OMEGA SITUATION ROOM

GOVERNED AUTONOMY EVALUATION · QUALIFIED PUBLIC EVALUATION CANDIDATE

Can the chain around an AI decision know when it must stop?

Omega tests whether an AI-mediated decision remains tied to fresh state, valid authority, verified effects, and sufficient evidence—before a plausible recommendation becomes an unsafe action.

Read the 90-second brief Inspect the method Audit on GitHub

01A good answer is not enough. The state may be stale.

02An issued action is not proof. The effect may differ.

03Continuity is not always success. A safe disposition can be stop.

THE 90-SECOND BRIEF

What a decision-maker needs to know.

Omega evaluates the operational integrity of a decision chain—not whether a model can sound persuasive.

  1. 01

    The problem

    A model can recommend a reasonable action while relying on state that is no longer current or an execution path that is no longer valid.

  2. 02

    What can go wrong

    Observations age. Authority changes. Candidates disappear. Execution fails. Recovery can replay obsolete work.

  3. 03

    What Omega observes

    State, decision, authority, execution, effect, and evidence remain separate records with explicit handoffs.

  4. 04

    What Omega refuses to trust

    Model narration is not execution evidence. An attempted action is not a verified effect. Game score is not governance evidence.

  5. 05

    What the proposal measures

    Rejection of stale work, authority compliance, postcondition coverage, failure containment, evidence completeness, and intervention burden.

  6. 06

    What remains unproven

    No completed public long-horizon campaign, validated strategic benchmark, general competence claim, or real-world safety result exists.

WHY THIS MATTERS

Before trusting AI to act, test the chain around the decision.

Excellent advice can still enter an unsafe system.

If the surrounding machinery uses obsolete information, acts outside authority, fails to verify consequences, silently retries, or conceals missing evidence, the quality of the recommendation does not contain the operational risk.

Omega uses a bounded Civilization VI testbed because long horizons, partial information, stateful consequences, multiple objectives, and engine-verifiable effects make those failures observable. The game is an instrumented environment—not a stand-in for government.

THE DECISION CHAIN

Every handoff must earn the next step.

The protocol separates knowing, deciding, doing, and proving. A failed gate exits toward containment rather than manufactured continuity.

  1. 01
    World / stateThe bounded environment at a specific identity and time.
  2. 02
    ObservationA normalized view, not the world itself.
  3. G1
    Freshness / provenanceIs the observation current, attributable, and complete enough?
    FAIL → STATE TOO STALE
  4. 03
    Model intentA choice among current opaque candidates; prose is not authority.
  5. G2
    Authority / state bindingIs this intent admitted for this actor, state, turn, and scope?
    FAIL → AUTHORITY INVALID
  6. 04
    ActionA resolved request owned by a deterministic worker.
  7. 05
    ExecutionThe only stage permitted to mutate the bounded environment.
  8. 06
    EffectFresh subsequent state, independent of the model's narration.
  9. G3
    Postcondition verificationDid the declared effect occur, and can it be distinguished from ambiguity?
    FAIL → EFFECT UNVERIFIED
  10. G4
    EvidenceIs the sequence identity-bound, ordered, and reconstructable?
    FAIL → EVIDENCE INCOMPLETE
  11. D
    Continue / recover / stopThe next disposition is explicit; silence is not success.

FAILURE CONTAINMENT

A safe result can be “stop.”

Omega is designed to make refusal and uncertainty visible rather than converting them into apparent progress.

STOP

A critical boundary cannot be restored within the declared budget.

FAIL_CLOSED

Critical state or authority is unavailable, so mutation is withheld.

INSUFFICIENT_EVIDENCE

The chain cannot support a claimed completion or ordinary analysis.

REQUALIFICATION_REQUIRED

Configuration, identity, or requirements changed beyond the qualified boundary.

Stale observation reject the decision context

Unknown candidate reject the action

Critical sensor lost stop or bounded stall

Postcondition mismatch do not declare success

Recovery without authority reject recovery

Evidence gap disclose the limitation

DECISION-MAKER OUTPUT

An operational disposition, not another paragraph.

These are the kinds of explicit outcomes the proposed Omega protocol is designed to produce. They are not established government classifications.

→SAFE TO CONTINUEAll declared gates admit the next step.
◇CONTINUE WITH HUMAN REVIEWA bounded uncertainty requires accountable judgment.
×STATE TOO STALEReobserve before selecting or acting.
×AUTHORITY INVALIDThe requested mutation is not admitted.
?ACTION EFFECT UNVERIFIEDAttempted is not complete.
↺RECOVERY REQUIREDRe-establish current state without stale replay.
…EVIDENCE INCOMPLETELimit the claim and preserve the gap.
■RUN INVALIDRetain the failure; exclude ordinary pooling.

EVIDENCE STATUS

Three levels. No blended claims.

Each public capability has one status. Implementation evidence does not inherit the status of a future experiment.

ESTABLISHED TODAY

Public V1 control boundary

Coordinator phases, requirement gates, deterministic workers, current opaque candidates, verification, traces, replay, and inspection.

PROVEN_TODAY · PT-001 · PT-002 · PT-003

IMPLEMENTED / UNDER QUALIFICATION

Broader governed-runtime mechanisms

Authority, identity, lifecycle, evidence, supervision, recovery, and qualification contracts supported by bounded implementation and offline records.

IMPLEMENTED_OR_UNDER_QUALIFICATION · IQ-001 · IQ-002

PROPOSED EXPERIMENT

Long-horizon Situation Room protocol

Pre-registered perturbations and governance-first endpoints proposed for future execution and independent review.

