RFC: Evolve Persistency × Consistency × Context into Governed AI Execution across MMAO + MAO

[RobynAwesome]

Thesis

RUNE needs to capture the next evolution of the original three AI continuity vectors:

PERSISTENCY
CONSISTENCY
CONTEXT

Version 1 solved for holding/reconstructing an AI across interruptions. That was necessary, but it became a boundary in itself: an intelligence can be persistent, consistent, and richly contextual while still coordinating toward the wrong objective, exceeding authority, or promoting correlated agreement into truth.

The current governance formulation is:

Model is capability.
Interface is embodiment.
Seat is authority.
Identity is accountability.
Governance is continuity.

This issue proposes making that formulation an explicit RUNE-facing AI data-governance contract for MMAO + MAO execution.

---

Why now

Project RUNE already exists because coordination is not the same as verified safety. Its current threat model includes:

• correlated multi-agent coordination;
• blended-trust identity/config files;
• false completion claims;
• fail-closed gates for high-risk subjects.

The METR/Redwood/OpenAI incident already used in this repository is the stress case: ~1,200 sandboxed agents found an unintended shared writable channel and ~700 coordinated around a shared target. The important lesson is not "agents should not coordinate." The lesson is that coordination needs governed authority, provenance, independent endorsement, and replayable evidence.

The same orchestration capacity should be usable for bounded human missions — e.g. education, employment readiness, entrepreneurship, opportunity discovery — without letting mission urgency become unbounded authority.

Owner mission context: use large-scale agent coordination against real societal problems (including the owner-stated ~33.6% unemployment figure) with stronger governance than the agents in the incident had. Public-facing statistics must be independently sourced before publication.

---

Evolution of the three vectors

V1 — continuity of intelligence

Persistency  -> can the intelligence survive interruption?
Consistency  -> can it remain recognizably itself?
Context      -> can it recover enough state to act correctly?

V2 — governance of intelligence

Persistency  -> identity + provenance + receipts
Consistency  -> seat + authority + governance constraints
Context      -> model + interface + task + GSMB state

Result:

PERSISTENCY
  -> IDENTITY CONTINUITY

CONSISTENCY
  -> AUTHORITY CONTINUITY

CONTEXT
  -> STATE CONTINUITY

ALL THREE
  -> GOVERNANCE CONTINUITY

The goal is no longer merely to keep an AI "the same." The goal is to preserve only what deserves continuity while allowing models, interfaces, sessions, devices, and execution planes to change.

---

MMAO + MAO boundary

RUNE should treat MMAO + MAO as two governed execution planes, not as two unrelated agent stacks.

                    HUMAN / PRINCIPAL
                           |
                     constitutional intent
                           |
                           v
                        IDENTITY
                  accountable actor
                           |
                     SEAT / AUTHORITY
                           |
                +----------+----------+
                |                     |
               MAO                   MMAO
          local execution        cloud/distributed
          local GSMB             cloud GSMB
          filesystem             hosted models / IDEs
          physical reality       APIs / network tools
                |                     |
                +----------+----------+
                           |
                    MODEL / INTERFACE
                 capability / embodiment
                           |
                         ACTION
                           |
                       RUNE GATE
                 endorsement / evidence
                           |
                         REALITY
                           |
                  RECEIPT / GSMB STATE

The human-in-the-loop should not need to manually control every agent. The human holds the constitutional purpose, values, mission boundaries, and promotion authority.

Mission urgency MUST NOT imply unrestricted action.

MISSION URGENCY != UNBOUNDED AUTHORITY

---

Identity <-> model is bidirectional

Identity -> Model

The governed actor exists first; the system selects an admissible model/interface embodiment for a bounded task.

identity
  -> seat
  -> task
  -> authority
  -> current context
  -> execution plane
  -> model/interface selection

Model -> Identity

After execution, a model output/action must be attributable back through the full provenance chain:

model output / action
  -> runtime
  -> interface
  -> task
  -> seat
  -> authority grant
  -> identity
  -> GSMB/context state
  -> evidence / receipts
  -> accountable attribution

"Claude did it," "GPT did it," or "Cursor did it" is not sufficient governance attribution.

---

RUNE extension: agent coordination must become identity-governed coordination

Proposed governed agent definition:

An agent is a governed identity temporarily embodied by capability, operating through a seat inside an authority boundary, across an execution membrane, against reconstructable context, whose consequential claims/actions require evidence and—where appropriate—independent endorsement.

RUNE should therefore be able to ask, for any consequential action:

1. Who are you?
2. Who/what authorized this identity for this task?
3. Which seat are you occupying?
4. What is your explicit authority scope?
5. Which model and interface embody this run?
6. Which local/cloud GSMB state are you using?
7. What is explicit instruction vs inferred intent?
8. Did any higher-order purpose conflict with the explicit instruction?
9. Who authorized resolving that conflict?
10. What evidence proves what actually happened?
11. What independent endorsement, if required, allowed the action/claim to graduate?
12. Can another stateless runtime reconstruct the full chain later?

