Malleus

A root ontology in LinkML, and the opinion that words have power.

Why this exists

I believe words have power. The closer we work with them, the more carefully we pin down what they mean and how they relate, the closer we get to something a machine can use without guessing. An ontology is that pinning-down, made explicit and machine-readable. Borges and Le Guin understood this long before software did: to name something precisely is to begin controlling it.

The practical bet: if you define your domain once, in an ontology, you can propagate that definition through every layer of a system. LinkML already compiles a schema down to JSON Schema, Pydantic, SQL DDL, OWL, SHACL, TypeScript, whatever you need. So the same five concepts, with the same constraints, can shape the frontend form, the backend validator, the ML training contract, Shelob's tool schema, the knowledge graph's node types, and the Prolog rules that reason over them. One source. All layers speaking the same vocabulary.

When that actually happens across a codebase, something unexpectedly useful shows up. Components stop drifting apart. The frontend and backend stop disagreeing about what a "Drug" is. A new contributor learns one vocabulary instead of five. Whole classes of bugs (the ones caused by definitions sliding between modules) just stop existing. Adding a new concept becomes one change in one file, flowing outward through whatever code generators you've wired up.

That's malleus: a small, stable root vocabulary, plus the mechanics to keep everything built on top of it honest.

Current package boundary: 0.7.0, stage-7c-policy-selected-authorization-control. See docs/IMPLEMENTATION_STATUS.md for implemented and explicitly pending capabilities. Code can inspect the same boundary through malleus.IMPLEMENTATION_STATUS.

The core primitives

Everything in malleus is one of five things:

• Entity: something that persists through time. A drug, a server, a person, a concept.
• Event: something that happens. A click, a deployment, an interaction detected.
• Signal: a derived quality computed from patterns. A risk score, a health status, a trend.
• Agent: a mixin capturing the capability to act or decide. Not a class, a trait.
• Relation: a typed, directed, reified edge between entities.

Plus four cross-cutting mixins so every typed thing can carry basics without reinventing them: Identifiable (id, name), Temporal (created_at, updated_at), Describable (description, tags), Statusable (ACTIVE, INACTIVE, DESTROYED).

Domains extend this root. CYP450 drug interactions, MITRE ATT&CK threat models, both come with examples in this repo. Writing your own is a YAML file.

Install

pip install malleus-dev

The Python package contains the logic compiler and verifier. Executing logic checks also requires a swipl executable on PATH; absence fails explicitly at check time.

Quick start

from malleus import (
    KnowledgeGraph,
    OntologyRegistry,
    ProposedOperation,
    bundled_ontology_path,
    stage_subgraph,
)

reg = OntologyRegistry(bundled_ontology_path("domains", "cyp450.yaml"))
kg = KnowledgeGraph(reg)

kg.create_entity("Enzyme", "enz-cyp3a4", {"name": "CYP3A4", "cyp_isoform": "CYP3A4"})
kg.create_entity("Drug", "drug-sim", {"name": "Simvastatin"})

candidate = stage_subgraph(kg, [
    ProposedOperation.relation(
        "SubstrateOfRelation", "rel-001", "drug-sim", "enz-cyp3a4",
        {"relation_type": "SUBSTRATE_OF"},
    )
])
assert candidate.valid
assert kg.edge_count == 0             # staging never mutates the base graph
print(candidate.candidate_digest)     # binds ontology, base state, and ordered writes
candidate.materialize_into(kg)        # explicit structural materialization

# Write-time validation. No structurally invalid write materializes.
op = kg.create_entity("NotAType", "x", {})
assert op.op_status.value == "REJECTED"
print(op.rejection_reason)   # "Unknown entity type: 'NotAType'"

The OntologyRegistry is the constructor parameter for the KnowledgeGraph. No registry, no KG. That's the rule, and it's the whole point: the graph can only ever hold things the ontology says exist.

STAGED means an operation passed validation inside an isolated candidate. COMMITTED means it was structurally materialized. Neither means the record is true, epistemically accepted, or authorized for action.

