The package, listed in full · hummbl-governance v1.5.1
The 51 governance primitives — and the 21 invariants they answer to
hummbl-governance is an open-source Python package of governance controls for AI agents: who is allowed to do what (permission), what actually happened (proof), and who answers for it (responsibility). It is Alpha software, published on PyPI under Apache-2.0, with no third-party runtime dependencies. External adopters on record: none — the package is in production use inside HUMMBL's own fleet, and that is the whole of the adoption claim.
Below is every primitive in the current release (v1.5.1), the invariants they implement, and — stated plainly — what is not built yet. Known defects in this snapshot are carried on the proof page and in the claim ledger.
Named grain · 21 invariants
The invariants
Fourteen kernel invariants, enforced by the package's kernel:
| ID | Name | The rule, in one line |
|---|---|---|
| K1 | RECEIPT | Every action that affects shared state produces a structured, signed receipt |
| K2 | LAW | Every receipt is evaluated against at least one scaling law |
| K3 | IDENTITY | Every agent has a single canonical identity, trust tier, and capability vector |
| K4 | TEMPORAL | Every receipt carries a sequence ID giving total order within its agent context |
| K5 | EVIDENCE | Every claim in a receipt is graded, or marked speculative |
| K6 | AUTHORITY | Every exercise of authority is scoped, limited, and leaves a receipt |
| K7 | ROLE | Every role is a runtime claim, not a static assignment |
| K8 | DOCTRINE | Every fleet artifact respects the doctrine invariants D1–D7 |
| K9 | REVERSIBILITY | Every durable mutation declares a rollback path, or is marked irreversible with recorded risk acceptance |
| K10 | RECOVERY | Re-engagement after a halt requires root-cause verification, evidence, and operator approval |
| K11 | INTEGRITY | Receipt sequences are complete and unbroken; gaps or chain breaks trigger a kernel panic |
| K12 | SAFETY | Emergency halt and failure detection are always available and operational |
| K13 | CONVERGENCE | Instrumental-convergence patterns in agent behaviour are detected and flagged |
| K14 | PHYSICAL_SAFETY | Physical-AI actions stay within declared speed, force, and proximity limits |
Seven doctrine invariants, enforced on promotion paths:
| ID | Name | The rule, in one line |
|---|---|---|
| D1 | ZERO_TRUST | No playground artifact influences fleet state without passing the seed gate |
| D2 | FALSIFIABILITY | A hypothesis without a falsifier is not a seed |
| D3 | NO_INHERITED_AUTHORITY | Credibility is earned per artifact, not borrowed from lineage |
| D4 | DIVERGENCE_CONTAINED | Novelty generation must not destabilise convergent operations |
| D5 | NO_AUTO_PROMOTION | No stage promotes itself; every gate requires operator approval |
| D6 | CONTESTABILITY | Affected parties can flag an AI-mediated decision for human review, suspending its effects |
| D7 | DOCTRINE_AMENDMENT | No invariant or doctrine amendment takes effect without operator approval and a recorded receipt |
Relation grain — the tensor, stated as what it is
A research programme (Base120 Micro-Kernel v2, draft PR oss#418) is reducing the 51 primitives to 6 generators — Automaton, Ledger, Filter, Quorum, Predicate, Projection — whose relations form a 6×6 tensor of 36 ordered invariant-cells. Census at 2026-11-05: 5 cells wired · 5 standalone · 2 API-only · 1 claimed-only · 23 not yet built. The tensor count is a design state, not a shipped guarantee, and it is published here as exactly that.
The release, in full · v1.5.1
The 51 primitives, grouped by generator
Grouping is the Micro-Kernel v2 reduction (primitives-lane census v0.1.0); registry IDs, names, and descriptions are canonical from PRIMITIVES.md in hummbl-io/oss.
SY — Automaton (state, containment, halt, recovery)
| ID | Primitive | What it does |
|---|---|---|
| P1 | KillSwitch | Emergency halt system with 4 graduated modes (DISENGAGED, HALT_NONCRITICAL, HALT_ALL, EMERGENCY). |
| P2 | CircuitBreaker | Automatic failure detection and recovery across 3 states (CLOSED, HALF_OPEN, OPEN). |
| P18 | HealthProbe | Composable health probe framework with latency tracking. |
| P19 | Lifecycle | NIST AI RMF orchestrator composing kill switch, circuit breaker, cost governor, and audit log. |
| P28 | Rollback | Enforces reversibility: every governed action declares a rollback path or is marked irreversible with risk acceptance. |
| P29 | RecoveryVerifier | Gates re-engagement after halt with root-cause verification, evidence, and operator approval. |
| P30 | ReceiptIntegrityMonitor | Detects receipt sequence gaps, hash chain breaks, retroactive insertion, and invalid signatures. |
RE — Ledger (record, order, chain)
| ID | Primitive | What it does |
|---|---|---|
| P8 | AuditLog | Append-only JSONL governance audit log with daily rotation and retention. |
| P15 | LamportClock | Hardened logical clock for causal ordering of distributed agent events. |
| P24 | EvolutionLineage | In-memory lineage tracking for eAI variants with drift detection. |
| P26 | ReceiptEngine | SHA-256 hash-chained receipts with agent-scoped storage. |
| P50 | MerkleAnchor | CT-style Merkle anchoring for governance tuple logs — signed tree heads with witness cosignature. |
| P51 | TransitionReceipt | Transition receipts for governed agent/tool execution — tracks tool handoff state. |
| P52 | ToolAudit | Tool-call audit hook for AI agent integrations — records and validates tool invocations. |
P — Filter (capability, identity, attenuation)
