Skip to main content

Glossary

This page provides definitions of key terms and concepts used throughout the OpenBox documentation.


Activity

A single unit of work inside a Workflow — calling an LLM, querying a database, invoking a tool, or making an HTTP request. Activities are where side effects happen in Temporal.

OpenBox connection: OpenBox captures the inputs and outputs of every Activity execution, evaluates governance policies against them, and records a Governance Decision for each one.

Learn more: Temporal 101


Risk Profile

A risk scoring framework that evaluates AI agents across three weighted categories:

CategoryWeightParameters
Base Security25%Attack vector, attack complexity, privileges required, user interaction, scope
AI-Specific45%Model robustness, data sensitivity, ethical impact, decision criticality, adaptability
Impact30%Confidentiality, integrity, availability, safety impact

The resulting score (0% – 100%) determines the agent's initial Trust Tier.

Learn more: Assess Phase


Alignment Score

A 0% – 100% metric measuring how well an agent's actions match its stated goals. Contributes 25% of the overall Trust Score.

RangeRatingMeaning
90% – 100%ExcellentActions strongly aligned with goals
70% – 89%GoodMinor deviations, acceptable
50% – 69%WarningNotable drift, review recommended
Below 50%MisalignedSignificant deviation, action required

Learn more: Verify Phase


Attestation

Cryptographic, tamper-proof evidence for every governance session. Each session's events are hashed into a Merkle Tree and digitally signed, creating a verifiable Proof Certificate that confirms no governance data was altered after the fact.

Signing providers:

  • AWS KMS — ECDSA NIST P-256 (default)
  • External Attestation — Your own signing service (e.g., TEE, HSM)

Learn more: Attestation & Cryptographic Proof


Behavioral Rules

Stateful authorization rules that detect multi-step patterns across an agent's session. Unlike Policies, behavioral rules track prior actions to identify sequences, frequencies, or combinations.

Pattern examples:

  • Sequence: PII access → External API call (without approval)
  • Frequency: More than 10 failed auth attempts in 1 minute
  • Combination: Database write + File export + External send

Rules are evaluated in priority order and stop at the first rule that triggers a verdict.

Learn more: Authorize Phase: Behavioral Rules


Behavioral Score

A 0–100 runtime compliance metric that starts at 100 for new agents and decreases when policy violations occur. Contributes 35% of the overall Trust Score.

Violation SeverityPenaltyTrust Score Impact
Minor-5 pts-1.75 pts
Major-15 pts-5.25 pts
Critical-25 pts-8.75 pts

Learn more: Trust Scores


Governance Decision

The outcome produced when OpenBox evaluates an agent operation. There are four possible decisions, listed in precedence order:

DecisionEffectImpact
HALTTerminates the entire agent sessionSignificant negative
BLOCKAction rejected, agent continuesNegative
REQUIRE_APPROVALOperation paused for HITL reviewNeutral (pending)
ALLOWOperation proceeds normallyPositive

Precedence: HALT > BLOCK > REQUIRE_APPROVAL > ALLOW

Learn more: Governance Decisions


Guardrail

Pre- or post-processing validation and transformation rules applied to agent inputs and outputs. Multiple guardrails execute as a chained pipeline — the output of one feeds into the next.

Types:

  • Input Guardrails — Validate/transform incoming data (PII detection, rate limiting)
  • Output Guardrails — Validate/transform responses (PII redaction, format enforcement)

Learn more: Authorize Phase: Guardrails


HITL (Human-in-the-Loop)

A workflow pattern where an agent operation is paused pending human approval before proceeding. Triggered by a REQUIRE_APPROVAL governance decision.

Approval actions:

ActionResultTrust Score Effect
ApproveOperation proceedsNo change
RejectOperation blocked (reason required)-2 pts
TimeoutOperation blocked after 5 minutes-1 pt

Learn more: Approvals


Merkle Tree

A cryptographic data structure used to combine individual event hashes into a single session root hash. Uses SHA-256 with sorted-pair hashing to ensure consistent tree construction regardless of processing order. Enables tamper-proof verification of governance events.

Learn more: Attestation & Cryptographic Proof


Policy (OPA/Rego)

Stateless permission checks written in OPA (Open Policy Agent) Rego language. Policies evaluate an input document at runtime and return a governance decision (CONTINUE or REQUIRE_APPROVAL) with an optional reason.

Unlike Behavioral Rules, policies are stateless — they evaluate each operation independently without tracking prior actions.

Learn more: Authorize Phase: Policies


Proof Certificate

A per-session attestation record produced after an agent session completes. Contains:

FieldDescription
Merkle RootSHA-256 root hash of the session's event Merkle Tree
SignatureDigital signature of the Merkle root (from AWS KMS or external provider)
Event CountNumber of governance events included in the attestation

Learn more: Attestation & Cryptographic Proof


Query

A synchronous, read-only request to inspect a running Workflow's state without affecting its execution. Queries return a value but never change the Workflow.

OpenBox connection: OpenBox uses Queries to inspect governance state during execution — for example, checking whether an approval is still pending. Queries do not trigger governance evaluation since they are read-only.

Learn more: Temporal 101


Session Replay

An interactive dashboard view showing the complete execution timeline of an agent session. Allows step-by-step walkthrough of every operation with its governance decision (ALLOW, BLOCK, REQUIRE_APPROVAL, etc.).

Learn more: Session Replay


Signal

An asynchronous message sent to a running Workflow from the outside. Signals let external systems inject data or trigger decisions mid-execution — for example, delivering a human approval result back to a paused agent.

OpenBox connection: OpenBox captures Signal data and evaluates governance policies on every Signal received. This is how HITL approvals flow back into the Workflow when a REQUIRE_APPROVAL decision pauses execution.

Learn more: Temporal 101


Task Queue

A named channel that connects Workflow/Activity starters to Workers. When you start a Workflow on a Task Queue, only Workers listening on that same queue will pick it up.

OpenBox connection: OpenBox preserves your existing Task Queue configuration. The wrapped Worker polls the same queue your original Worker used — governance is transparent to task routing.

Learn more: Temporal 101


Trust Layer

The governance layer OpenBox adds alongside your workflow engine. It provides trust scoring, policy enforcement, monitoring, and compliance evidence — without modifying your existing workflow code. Your workflow engine remains the system of record for execution.

Learn more: What is OpenBox?


Trust Lifecycle

OpenBox's 5-phase governance model for establishing, maintaining, and evolving trust in AI agents:

  1. Assess — Establish baseline risk profile
  2. Authorize — Define guardrails, policies, and behavioral rules
  3. Monitor — Observe runtime behavior and telemetry
  4. Verify — Check goal alignment and session integrity
  5. Adapt — Evolve trust based on observed patterns

Learn more: Trust Lifecycle Overview


Trust Score

A 0–100 metric representing an agent's overall trustworthiness, calculated from three components:

ComponentWeightDescription
Risk Profile Score40%Inherent risk profile
Behavioral Score35%Runtime compliance
Alignment Score25%Goal consistency

Formula: Trust Score = (Risk Profile × 40%) + (Behavioral × 35%) + (Alignment × 25%)

Learn more: Trust Scores


Trust Tier

One of four trust levels derived from the Trust Score that determines how strictly an agent is governed:

TierRisk Profile RangeRisk LevelDescription
Tier 10% – 24%LowMinimal oversight, most operations auto-approved
Tier 225% – 49%MediumStandard controls, approval for sensitive operations
Tier 350% – 74%HighEnhanced monitoring, stricter enforcement
Tier 475% – 100%CriticalStrict controls, frequent HITL, rate limiting

Learn more: Trust Tiers


Worker

A process that hosts your Workflow and Activity code and polls Temporal for tasks to execute. You start a Worker, register your Workflows and Activities on it, and it handles execution.

OpenBox connection: The Worker is the single integration point. You replace Temporal's Worker with create_openbox_worker — one code change that wraps the Worker with the Trust Layer. No changes to your Workflows or Activities.

Learn more: Temporal 101


Workflow

A durable function that orchestrates a sequence of steps. If the process crashes mid-execution, Temporal replays the Workflow from its event history so it can resume exactly where it left off.

OpenBox connection: When a Workflow starts, OpenBox creates a governance session. When it completes or fails, OpenBox closes the session and triggers Attestation. Every Workflow execution maps 1:1 to a governance session in your dashboard.

Learn more: Temporal 101