Distributed convergence

Every OntologyRegistry has a deterministic content hash and a fingerprint of atomic facts. Two peers running the same schema produce the same hash, no coordination needed. Two peers running different versions can verify compatibility without exchanging full schemas.

reg = OntologyRegistry(bundled_ontology_path("domains", "cyp450.yaml"))
print(reg.content_hash())        # 64-char SHA-256, deterministic
print(len(reg.fingerprint()))    # frozenset of atomic facts

result = reg.check_compatibility(foreign_hash, foreign_fingerprint)
# "identical" | "superset" | "subset" | "divergent"

Under additive-only evolution (add types, enum values, or slots; relax required to optional), a newer ontology's fingerprint is always a strict superset of an older one's. Peers can tag every write with the hash they used, and receivers can decide: accept (we're compatible), quarantine (we'll understand this after we upgrade), or reject (we've forked, this is a bug).

This matters in fleets running rolling updates. Without it, CRDT sync during the upgrade window can silently drop properties the older node doesn't recognize. With it, the older node says "I can't validate this yet, hold it" and nothing is lost.

One caveat, worth saying plainly. Relaxing a slot from required to optional is additive on the producer side (you're loosening a guarantee) but subtractive on the consumer side (code that hardcoded the field's presence will crash when a new producer omits it). The default check_compatibility() answers the producer question: can data flow safely between us? For the consumer question, use strict_fingerprint() and check_compatibility_strict(), which include required-constraint facts. A relaxation shows up there as divergence, surfacing the risk that would otherwise stay hidden.

Domain extensions

Two examples ship with the library. Write your own the same way:

# your_domain.yaml
id: https://example.org/schema/your_domain
name: your_domain
imports:
  - malleus
  - linkml:types

classes:
  YourEntity:
    is_a: Entity
    slot_usage:
      your_slot:
        required: true
        range: YourEnum

  YourRelation:
    is_a: Relation
    slot_usage:
      relation_type:
        range: YourRelationType
        required: true
        equals_string: CONNECTS
      source_id:
        range: YourEntity
      target_id:
        range: YourEntity

enums:
  YourEnum:
    permissible_values:
      VALUE_A: {}
      VALUE_B: {}

  YourRelationType:
    permissible_values:
      CONNECTS: {}

slots:
  your_slot:
    range: YourEnum

Relations use concrete classes with explicit source and target ranges. Malleus rejects unknown properties, missing required fields, malformed values, duplicate identifiers, mismatched predicates, and invalid endpoint types before graph mutation.

Pinned Prolog verification

GraphFactCompiler converts any Malleus graph into a fixed typed fact vocabulary. A LogicContract pins the ontology hash, exact trusted rule bytes, declared rule IDs, versions, and subprocess wall-clock timeout. PrologVerifier evaluates caller-supplied context plus an isolated candidate in a fresh SWI-Prolog process. Stage 5 does not claim that the context is protocol-accepted state.

from malleus import LogicContract, PrologVerifier, ProposedOperation, stage_subgraph

contract = LogicContract.load("your_logic_contract.yaml")
verifier = PrologVerifier(contract)
candidate = stage_subgraph(kg, [
    ProposedOperation.relation(
        "InhibitsRelation", "rel-002", "drug-sim", "enz-cyp3a4",
        {"relation_type": "INHIBITS", "inhibition_strength": "STRONG"},
    )
])
result = verifier.verify_candidate_subgraph(candidate)
if not result.valid:
    for violation in result.violations:
        print(violation.rule_id, violation.violation_code, violation.witness_record_ids)
else:
    # Structural materialization only. This is not epistemic acceptance.
    candidate.materialize_into(kg)

The rule program exposes only two required predicates:

malleus_rule(RuleId).
malleus_violation(RuleId, ViolationCode, WitnessRecordIds).

The verifier enumerates every violation, rejects malformed or unknown witnesses, and never mutates the base graph. Consult errors, timeouts, manifest mismatches, and malformed results raise LogicExecutionError; they never become SATISFIED. logic_monitor_failure_records() converts such a failure into an atomic MonitorFailure and UnavailableAssessment pair bound to the logical contract and ruleset. Completed checks can be serialized as content-addressed LogicCheckRecord and ViolationWitness records.

