Multi-Agent Epistemic Control Plane v0.1 — local beliefs, shared state, dissent, and coalitions

[hummbl-dev]

Parents

• hummbl-dev/hummbl-dev#151 — Multi-Actor World Models v0.1
• hummbl-dev/hummbl-dev#149 — User-Driven World Model Generation v0.1

Purpose

Define control-plane patterns for coordinating multiple agents without forcing their beliefs into one global consensus graph.

The control plane should maintain:

shared evidence and operational ledger
+ agent-local belief states
+ authority and policy state
+ unresolved dissent
+ coalition and lifecycle state

Required separations

• observation versus inference;
• local belief versus shared operational fact;
• ratified state versus action-ready state;
• evidence agreement versus conclusion agreement;
• consensus versus independent corroboration;
• merge proposal versus accepted merge;
• coalition membership versus authority;
• coordination state versus private user model state.

Candidate control-plane objects

SharedWorldState
LocalBeliefPointer
ClaimLineage
IndependenceLineage
DissentRecord
ChallengeRecord
MergeProposal
ForkRecord
ConsensusPosture
DecisionGate
CoalitionSpec
CoalitionMembership
CoalitionAuthority
CoalitionResourceBudget
CoalitionLifecycle
QuarantineRecord
ReconciliationRun
StateSnapshot

Consensus posture

Represent at least:

unassessed
single_lineage_support
correlated_support
independent_support
contested
context_split
blocked_on_evidence
accepted_for_bounded_action
superseded
falsified

Do not use agent majority as the default truth rule.

Independence analysis

Track and deduplicate support based on:

• model/provider/version;
• system prompt/policy lineage;
• source lineage;
• retrieval path;
• tool chain;
• execution environment;
• evaluator/judge lineage;
• common upstream agent or handoff.

A set of correlated agents should be classified as correlated support, not independent replication.

Dissent and merge/fork semantics

The system must support:

• persistent dissent without forced resolution;
• contextual variants where conflicting claims are locally valid;
• model forks that preserve common ancestry;
• merge proposals requiring declared resolution criteria;
• minority reports attached to accepted bounded actions;
• re-opening a decision when new evidence arrives;
• no silent deletion of rejected alternatives.

Coalition semantics

A coalition must declare:

• purpose and bounded task;
• root principal;
• members and versions;
• delegated authority;
• shared resources and cost ceiling;
• evidence and voting/decision rules;
• dissent preservation;
• creation and expiry times;
• delegation/subagent limits;
• dissolution, quarantine, and receipt requirements.

Coalitions must not gain authority merely by accumulating members.

Required artifacts

• v0.1 control-plane design note;
• schemas for shared state, dissent, merge/fork, independence lineage, and coalition spec;
• valid/invalid/adversarial fixtures;
• reconciliation pseudocode or minimal deterministic reference implementation;
• reason-code taxonomy;
• crosswalk to agent-as-code, agent-handoffs, hummbl-tuples, and runtime admission;
• receipt linked to parent issues.

Adversarial fixtures

• five agents using the same provider/source counted as five votes;
• coalition self-expands authority through membership;
• dissent is removed after a majority decision;
• agent-local hypothesis enters shared fact state without admission;
• stale shared-state snapshot overwrites newer correction;
• compromised agent poisons shared state without provenance;
• fork loses ancestry or ownership;
• model merge hides unresolved contradiction;
• coalition persists after expiry;
• private user state leaks into organization-wide shared state.

Acceptance criteria

• Local beliefs and shared state are separately represented.
• Consensus posture distinguishes correlated from independent support.
• Dissent, context splits, forks, and merge proposals are first-class.
• Shared-state writes require authority and admission.
• Coalitions have bounded purpose, authority, resources, expiry, and dissolution.
• Majority does not create truth or authority by default.
• Reconciliation is deterministic where rules permit and indeterminate otherwise.
• Adversarial fixtures cover poisoning, capture, staleness, and privacy leakage.
• Receipt links to hummbl-dev/hummbl-dev#151.

Non-goals

• No distributed-consensus protocol claim.
• No blockchain requirement.
• No universal ontology.
• No production private-user-state store.
• No separate agent epistemic network repository.

[hummbl-dev]

Portfolio issue triage — reviewed open state at 2026-07-13T07:56:40Z.

Classification:

• Type: security; suggested priority: P2
• Labels: none
• Assignee: hummbl-dev
• Last activity: 07/10/2026 21:29:47 (2 days ago); milestone: none

Findings:

• P2: no authoritative evidence or source link is present.
• P3: issue is unlabeled; routing and priority are ambiguous.

Suggested labels: security, p2. Next action: Document threat, affected surface, reproduction, mitigation, and security regression tests.

Gate: NEEDS_REFINEMENT. This is portfolio triage, not implementation authorization or closure approval.

[hummbl-dev]

---

Automated triage run (2026-07-24)

• Category: research
• Priority: P2
• Severity: medium
• Staleness: fresh
• Recommended next action: Define control-plane objects for shared evidence, local beliefs, and dissent; implement independence analysis to prevent correlated agents from appearing as independent corroboration.

Notes: Multi-Agent Epistemic Control Plane v0.1 — maintains shared evidence, agent-local belief states, authority state, and unresolved dissent. Do not use agent majority as the default truth rule.