level Objective)

A measurable target for service performance. Can be missed (rarely) unlike guarantees.

T

Trust Score

A measure of an agent's reliability based on historical behavior. Automatically collected evidence.

V

Value Attribution

The process of measuring and attributing value to governance decisions. Provides observable impact metrics.


Term Usage Guide

Say
Governance
Don't say
Control, Management, Oversight
Governance implies systematic policy application. Control is too vague. Management suggests ongoing administration. Oversight suggests passive watching.
Say
Intent
Don't say
Request, Action, Command
Intent is the governance unit before evaluation. Request is too generic. Action is what happens after Allow. Command implies execution.
Say
Judgment
Don't say
Decision, Result, Outcome
Judgment is the formal output of evaluation. Decision implies discretion. Result and outcome are too generic.
Say
LIM
Don't say
Plugin, Module, Handler
LIM specifically means Ledger-Integrated Module with defined properties. Other terms lack precision.

See Also

For detailed explanations: