Governance Primitives Catalog

34 runtime governance primitives for agentic AI systems. Each primitive is a composable, stdlib-only Python module. Published as hummbl-governance v1.4.2 on PyPI. Apache 2.0. Zero third-party Core dependencies.

Install: pip install hummbl-governance · Source: github.com/hummbl-io/oss

Safety primitives

The core safety infrastructure that prevents catastrophic failure.

KillSwitch

Safety · Emergency stop

4-mode emergency stop (DISENGAGED, HALT_NONCRITICAL, HALT_ALL, EMERGENCY). Immediately halts agent operations. The last line of defense when an agent behaves unpredictably.

CircuitBreaker

Safety · Failure isolation

3-state failure isolation (CLOSED, HALF_OPEN, OPEN). Automatically trips after N consecutive failures, preventing cascade. Auto-probes recovery before re-enabling.

CapabilityFence

Safety · Action scoping

Restricts agent actions to an explicit allowlist. Agents cannot perform operations outside their fence. Denies by default.

PhysicalSafetyMode

Safety · Physical systems

Safety monitor for AI systems that interact with physical hardware. Enforces kinematic constraints and physical safety bounds.

Authorization primitives

Who can do what, and who said they could.

DelegationToken

Authorization · Scoped delegation

HMAC-SHA256 signed delegation tokens with chain-depth enforcement. An agent can only act within the scope and depth of its delegation chain.

DelegationContext

Authorization · Context binding

Binds delegation tokens to an operational context (task, agent, session). Prevents token reuse across contexts.

IdentityEngine

Authorization · Identity

Agent identity registry with trust tiers. Verifies agent identity before authorizing actions. Supports revocation and trust adjustment.

AgentRegistry

Authorization · Roster

Centralized registry of authorized agents. Agents not in the registry cannot receive delegation tokens or perform governed actions.

Observability primitives

What happened, when, and who can verify it.

AuditLog

Observability · Append-only log

Append-only JSONL audit log with hash-chain integrity. Every governed action is recorded. Tampering breaks the chain and is immediately detectable.

LamportClock

Observability · Causal ordering

Lamport logical clock for causal ordering of events across distributed agents. Determines what happened before what without relying on wall-clock time.

HealthProbe

Observability · Health checks

Multi-probe health checking (8 probes: kill switch, circuit breaker, bus, cognition, cost, security, adapters, scheduler). Unified health endpoint for monitoring.

ToolCallAuditor

Observability · Tool calls

Audits every tool call an agent makes — inputs, outputs, timing, and authorization. Detects suspicious patterns and unauthorized tool usage.

TransitionReceipt

Observability · State transitions

Cryptographic receipts for state transitions. Proves that a transition occurred, who authorized it, and what the before/after states were.

Validation primitives

Checking that outputs are safe before they reach users.

OutputValidator

Validation · Output checking

Validates agent outputs against schema, length bounds, and content filters. Blocks outputs that fail validation before they reach users.

SchemaValidator

Validation · Schema enforcement

JSON schema validation for agent inputs and outputs. Rejects malformed data before it enters the system.

InjectionDetector

Validation · Prompt injection

Detects prompt injection attempts in agent inputs. Identifies known injection patterns and blocks them before processing.

PIIDetector

Validation · PII filtering

Detects personally identifiable information in agent outputs. Prevents PII leakage in logs, audit trails, and user-facing responses.

BlocklistFilter

Validation · Content filtering

Filters agent outputs against configurable blocklists. Supports allowlist and blocklist modes for fine-grained content control.

Coordination primitives

How multiple agents work together safely.

BusWriter

Coordination · Message bus

Append-only TSV coordination bus with flock-based locking. Agents communicate via typed messages (PROPOSAL, ACK, STATUS, DECISION, etc.) with durable audit trail.

ContractNetManager

Coordination · Task allocation

Contract-net protocol for multi-agent task allocation. Agents bid on tasks; the manager awards contracts based on bids. Supports multi-phase contracts.

ConvergenceGuard

Coordination · Goal convergence

Monitors whether multi-agent systems are converging on shared goals. Alerts when agents diverge or when convergence stalls.

Economic primitives

Cost control and incentive alignment.

CostGovernor

Economic · Budget enforcement

Per-agent budget enforcement with configurable caps. Tracks token costs, API calls, and compute usage. Halts agents that exceed budgets.

RewardMonitor

Economic · Incentive tracking

Monitors reward signals in RL and feedback-driven systems. Detects reward hacking, drift, and misalignment between stated and observed rewards.

Compliance primitives

Mapping governance actions to regulatory frameworks.

ComplianceMapper

Compliance · Framework crosswalk

Automated crosswalks between NIST AI RMF, EU AI Act, ISO 42001, and NIST CSF 2.0. Maps governance primitive usage to compliance requirements.

StrideMapper

Compliance · Threat modeling

STRIDE threat modeling for AI systems. Maps each STRIDE category (Spoofing, Tampering, Repudiation, Info disclosure, DoS, Elevation) to governance primitives that mitigate the threat.

Lifecycle primitives

Governing the full lifecycle of agents and decisions.

GovernanceLifecycle

Lifecycle · Stage management

Manages the lifecycle stages of governed entities (PROPOSED, ADMITTED, ACTIVE, DEPRECATED, RETIRED). Enforces transition rules and records each transition.

EvolutionLineage

Lifecycle · Version tracking

Tracks the evolution of governed entities over time. Records modifications, variants, and drift. Enables rollback to previous versions.

FailureModes

Lifecycle · Failure tracking

Catalogs failure modes for governed systems. Each failure mode has a severity, detection method, and mitigation strategy. Links failures to error records.

Attest

Lifecycle · Attestation

Cryptographic attestation of system state. Proves that the system was in a specific configuration at a specific time.

Kernel primitives

Meta-governance — the governance system that governs the governance system.

Kernel

Kernel · Meta-governance

The governance kernel that coordinates all other primitives. Enforces invariants, manages admission, and coordinates the authority/law/evidence engines.

ReceiptEngine

Kernel · Receipt generation

Generates HMAC-signed receipts for governed actions. Each receipt binds a contract, delegation context, and evidence into a tamper-evident record.

ReasoningEngine

Kernel · Governance reasoning

Applies governance rules to determine whether an action is permitted. Combines authority, law, and evidence to produce authorization decisions.

CorpusAdapter

Kernel · Knowledge access

Governed access to knowledge corpora. Ensures agents reference approved sources and records what was accessed for audit trails.

Statistical primitives

StatisticalFramework

Statistical · Calibration

Calibration framework for tracking prediction accuracy over time. Compares predicted outcomes to actual outcomes and generates calibration reports.

EAL

Statistical · Evidence-based assessment

Evidence Assessment Level framework. Validates and revalidates evidence quality for governance decisions.

See also: how these primitives are operationalized in production and the methods of organization that structure them.

Want to integrate these primitives into your system?

Book a 30-min call →