Approved exact-plan work
Authorization, plan alignment, capability, constraints, execution conditions, and required evidence support continuation.
This document turns the DSG ONE architecture into an explorable 3D model. Drag the scene, select a trust domain, inspect its authority boundary, then follow the implementation and evidence sections below.
The authority layer distinguishes exact-plan work, capability gaps, and unsupported or out-of-plan actions.
Authorization, plan alignment, capability, constraints, execution conditions, and required evidence support continuation.
The step remains inside the approved plan, but execution pauses until the required permission or capability is provisioned.
Missing authority, failed constraints, incomplete proof, or scope violation does not become success.
Admission requires composite improvement, real-efficiency improvement, and no regression in protected metrics such as P99 latency, error rate, throughput, cost efficiency, audit completeness, human override rate, SLA status, and zero-data-loss guarantees.
ADMIT = scoreImproved ∧ efficiencyImproved ∧ noProtectedRegressionNatural-language hints do not become solver code. Search completion and whole-space search are tracked independently.
missing constraints.aimoEncoding → REVIEWsearchComplete ≠ wholeSpaceSearchedAfter a candidate is found, an independent solver asks whether any assignment exists with strictly lower energy.
verifier_objective < candidate_energyUNSAT → VERIFIED_GLOBAL_OPTIMUM. Better witness → COUNTEREXAMPLE_FOUND. Timeout → VERIFICATION_TIMEOUT.
candidateAuthority = SIMULATION_ONLYpromotionAuthority = DSG_CONTROL_PLANEselfPromotionAllowed = falsePlan hash, commits, allowed paths, constraints, tests, build, objective improvement, evidence, and Cinema proof binding are checked before promotion.
BLOCKED and FAILED remain explicit. Required lanes need a joinable state, commit SHA, evidence, and no blockers.
Sessions bind to an approved plan, approved step, and agent identity. Verifier-side access is observation-only; delegated secrets use opaque references outside the model/evidence path.
secret_ref / otp_refHTTPS, private-address/DNS protections, SSRF-style rejection, and expiring AES-GCM session sealing protect the remote boundary.
PASS → COMMIT_NEXT_BASELINEREVIEW → HOLD_REVIEWBLOCK → EXECUTE_ROLLBACKMonitoring observes; canonical authority owns mutation. Unbound production targets fail closed.
PRODUCTION_TARGET_UNBOUND → BLOCKDSG-Agent-v0 can synthesize repair candidates, but deterministic SDK validation and Spacetime governance decide whether anything may proceed.
repair.synthesize → PROPOSAL_ONLYmissing authoritative input → BLOCKThe installable DSG Plugin points coding agents to the canonical Skills and Spacetime MCP surface rather than duplicating vendor-specific repair logic.
Scoped PATs remain management-plane credentials. W3C traceparent/tracestate/baggage can carry a trace identity into DSG evidence. Realtime stays behind project/session authorization, RLS, explicit schema/table selection, row filters, and selected columns.
PAT ≠ Realtime session credentialunknown schema/permission → BLOCKDSG ONE V1 on AWS reports source ea70fbecdde020b0a4d4e44180798418af209f27 with image digest sha256:2b79455abf1a4d8b29abccb384e55fe83f82d8b44a117fe86133a7024dfa06ad, source/digest binding true, and Microsoft Agent Framework 1.18.0 ready. DSG Spacetime production source-of-truth remains the private tdealer01-crypto/DSG-Private-Production-Runtime repository. Continuity main 07ad14724f0a8ca33dc1979a9f2913e7d004da63 passed exact-main CI 36534873806. The serving AWS repair source 5c8f75c167af9f019f31611597eea0b425fd7900 passed governed activation 36529273403 and external public MCP E2E 36529526202. That E2E proved read-only Cinema status ALLOW with a valid evidence chain and an unapproved connect BLOCK / APPROVAL_REQUIRED without an execution evidence increment. Approved browser execution and the installed ChatGPT connector binding remain separately gated. Previous receipts remain historical. Current AWS Spacetime signed GitHub read-only OAuth E2E is verified 11 Oct 2026; Cinema Connect/Execute and Quest/Unity E2E remain unverified.
line coverage = NOT MEASUREDbranch coverage = NOT MEASURED