AZURE UI → CINEMA → EXACT Z3 → PROOF RECEIPT

Separate who proposes, proves, authorizes, executes, and observes.

DSG ONE is a governed execution architecture. Simulation may search. Cinema may verify. Monitoring may observe. Executors may act. Only the control authority may authorize promotion or post-deploy mutation.

Truth boundary: this page distinguishes implemented code paths from marketplace approval, certification, and external review. A Proof Receipt proves one bounded verification request; it is not SOC 2, ISO, regulatory, legal, or third-party certification.

SIMULATION_ONLYcandidate authority
DSG_CONTROL_PLANEcanonical authority
VERIFIED_GLOBAL_OPTIMUMexact proof status
01 · PROPOSAL DOMAINGenerate candidates without granting authority

User intent, agents, and deterministic simulation can propose work. Simulation cannot promote its own candidate.

Canonical architecture

Five trust domains. One authority boundary.

The system that produces a candidate does not get to approve its own candidate. The verifier does not execute. Monitoring does not mutate production. External side effects only occur after plan-bound authorization.

01 · PROPOSAL

Intent / Plan / Candidate

User goals, agents, and AGI Simulation create candidate actions or improved configurations. Candidate admission requires real objective improvement with no protected-metric regression.

PROPOSE ONLY
02 · PROOF

Cinema Proof Agent + Z3

Raw metric, test, build, commit, and plan bindings are independently checked. Exact QUBO verification only claims a global optimum when a separate Z3 proof obligation establishes that no better assignment exists.

VERIFY ONLY
03 · AUTHORITY

DSG Control Plane

Canonical gate checks plan hash, commit binding, path scope, constraints, tests, build, evidence completeness, and independent Cinema proof before promotion or post-deploy control is authorized.

CANONICAL AUTHORITY
04 · EXECUTION

Controlled Executor

API, MCP, browser, and deployment adapters perform only the approved side effect. Browser sessions are plan-bound; verifier controllers are read-only; sensitive identity material uses opaque references outside the model/evidence path.

EXECUTE IN SCOPE
05 · OBSERVATION

Evidence / Replay / Monitoring

Observed production results determine ACCEPT_NEXT_BASELINE, HOLD_REVIEW, or rollback recommendation. Monitoring remains observation-only; the control authority decides the mutation.

OBSERVE ONLY
CORE RULE

Proposal ≠ Authority

Simulation and agents can search and propose, but selfPromotionAllowed=false.

CORE RULE

Observation ≠ Authority

Monitoring can report PASS / REVIEW / BLOCK, but execution authority remains the control plane.

CORE RULE

Proof before claim

Timeout, incomplete shards, missing raw evidence, or unbound production targets do not become success claims.

Code-to-architecture map

Every major box points to code.

These references are the implementation anchors used for this architecture. They are not decorative labels.

Candidate admissiondsg-agi-simulation/src/candidate-admission.tsInspect candidate gate →
Deterministic QUBO / Ising searchdsg-agi-simulation/src/aimo/solver.tsInspect solver →
Exact Z3 verifierDSG-Cinema-Proof-Agent/z3_main.pyInspect verifier →
Plan-bound browser executionDSG-Cinema-Proof-Agent/api_v1/remote_browser.pyInspect remote browser →
Canonical plan authorizationDSG-Cinema-Proof-Agent/api_v1/decision_core.pyInspect decision core →
Durable guarded evidenceDSG-Cinema-Proof-Agent/api_v1/guarded_store.pyInspect guarded store →
Live verification

Run the real Cinema + exact Z3 path.

Choose one bounded scenario. The browser obtains a free capped API key when needed, calls production /verify/evaluate, and displays a decision only when the backend returns a structurally valid proof receipt with verified=true and VERIFIED_GLOBAL_OPTIMUM.

Inputs are fixed and bounded so repeated checks are comparable.
Do not enter secrets or customer data.
The free activation path requires no card and does not create a charge.
Production CinemaCHECKING
Result appears here

See ALLOW / REVIEW / BLOCK, why it happened, exact verification status, proof hashes, and the next action.

User benefit gate

The interface answers six operational questions.

USE

Where do I start?

Select a bounded scenario and press Run live verification.

SEE

Where is the result?

The result panel shows the decision and proof in the same screen.

OUTCOME

What did I get?

Plain ALLOW / REVIEW / BLOCK plus machine-readable proof status.

FIX

What must change?

The reason and next action identify the failing boundary instead of forcing log hunting.

PROOF

Is it supported?

Proof, request, and context hashes plus downloadable receipt bind the result.

EFFORT

Is it simpler?

One bounded run replaces separate policy, solver, evidence, and replay interpretation steps.

Distribution truth

Separate live delivery from external approval.

A repository package is not the same as marketplace approval. These labels describe the boundary without implying third-party certification.

Direct API

LIVE

Production Cinema API and self-serve free activation are publicly reachable.

GitHub Marketplace

LISTING / CUTOVER CHECK

Repository integration exists; paid entitlement cutover must be verified independently.

OpenAI Skills

PACKAGE READY

Skill package exists in the repository. External review remains separate.

Microsoft Marketplace

OFFER PACK READY

Offer metadata exists; this page does not claim Microsoft approval or certification.

AWS Marketplace

EXTERNAL WORK REMAINS

Seller onboarding and marketplace billing infrastructure remain external requirements.

JetBrains Marketplace

SPEC / PACKAGE BOUNDARY

Specification exists; installable artifact and marketplace review must be verified separately.

Free activation: checking…Billing status: checking…Checkout status: NOT VERIFIED / NOT LINKED
Trust boundaries

Know what passed, why, and prove it.

The architecture is deliberately asymmetric: candidate generators cannot promote themselves; verifiers cannot mutate; observers cannot execute; and unbound production targets fail closed.

PASS

Supported claim

Claim only what the bound plan, raw evidence, exact proof, execution result, and production binding support.

REVIEW

Incomplete evidence

Timeouts, partial search, missing replay, incomplete evidence, or external review remain visible as unresolved work.

BLOCK

Unsupported mutation

Out-of-plan actions, missing authority, constraint violations, or unbound targets do not execute.