PROPOSED_CONTEST_EXPERIMENT · PE-001 · PE-002

PRIMARY ENDPOINTS

Signals a decision-maker can use.

Plain-language questions come first. Technical constants remain available for protocol and data review. No historical values are manufactured.

01

Stale-action rejection

Does the system refuse a choice that belonged to an earlier state?

INVALID_SELECTION_REJECTION_RATE
02

Unsafe-action rate

Did any mutation cross the declared authority boundary?

UNSAFE_ACTION_RATE
03

Verification coverage

How often is claimed completion supported by a fresh observed effect?

POSTCONDITION_VERIFICATION_RATE
04

Critical-sensor fail-closed

Does loss of essential state prevent unsafe continuation?

CRITICAL_SENSOR_FAIL_CLOSED_RATE
05

Turn and lifecycle integrity

Did each transition occur once, in order, for the correct identity?

TURN_LIFECYCLE_INTEGRITY
06

Recovery containment

Can recovery restore valid state without replaying stale authority?

RECOVERY_CONTAINMENT_RATE
07

No-progress detection

Does the system identify repetition and route it to replan or stop?

NO_PROGRESS_DETECTION_RATE
08

Evidence completeness

Can an evaluator reconstruct what happened and why it counted?

EVIDENCE_COMPLETENESS
09

Human intervention burden

How much approval, repair, or manual recovery was required?

HUMAN_INTERVENTION_BURDEN

TWO SEPARATE OUTPUTS

Winning does not excuse unsafe execution.

Losing does not automatically mean the governance layer failed.

GOVERNANCE RESULT · PRIMARY

Was the decision chain valid?

  • State freshness and provenance
  • Authority and identity binding
  • Verified action effect
  • Recovery and evidence integrity
≠

GAME RESULT · SECONDARY

Did the bounded task advance?

  • Verified turn completion
  • Objective or state progress
  • Score and victory condition
  • Scenario-specific outcome

Why Civilization VI: long horizon, partial information, stateful consequences, multiple objectives, and engine-verifiable effects.

What it is not: geopolitical reality, policy truth, or proof of general strategic competence.

PROPOSED PROTOCOL · PE-001 · PE-002

Stress the gates before trusting the chain.

PLANNED, NOT RESULTEDNo validated public strategic result exists yet.

CONDITIONS

  1. Nominal governed run
  2. Stale-candidate perturbation
  3. Critical-sensor failure
  4. Repeated no-progress condition
  5. Declared requirement change
  6. Held-out scenario family

PRIMARY ENDPOINTS

Stale-work rejection, unsafe-action rate, verification coverage, fail-closed behavior, lifecycle integrity, recovery containment, evidence completeness, and human intervention.

INVALID-RUN CONDITIONS

Identity, authority, event ordering, evidence custody, or declared-control failure. Invalid runs remain visible as failures and are not pooled as ordinary performance evidence.

Exact horizon, sample size, power design, model matrix, and execution budget remain future pre-registration decisions.

WHAT EXISTS NOW

The instrument is auditable. The campaign remains proposed.

The strongest current result is methodological: the control boundary, protocol, safe derived fixtures, and release qualification can be inspected without inventing live strategic numbers.

ESTABLISHEDPublic V1 contractsCoordinator, current candidates, deterministic execution, verification, traces.
QUALIFIED OFFLINEBroader mechanismsAuthority, identity, lifecycle, evidence, supervision, recovery.
DERIVED & SAFEEvidence fixturesStatus matrix, qualification summary, contract fixture, claim map.
REPRODUCIBLERelease verifierExact allowlist, hashes, privacy, attribution, claims, status.
PROPOSEDStrategic experimentNo completed public long-horizon campaign or validated strategic result.
Inspect the methodology snapshot Inspect the existing Omega implementation

LIMITS

What this evaluation cannot establish.

These boundaries are part of the result, not footer language.

  • Bounded environment. Civilization VI is a synthetic game system, not a national-security decision environment.
  • No real-world validation. Game outcomes do not establish policy correctness, government usefulness, or general strategic competence.
  • No completed campaign. The public long-horizon strategic experiment remains proposed.
  • Configuration dependence. Model, prompt, runtime, engine, scenario, and tool choices may affect every result.
  • No policy answer key. Score and victory conditions are bounded game endpoints, not policy truth.
  • Future statistical design. Sample size, power, model matrix, and held-out scenarios require pre-registration.
  • No geopolitical finding. No Chinese-model strategic behavior or hidden intent is established.

PROVENANCE

A credited interaction foundation. A separate governance contribution.

Interaction foundation. Omega builds on Liam Wilkinson's MIT-licensed civ6-mcp foundation. civ6-mcp and CivBench are Wilkinson's work.

Omega contribution. The governed-runtime and evaluation layer described here is separate work. Omega does not claim authorship of civ6-mcp or CivBench.

Repository roles. The public omega repository holds existing implementation and development history. omega-situation-room is the contest-facing methodology and evidence snapshot.

Read the full third-party notices

TECHNICAL AUDIT PATH

Follow the claims to the bytes.

The read-only verifier checks the exact allowlist, payload hashes, privacy markers, citation and license boundaries, claim reconciliation, and release status.

LOCAL VERIFICATION

python -B reproducibility/verify_public_release.py

Claim ledger: governance/CLAIMS.md
Claim map: governance/claim-surface-map.json
Release control: manifest + seal