The package ships the CYP450 contract and rules as an example. Stage 5 accepts only trusted, pinned local rule programs. The timeout bounds the Prolog subprocess wall clock, not graph compilation, output size, memory, or CPU. It does not sandbox untrusted Prolog or issue formal proof certificates.

Policy-selected monitoring and control

Stage 6 replaces opaque monitor and epistemic-policy artifacts with typed, content-addressed records. A monitor specification binds its assessment kind, implementation hash, and input artifacts. An epistemic policy names the exact monitors it requires and maps each VIOLATED or UNKNOWN result to an epistemic control. Each proposal binds one exact policy before monitoring begins, so a controller cannot choose a favorable policy after seeing outputs.

evaluate_epistemic_policy() requires exactly one assessment from every selected monitor. It returns the ordered assessment IDs, control-triggering assessment IDs, selected verdict, and canonical evaluation hash. Protocol replay recomputes those values before accepting an EpistemicDecision.

The control rules are intentionally small:

• All required assessments SATISFIED selects ACCEPT.
• VIOLATED selects the monitor-specific REJECT, DEFER, or CONTEST mapping.
• UNKNOWN selects only DEFER or CONTEST.
• Explicit policy precedence resolves multiple triggered controls.
• Omitted, duplicate, or unrequired monitor outputs block the decision.
• An exact monitor can produce only one output per proposal. A logical monitor can also record only one completed check for that proposal.

A monitor that did not complete is not silently omitted. The caller records MonitorFailure plus UnavailableAssessment atomically, using monitor_failure_records() for non-logical monitors or logic_monitor_failure_records() for logical execution. Stage 6 validates these outputs and controls; it does not execute every domain-specific monitor or claim to reproduce its result.

Recording assessments without appending an epistemic decision leaves the proposal open. This separates monitoring-only C3 from monitoring-plus-control C4 without maintaining two code paths.

Core assessment kinds use their declared concrete record types. Domain-defined assessment subclasses cannot claim a core kind while omitting that kind's required evidence. Version 0.4.0 therefore does not replay 0.3.0 proposals unchanged: each proposal must explicitly name and source its policy record. This precommitment prevents ex-post selection. It does not prove that the proposer had authority to choose that policy or that the policy applies to the proposal's domain; those checks remain outside Stage 6.

Accepted graph and bitemporal replay

Stage 7b makes the proposed graph mutation replayable and binds it to assent. A GraphBaseArtifact commits an externally supplied base graph. A CandidateSubgraphArtifact stores exact ordered writes, an explicit valid-time interval for every write, supersession links, ontology hash, acceptance and materialization heads, and pre-state and post-state digests. ProposedSubgraph and EpistemicDecision both bind that candidate by ID, record hash, and candidate digest.

A candidate-bound ACCEPT requires exactly one AcceptedGraphApplication in the same decision event. REJECT, DEFER, and CONTEST require no application. Replay restages the writes and recomputes every binding before it updates the derived graph. Direct use of CandidateSubgraph.materialize_into() remains a structural operation and cannot change the ledger's accepted projection.

from malleus import AcceptedGraphProjector

projector = AcceptedGraphProjector(protocol_ledger)
current = projector.current(valid_as_of="2026-08-12T08:00:00+00:00")
historical = projector.as_of(
    transaction_as_of="2026-08-12T09:00:00+00:00",
    valid_as_of="2026-01-01T00:00:00+00:00",
)

Valid intervals are half-open. Valid time is always explicit and is never inferred from transaction time. A later retroactive supersession affects only transaction views that include the later event. The JSONL ledger is the authority; NetworkX is rebuilt as a defensive projection. Accepted projections omit the local KnowledgeGraph.operations audit because those operation timestamps are execution-local and are not ledger commitments.

