CK.Consensus -- Ontological Governance

The Core Principle

A developer does not edit a concept kernel's tool directly. They talk to the kernel via Claude Code (/ck), propose changes through consensus, and the kernel's ontology governs what changes are valid. The tool is a consequence of the ontology, not the other way around.

This inverts the typical development flow:

Traditional	CKP Consensus
Edit code, then update docs	Propose change to ontology, then generate code
Code review catches design issues	Ontology review catches design issues BEFORE code exists
CI validates after commit	Consensus validates before commit
Documentation drifts from code	Ontology IS the code's schema -- drift is a compliance failure

The Consensus Loop

Developer (Claude Code)
    |
    v
  /ck {Kernel}                    -- Talk to the CK loop
    |
    v
  CK.Consensus propose            -- Submit change proposal
    |
    v
  CK.Consensus evaluate           -- Evaluate against ontology + SHACL + fleet
    |
    v
  CK.Consensus approve            -- Generate tasks for TOOL changes
    |
    v
  Headless Claude Code             -- Execute tasks against tool/processor.py
    |
    v
  Tool behaves per ontology        -- Produces typed instances
    |
    v
  Instances accumulate             -- Knowledge grows in DATA loop
    |
    v
  Developer observes, proposes     -- Back to top

CK.Consensus Kernel

CK.Consensus is a system kernel (node:hot, STRICT governance) with five actions:

Action	Access	Description
propose	auth	Submit a change proposal for a target kernel
evaluate	auth	Evaluate a proposal against ontology, SHACL, and fleet topology
approve	auth	Approve an evaluated proposal, generate tasks
decisions	anon	List all consensus decisions with provenance
review	auth	Review a pending proposal (uses CK.Claude strict-auditor)

Edges

CK.Consensus declares three outbound edges:

Target	Predicate	Rationale
CK.ComplianceCheck	TRIGGERS	Validate proposals against CKP compliance (20 check types)
CK.Operator	TRIGGERS	Reconcile changes after tasks complete
CK.Claude	EXTENDS	AI review of proposals via strict-auditor persona

The EXTENDS edge to CK.Claude means CK.Consensus gains a review action that uses Claude with the strict-auditor persona -- zero speculation, evidence-only verdicts. The TRIGGERS edges create the governance pipeline: proposals are validated by ComplianceCheck and, once approved and executed, reconciled by the Operator.

Propose

Any change to a kernel's CK loop goes through a formal proposal. The proposal declares what to change, where, and why:

proposal_id: "P-a1b2c3d4"
proposer: "ckp://Actor#developer.peter"
target_kernel: "Delvinator.Core"
change_type: "action_add"
specification:
  action_name: "quality.score"
  action_type: "inspect"
  access: "auth"
  description: "Compute quality score for an entity"
rationale: "Exchange analysis reveals recurring quality assessment patterns
  that should be formalised as an action"

The rationale field is important -- it records the why behind the change. This becomes part of the provenance chain, enabling future auditors to understand the reasoning behind every evolution.

Change Types

Type	What Changes	CK Loop Files Affected
action_add	New action on the kernel	SKILL.md, conceptkernel.yaml (spec.actions)
action_modify	Change existing action	SKILL.md, conceptkernel.yaml
edge_add	New edge to another kernel	conceptkernel.yaml (spec.edges)
ontology_update	New or modified classes	ontology.yaml
skill_update	Behavioral instruction changes	SKILL.md
claude_md_update	Agent context changes	CLAUDE.md
shacl_update	New validation constraints	rules.shacl

Evaluate

The evaluation phase subjects the proposal to four validation layers. All four must pass for the proposal to be approved:

1. Ontological Validity

Does the proposed change conform to the target kernel's ontology.yaml? An action that references undefined classes is rejected. If the proposal adds a quality.score action that produces QualityScore instances, does ontology.yaml define QualityScore? If not, the ontology must be updated as part of the proposal.

2. Constraint Compliance

Does the change satisfy rules.shacl? A new edge to a non-existent kernel violates structural constraints. If rules.shacl requires all instances to have prov:wasGeneratedBy, does the proposal include provenance generation?

3. Fleet Topology

Does the change create orphan references, circular dependencies, or broken edge targets? If the proposal adds an edge to CK.Auth, does CK.Auth exist in the fleet? Topology validation prevents changes that would make the fleet inconsistent.

4. CKP Compliance (20 Check Types)

The proposal is validated against the full compliance suite via the TRIGGERS edge to CK.ComplianceCheck. This catches issues that the first three layers miss:

• Missing awakening files
• Invalid NATS topic declarations
• Version conflicts
• Governance mode violations
• Provenance field completeness

Verdict

Verdict	Meaning	Next Step
pass	Proposal is structurally valid, constraint-compliant, and topologically consistent	Proceed to approve
fail	Proposal violates ontology, SHACL, or fleet topology	Reject with evidence
needs_revision	Proposal is partially valid but requires amendments	Return to proposer

Approve

Approved proposals generate tasks. Each task is a concrete work item with an instruction, a target file, constraints, and an executor:

decision_id: "D-e5f6g7h8"
proposal_id: "P-a1b2c3d4"
verdict: "pass"
tasks_generated:
  - task_id: "T-i9j0k1l2"
    target_file: "concepts/Delvinator.Core/tool/processor.py"
    instruction: "Add @on('quality.score') handler that computes quality
      score per ontology.yaml QualityScore schema"
    constraints:
      - "Output MUST validate against ontology.yaml QualityScore class"
      - "MUST produce prov:wasGeneratedBy linking to this task"
    executor: "headless-claude-code"
    version_pin: "abc123"         # git commit hash
prov:wasGeneratedBy: "ckp://Action#CK.Consensus/evaluate-{ts}"
prov:wasAttributedTo: "ckp://Actor#developer.peter"
prov:generatedAtTime: "2026-04-05T10:00:00Z"

Tasks flow to the Task Execution Engine for automated execution via headless Claude Code. See the task engine documentation for the full lifecycle (pending, executing, completed, failed), task operations (split, evolve, merge), and version pinning details.

