Agent Epistemic Profile v0.1 — identity, roles, permissions, beliefs, and calibration

[hummbl-dev]

Parents

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

Purpose

Define an auditable, versioned agent profile for agents that participate in user-owned, organizational, or bounded agent-only world-model systems.

An agent profile must distinguish what the agent can do, may know, may propose, may write durably, and may execute. It must also identify model/provider/version lineage so correlated agents are not mistaken for independent evidence.

Candidate core objects

AgentIdentity
AgentVersion
PrincipalBinding
RoleAssignment
Capability
EpistemicPermission
DataAccessPermission
DurableWritePermission
ActionPermission
DelegationLimit
LocalBeliefContract
SharedStateAccess
CalibrationProfile
LifecycleState
Attestation

Required distinctions

capability != permission
permission != authority
epistemic permission != durable-write permission
durable-write permission != action authority
agent role != principal identity
agent version != stable agent identity

Required profile fields

• agent ID and version;
• implementation/model/provider lineage;
• prompt/policy/configuration hash or reference;
• root principal and represented entity;
• assigned roles;
• capabilities and unavailable capabilities;
• data/source/tool permissions;
• epistemic permissions:
  • observe;
  • infer;
  • propose;
  • challenge;
  • arbitrate;
  • ratify, if ever allowed;
• durable-state permissions;
• action permissions;
• delegation/subagent creation limits;
• scope, expiry, and revocation;
• local-belief retention policy;
• shared-state publication policy;
• required receipts;
• calibration references;
• lifecycle and quarantine state.

Epistemic role profiles

Create bounded example profiles for:

• observer;
• compiler;
• translator;
• critic;
• researcher;
• simulator;
• planner;
• executor;
• historian;
• consent steward;
• arbiter.

No example should automatically combine observation, interpretation, approval, execution, and receipt authority.

Independence lineage

The profile must expose lineage relevant to evidence independence:

• foundation model/provider;
• model/version/weights where known;
• system prompt/policy lineage;
• source and retrieval adapters;
• tool chain;
• execution environment;
• evaluator/judge lineage.

The profile should enable downstream systems to recognize correlated agents and avoid vote/evidence inflation.

Calibration profile

Support domain/task-specific references such as:

agent_id
agent_version
domain
task_type
prediction_count
evaluated_count
calibration_posture
error_profile
abstention_quality
correction_behavior
receipt_completeness
authority_violations
last_evaluated_at
staleness_posture

Do not define a universal scalar reputation score.

Required artifacts

• v0.1 boundary/design note;
• Draft 2020-12 schema;
• valid/invalid/adversarial fixtures;
• deterministic validator;
• example role profiles;
• crosswalk to delegation/identity primitives in hummbl-governance and the actor-neutral runtime admission packet;
• receipt linked to parent issues.

Adversarial fixtures

• agent claims action authority from epistemic permission;
• agent changes provider/version without version event;
• two agents falsely claim independence despite shared lineage;
• agent delegates beyond permitted depth;
• executor also self-ratifies and self-receipts a high-impact action;
• expired or revoked role remains active;
• agent profile omits principal binding;
• calibration is generalized from one domain to all tasks.

Acceptance criteria

• Agent identity and version are distinct.
• Principal, role, capability, permission, and authority are separate.
• Epistemic, durable-write, and action permissions are separately represented.
• Independence lineage is machine-readable.
• Calibration is domain/task specific.
• Delegation depth and subagent creation are bounded.
• Lifecycle, expiry, revocation, and quarantine are represented.
• Examples avoid role concentration.
• Adversarial fixtures fail deterministically.
• Receipt links to hummbl-io/hummbl-io#151.

Non-goals

• No agent consciousness or moral-status claims.
• No universal agent reputation score.
• No production user-model database.
• No autonomous authority derived from technical capability.
• No new agent-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 21:29:10 (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: governance
• Priority: P2
• Severity: medium
• Staleness: fresh
• Recommended next action: Define the auditable agent profile with capability/permission/authority distinctions and model/provider/version lineage tracking for correlated-agent detection.

Notes: Agent Epistemic Profile v0.1 — distinguishes what an agent can do, may know, may propose, may write durably, and may execute. Requires 15 core object types and bounded epistemic role profiles.