This is an accepted knowledge commitment, not a truth guarantee or action authorization. The caller must supply the exact graph committed by the graph base artifact. Remote graph-base resolution, typed retraction, and multi-writer serialization remain outside version 0.7.0.

Architecture

For the layer-by-layer walkthrough (vocabulary, typed graph, ground truth loading, logic engine, distributed convergence), see docs/ARCHITECTURE.md.

Adoption guides:

• docs/PRINCIPLES.md: what malleus claims and what it does not, the six principles the rites defend, and the future work that is reserved rather than asserted
• docs/ADOPTION_GUIDE.md: start here. How to adopt malleus from any project and keep it alive, written for a human and their coding assistant together
• docs/IMPLEMENTATION_STATUS.md: the current machine-checked capability boundary
• docs/ONTOLOGY_PROTOCOL.md: how to add malleus to an existing project
• docs/KNOWLEDGE_GRAPH_PROTOCOL.md: how the ontology shapes the KG
• docs/ASSENT_PROTOCOL.md: how proposals, assessments, decisions, authorization, and replay remain separate
• docs/DELIMITATIONS.md: what malleus reuses, rejects, and can honestly claim against OWL, SHACL, TypeDB, XTDB, nanopublications, and the rest of the field
• docs/RECIPES.md: six recipes for capturing and using domain knowledge in a KG, each grounded in working code from a surveyed fleet of adopting projects

The value is prevention

Be clear-eyed about what malleus buys you, because it is easy to underestimate. The value is prevention: whole classes of bugs (definitions sliding between modules, invalid records entering the store, a rule silently citing an axiom that no longer exists) stop being possible. Prevention is invisible by nature; you never see the bug that could not happen, so the investment is hard to quantify from inside a healthy project. It becomes visible in exactly two places: in projects that adopted the vocabulary but not the enforcement and paid a measured cost for the gap, and in rebuilding a stuck project with these recipes and watching the difference. The recipes and delimitations documents exist to make that argument with evidence instead of conviction.

The Ordo Malleus

Discipline decays without an auditor, so malleus ships its own inquisition. (An ontology named after a hammer was always going to attract inquisitors; we let it, within reason.)

Three tiers:

• malleus-inquisitor <schema.yaml>: the mechanical rites, a CLI that any machine can judge. Does the schema construct, is the imported root current against the installed malleus (staleness is detected via check_compatibility_strict, the consumer-side check: the producer-side check_compatibility is blind to a dropped required constraint, which is the most silent drift there is), are the type-slots constrained, are relation endpoints narrowed, are Signals genuinely derived, are formula-shaped slots backed by an executor. Exit 0 grants the purity seal, 1 records heresies, 2 means the instrument itself is broken and nothing was judged. Severities are data: copy rubric.yaml, tune it, and pass --rubric PATH. Every run prints the rubric it used and how many rites were disabled, because a seal is only as wide as the rubric that granted it.
• The malleus-inquisitor skill (.claude/skills/malleus-inquisitor/): the judgment rites a coding assistant applies to a whole repo: write-path enforcement, reader census, citation integrity, provenance quality, fail-closed rules. It writes a ranked MALLEUS_INQUISITION.md into the inspected project.
• The rubric (src/malleus/inquisition/rubric.yaml): the single source both tiers read. Every rite records the generic field lesson that paid for it, no project named. It is data on purpose: tune it, extend it, and send generic lessons back as issues or PRs. That is how the Ordo learns.

And for every project that uses malleus, there is the acolyte: malleus-inquisitor install-skills --user (or --project .) copies the shipped agent skills into your .claude/skills, so any coding-assistant session gets the malleus-acolyte companion: the adoption playbook as standing orders, self-inquisitions, and, unlike the central inquisitor, license to fix its own project's findings. Generic lessons flow upstream as issues and PRs; releases carry the grown rubric and skills back down. Re-run the installer after upgrading.

Tests

pip install -e .[dev]
pytest tests/ -v

License

Apache-2.0. See LICENSE.

A note on the name

Malleus is the Latin for "hammer". The tool that shapes. Use it to shape your own domains.