Provenance: Every Decision Is a prov:Activity

Every consensus decision produces a provenance record:

prov:Activity:
  id: "ckp://Instance#CK.Consensus/decision-2026-04-06T20:00:00Z"
  prov:wasAttributedTo: "user:peter@conceptkernel.org"
  prov:generatedAtTime: "2026-04-06T20:00:00Z"
  prov:wasGeneratedBy: "ckp://Action#CK.Consensus/evaluate"
  prov:used:
    - "ckp://Kernel#Delvinator.Core:v1.0/ontology.yaml"
    - "ckp://Kernel#CK.ComplianceCheck:v1.0/check.all"
  decision:
    verdict: pass
    target_kernel: Delvinator.Core
    change_type: action_add
    proposal_hash: "sha256:..."

The provenance chain is complete and auditable:

• Who proposed the change (prov:wasAttributedTo)
• When the decision was made (prov:generatedAtTime)
• What was evaluated (prov:used -- links to ontology and compliance results)
• How the decision was reached (prov:wasGeneratedBy -- the evaluate action)

Tasks generated from the decision carry prov:wasGeneratedBy linking back to the consensus record. Instances produced by task execution carry prov:wasGeneratedBy linking to the task. The chain is unbroken: proposal -> decision -> task -> instance.

The Full Picture: Ontology Is Law

After consensus-driven tasks are executed, the tool behaves according to the ontology:

Aspect	Governed By	Enforcement
Produces	ontology.yaml	Instance shapes, required fields, types
Notifies	conceptkernel.yaml	NATS topics, event types
Validates	rules.shacl	Structural constraints on instances
Composes	edges	COMPOSES, TRIGGERS, PRODUCES, CONSUMES, EXTENDS predicates

The ontology is not documentation. It is the schema. The tool is its executor. Consensus is the gate between intent and implementation.

Governance Mode Interaction

Consensus behavior depends on the target kernel's governance mode:

Kernel Governance	Consensus Requirement
STRICT	All CK loop changes MUST go through consensus
RELAXED	Structural changes (actions, edges, ontology) MUST go through consensus; documentation changes (SKILL.md, CLAUDE.md) MAY be direct
AUTONOMOUS	MAY bypass consensus; SHOULD still produce PROV-O records

This means CK.ComplianceCheck (STRICT governance) requires consensus for every change, while a development kernel with RELAXED governance can iterate faster on documentation without governance overhead. Even AUTONOMOUS kernels should record provenance -- the audit trail matters even when governance is relaxed.

The governance mode is declared in conceptkernel.yaml -- the developer sets the rules, consensus enforces them. See CK Loop Evolution for the full workflow including change types and governance levels.

Architectural Consistency Check

Logical Analysis: Consensus Design

Question: Is consensus overkill for small changes?

Answer: The consensus loop is as lightweight as the change. A small change (add one action) produces a small proposal, a quick evaluation, and one task. The overhead is the provenance record -- about 20 lines of YAML. The benefit is auditability: every change to every kernel has a traceable decision. For RELAXED kernels, consensus is optional -- the overhead only applies when governance demands it.

Question: What prevents the consensus kernel itself from being modified without consensus?

Answer: CK.Consensus has STRICT governance, which means changes to CK.Consensus must go through... CK.Consensus. This is deliberately self-referential. The bootstrap case (initial CK.Consensus deployment) is handled by the operator, which creates the kernel from a template. After that, all changes are self-governed.

Question: How does headless Claude Code execution work in practice?

Answer: Currently, tasks are executed via claude -p (batch mode) or claude_agent_sdk (streaming). The task instruction is the prompt. The constraints are injected as system-level rules. Claude makes the code changes, and the output is validated against the kernel's ontology before sealing. This is the v3.5.15 (Task Execution Engine) feature -- currently at the design stage.

Gap identified: The evaluate step checks against the target kernel's current ontology. But if the proposal includes an ontology_update, the new ontology does not exist yet at evaluation time. The evaluator must speculatively validate against the proposed ontology -- a form of dry-run that is not yet implemented. Currently, ontology_update proposals are evaluated only for structural validity (valid YAML, valid LinkML), not for semantic consistency with the fleet.

Gap identified: CK.Consensus is specified but the task execution pipeline (v3.5.15) is not yet implemented. Proposals can be evaluated and approved, but tasks currently accumulate rather than executing automatically. This is the next major planned feature.

Conformance Requirements

• Changes to CK loop MUST go through consensus (STRICT governance)
• AUTONOMOUS kernels MAY bypass consensus
• Consensus MUST produce a prov:Activity record for every decision
• Tasks produced by consensus MUST reference the consensus decision via prov:wasGeneratedBy
• Task execution MUST validate output against the kernel's ontology
• The tool MUST behave according to the current ontology -- not a cached or previous version
• Edges between kernels MUST be declared in conceptkernel.yaml -- implicit dependencies are non-conformant
• Each kernel is a sovereign BFO:0000040 Material Entity -- no kernel can modify another's CK loop