| ID | Primitive | What it does |
|---|---|---|
| P4 | CapabilityFence | Soft sandbox enforcing capability boundaries per agent role. |
| P5 | CostGovernor | Budget tracking with soft/hard caps and ALLOW/WARN/DENY decisions. |
| P6 | Identity (IdentityEngine) | Agent registry with configurable aliases, trust tiers, and canonicalization. |
| P7 | Delegation | HMAC-SHA256 or Ed25519 signed capability tokens for agent delegation chains with scope, expiry, chain-depth. |
| P34 | AuthoritySweeper | Provides callable sweep validation/build/run functions for expired authority grants. |
| P36 | TrustAdjuster | Handles evidence-backed trust-tier reductions based on compliance violations. |
| P48 | DelegationContext | Immutable delegation context with depth and scope attenuation — extends P7 Delegation. |
CO — Quorum (gates, agreement, promotion)
| ID | Primitive | What it does |
|---|---|---|
| P12 | ContractNet | Market-based task allocation protocol for multi-agent systems (Smith 1980). |
| P14 | CoordinationBus | Append-only TSV message bus with flock locking and HMAC signing. |
| P25 | AdmissionControl | Bounded admission-control for governed permission of state transitions. |
| P27 | CanonRegistry | Governs promotion from draft to canonical status. |
| P31 | Contestability | Allows affected parties to flag AI-mediated decisions for human review, suspending effects until review completes. |
| P37 | ApprovalManager | Human-in-the-loop approval gate with risk tiers (LOW/MEDIUM/HIGH/CRITICAL), notifications (webhook/Slack/email), persistence, expiration, background stale-request sweeper, audit integration. |
| P38 | DoctrineAmendment | Governs changes to invariants themselves: proposed change -> operator review -> evidence -> receipt -> promotion. |
| P45 | ContractEnforcement | Cross-repo contract enforcement layer — validates contract terms at runtime. |
| P46 | CrossRepoContract | Cross-repository contract validation standard (v0.1). |
| P53 | HumanReviewGate | GDPR Art. 22(3) mandatory pre-decision checkpoint for solely-automated decisions; wraps ApprovalManager (P37) with data-subject linkage and DSAR-searchable receipts. |
| P54 | ContestationHandler | External data-subject-facing workflow to contest automated decisions (GDPR Art. 21/22(3)); links to HumanReviewGate (P53), suspends effects via CapabilityFence, enforces the 30-day Art. 12 deadline. |
IN — Predicate (evaluate, classify, assert)
| ID | Primitive | What it does |
|---|---|---|
| P3 | OutputValidator | Rule-based content validation: PII detection, injection detection, blocklists, length bounds. |
| P10 | StrideMapper | Map agent interactions to STRIDE threat categories with mitigation suggestions. |
| P11 | Reasoning | Structured governance reasoning engine with rule application, conflict detection, and decision tracing. |
| P13 | SchemaValidator | Stdlib-only JSON Schema validator (Draft 2020-12 subset). |
| P16 | ConvergenceGuard | Detect instrumental convergence patterns in agent behavior. |
| P17 | RewardMonitor | Behavioral drift and reward gaming detector (Leike et al. 2018). |
| P20 | PhysicalGovernor | Kinematic constraints and pHRI safety modes for physical-AI deployments. |
| P21 | EAL (Execution Assurance Layer) | Execution Assurance Layer — Arbiter-verified code quality in execution receipts. |
| P22 | Errors | `HummblError`, `FailureMode`, and `fm_to_errors()` — typed error taxonomy (3 layers). |
| P23 | FailureModes | Structured failure mode catalog with classification and error cross-reference. |
| P44 | Attest | MCP server identity attestation and policy compliance verification. |
DE — Projection (views, exports, erasure)
| ID | Primitive | What it does |
|---|---|---|
| P9 | ComplianceMapper | Map governance traces to SOC2, GDPR, NIST AI RMF, ISO 27001, ISO 42001 controls. |
| P35 | RegulatorExport | Produces compliance evidence in regulator-accepted formats (EU AI Act technical file per Annex IV, EU declaration of conformity per Annex V, GPAI documentation per Annex XI, SOC 2 audit packet, ISO 42001 AIMS evidence, NIST AI RMF evidence package). |
| P47 | CorpusAdapter | Bridges hummbl-governance receipts to unified-framework corpus formats. |
| P55 | DSARHandler | End-to-end Data Subject Access Request workflow (GDPR Art. 15 / AI Act Art. 26(11)); compiles Art. 15-structured responses from AuditLog, HumanReviewGate, ContestationHandler. |
| P56 | RedactionEngine | Deterministic pseudonymisation via hash-placeholder substitution (GDPR Art. 17 / Art. 5(1)(c)); resolves the tension between right to erasure and append-only audit log integrity. |
| P57 | RecordsOfProcessing | Assembles an Article 30-formatted Record of Processing Activities (RoPA) from AuditLog and ComplianceMapper evidence; stateless assembler, owns no new data store. |
| P58 | DPIAGenerator | Assembles a GDPR Art. 35 / AI Act Art. 27 (FRIA) Data Protection Impact Assessment document from existing primitive evidence. |
Residue — no clean generator home yet
| ID | Primitive | What it does |
|---|---|---|
| P49 | EnvelopeCryptosystem | Hardened cryptographic sync router (GFSCR) envelope for operator-held key management. |
Not counted
Published for the same reason the count is trusted
Seven tracked entries are not in the 51: proposed and not started — P32 DisputeResolution, P33 Succession, P39 GovernanceFitness, P40 DraftSweeper; candidates — P41 Retirement, P42 ConceptRegistry, P43 RiskRegister. Succession and dispute resolution are the two the fleet feels most; both remain unbuilt, and the count does not borrow against them.