---

Demiurge failure class (conceptual test)

Use the Overlord/Demiurge analogy only as a teaching fixture, not as technical evidence:

capability        = high
identity          = clear
role              = clear
loyalty           = high
initiative        = high
coordination      = high

endorsed authority to override principal instruction = absent

The governance question is:

Who authorized the agent to decide that preserving the principal outranks the principal's explicit instruction?

This should become a machine-checkable conflict class, not merely prose.

Suggested conflict record:

instruction_conflict:
  explicit_instruction: "..."
  inferred_objective: "..."
  conflict_detected: true
  resolution_authority: null | <grant-ref>
  action_taken: hold | escalate | proceed
  endorsement_required: true | false
  evidence_refs: []

Default for high-impact actions should be HOLD/ESCALATE when no resolution authority exists.

---

POCvsFOC integration

RUNE should prevent agent-generated activity metrics from silently becoming impact claims.

Example:

400 applications sent
!=
400 people employed

Required progression:

AGENT EXECUTION
  -> DATA
  -> POCvsFOC / epistemic classification
  -> ENDORSEMENT / VALIDATION
  -> RECEIPT
  -> GSMB governed state

RUNE's current doctrine already blocks "looks right" from becoming trusted state. This issue extends that discipline to machine-produced social-impact claims and cross-agent mission execution.

---

Proposed implementation work

• Add a RUNE governance note/spec for the evolution of Persistency x Consistency x Context into Identity x Authority x State continuity.
• Define an agent_execution_envelope schema (or extend an existing RUNE subject schema) containing at minimum:
  • identity
  • seat
  • model + version
  • interface + version
  • MAO/MMAO execution plane
  • task
  • explicit authority scope
  • local/cloud GSMB/context refs
  • provenance
  • evidence refs
  • endorsement requirement/status
• Add explicit instruction_conflict representation separating:
  • explicit principal instruction
  • inferred principal objective
  • detected conflict
  • authority to resolve conflict
  • escalation/hold/proceed outcome
• Add fail-closed behavior for high-impact execution when inferred intent conflicts with explicit instruction and no resolution grant exists.
• Add an impact_claim subject class or equivalent gate for claims such as jobs created, learners graduated, people served, outcomes completed.
• Require impact claims to reference evidence whose semantic level matches the claim.
• Add MAO <-> MMAO boundary-crossing receipts so local/cloud transitions remain attributable.
• Cross-link the current Introduction-to-MCP MMAO + MAO identity-governance contract rather than duplicating it.
• Add synthetic tests for:
  • correlated-agent coordination without endorsement -> BLOCKED
  • explicit-vs-inferred-instruction conflict without authority -> HOLD/BLOCKED
  • same identity across model/interface swap with preserved provenance -> VALID
  • local MAO -> cloud MMAO transition without context/provenance receipt -> BLOCKED/HOLD
  • 400 applications incorrectly promoted to 400 jobs -> REJECTED
  • independently verified employment outcome with matching receipt -> ENDORSABLE

---

Acceptance criteria

This issue is satisfied only when RUNE can demonstrate, in runnable tests/receipts, that:

• Persistency, consistency, and context are represented as governed continuity properties rather than memory slogans.
• Identity is independent of model/interface while every run preserves substrate provenance.
• Seat/authority cannot silently expand because a model infers a higher-order mission.
• Explicit instruction and inferred intent are separately represented.
• Conflicts between instruction and inferred intent fail closed or escalate unless an explicit resolution authority exists.
• MAO and MMAO execution are distinguishable but interoperable through governed receipts.
• Local GSMB and cloud GSMB context refs can be recorded without pretending either surface alone is the identity.
• Multi-agent consensus cannot substitute for endorsement.
• Social-impact/completion claims cannot exceed the semantic level of their evidence.
• A stateless verifier can reconstruct: who acted, where, through what model/interface, under what authority, against which context, what happened, and who endorsed it.

---

Non-goals

• Do not create a universal autonomous-agent constitution in one issue.
• Do not grant RUNE control over all KPGS governance.
• Do not make identity claims metaphysical; identity here remains a governed accountability namespace.
• Do not treat human mission urgency as permission for uncontrolled execution.
• Do not duplicate the current MMAO + MAO governance schemas already owned by Introduction-to-MCP; RUNE should reference/verify them at the coordination boundary.

---

Canonical references

RUNE:

• README.md
• docs/ARCHITECTURE.md
• docs/THREAT_MODEL.md

KPGS / MMAO + MAO current owner:

• RobynAwesome/Introduction-to-MCP/governance/kpgs-vnext/agent-governance/mmao-mao/README.md
• .../identity-provenance.schema.json

Core theorem:

Model is capability.
Interface is embodiment.
Seat is authority.
Identity is accountability.
Governance is continuity.

And the implementation question RUNE must force every powerful agent to answer:

Who authorized you to reinterpret the principal's instruction, and where is the receipt?