# CDI Runtime + AIVS — Engineering Specification v0.1 **Status:** IMPLEMENTABLE MVP / synthetic core tested **Date:** 2026-08-10 **Python:** 3.11+ **Dependencies:** Python standard library only **Primary OS target:** Windows 10/11 **Current test environment:** synthetic, non-Windows container **Integration target:** `mssp-game-computer-runtime-mvp` v0.8.x --- # 1. Build Order Do **not** start by writing an AI that rewrites a game binary. Implement in this order: ```text M0 Core State → M1 Windows Evidence Import → M2 Serialization Advisor → M3 Source-visible Shadow → M4 Game GEC → M5 Low-risk Promotion → M6 Live AIVS ``` --- # 2. Directory Layout ```text CDI_AIVS_series_paper06/ ├─ 06_CDI_Runtime_AIVS_工程架構協議與MVP_v0.1.md ├─ CDI_Runtime_MVP_spec_v0.1.md ├─ LOCAL_AI_HANDOFF.md ├─ SERIES_FINAL_INDEX_6of6.md ├─ 06_source_notes_2026-08-10.md ├─ README.md ├─ schemas/ │ ├─ sqlite_schema.sql │ ├─ vertical_sync_packet.schema.json │ ├─ compute_candidate.schema.json │ ├─ paradigm_profile.schema.json │ └─ game_equivalence_contract.schema.json ├─ examples/ │ ├─ runtime.example.json │ └─ scenario.game.example.json ├─ src/ │ └─ cdi_runtime_mvp.py └─ tests/ ├─ smoke_test.py ├─ MVP_SMOKE_TEST.json └─ cdi_smoke.db ``` --- # 3. Runtime Ports Define conceptual ports even if v0.1 uses one Python process: ```text EvidencePort PersistencePort RelayDecisionPort GovernorDecisionPort CandidatePort FencePort RoutePort BenchmarkPort GameStatePort ``` --- # 4. Evidence Adapter Contract Normalized event: ```json { "event_type": "cpu.region", "scope": "npc.paths", "timestamp": "...", "run_id": "...", "payload": {}, "evidence_ref": "..." } ``` MVP accepts generic JSONL. Future adapters: - `WpaCsvAdapter` - `EtwRealtimeAdapter` - `PixTimingAdapter` - `VtuneCsvAdapter` - `MsspDmsAdapter` --- # 5. Windows Collection v0.1 Official-tool-first: ```powershell wpr -start GeneralProfile -filemode # reproduce benchmark wpr -stop "cdi_trace.etl" "CDI capture" ``` Then export selected WPA tables: ```powershell wpaexporter.exe -i "cdi_trace.etl" ` -profile ".wpaProfile" ` -outputfolder "wpa_export" ``` Production work must define a CDI `.wpaProfile` and WPR profile. --- # 6. Why Not Native ETW First? Native ETW is a later adapter. The current research uncertainty is not “can Windows emit events?” It can. The research uncertainty is: ```text trace → causal region → serialization finding → candidate route → equivalence → promotion ``` Validate that first. --- # 7. State Model Run state: ```json { "epoch": 1, "topology_version": 1, "policy_version": 1, "current_state_version": 0 } ``` These versions MUST NOT be merged. --- # 8. Candidate Contract Every candidate must carry at least: ```text input_state_version read_keys write_keys epoch topology_version policy_version effect_class idempotency_key authority_scope/fencing_token when applicable ``` --- # 9. Relevant Conflict Algorithm Given candidate read set `R` and changed keys since read version `ΔW`: ```text if R ∩ ΔW != ∅: reject relevant_state_conflict else: version mismatch may continue ``` This is deliberately stronger than `if version changed: reject everything`. --- # 10. Commit Validation Order Use cheap deterministic checks first: ```text epoch → topology → policy → dependency status → invariant → effect barrier → fencing → relevant conflict → semantic AI only if unresolved ``` --- # 11. AIVS Regimes MVP: ```text R0 < 0.25 R1 < 0.50 R2 < 0.75 ESCALATE >= 0.75 ``` Pressure is synthetic weighted mean of: ```text anomaly drift conflict uncertainty risk novelty ``` This formula is a placeholder policy, not a learned truth. --- # 12. Relay Policy Default actions: ```text R0 → NOOP / machine check R1 → inspect local evidence R2 → hold candidate / deeper verifier ESCALATE → governor ``` --- # 13. Governor Policy Governor handles: ```text cross-domain conflict topology mutation policy mutation global route promotion high-risk external effect repeated relay failure ``` --- # 14. Paradigm Profile Do not require full 72 classification for every region. Progressive: ```text S/J/P/R → B/U/O if needed → L=F/K/Q if needed → modifiers if needed ``` Allow: ```text UNKNOWN HYBRID KEEP_ORIGINAL ``` as runtime sentinels. --- # 15. Route Candidate A route candidate is not active until: ```text feasible → shadow → equivalent → faster / otherwise useful → promotion receipt ``` --- # 16. Game Equivalence Contract At minimum define: ```text state fields RNG policy event ordering policy numeric tolerance required shadow runs minimum speedup ``` --- # 17. Game Epochs Do not equate frame and simulation tick. Future adapter: ```text sim_epoch render_epoch asset_epoch network_epoch ``` --- # 18. MSSP Integration Current MSSP is semantic/control evidence, not performance profiling. Preferred bridge: ```text MSSP DMS / structured events → JSONL → CDI import-jsonl ``` Later: ```text shared run_id shared epoch references MCP/local transport ``` --- # 19. Smoke Test Run: ```bash python tests/smoke_test.py ``` or: ```bash python src/cdi_runtime_mvp.py smoke --db tests/cdi_smoke.db ``` Expected: ```text PASS ``` --- # 20. Smoke Assertions The test MUST fail if any of these stop working: ```text R0 normal sync ESCALATE high pressure valid commit non-conflicting stale-version commit relevant stale conflict reject idempotent retry stale fencing reject irreversible speculation reject equivalent faster shadow promote non-equivalent faster shadow reject serialization gap calculation ``` --- # 21. First Windows Milestone Deliver: ```text real ETL file WPA export normalized JSONL CDI import top 20 regions by wall time thread wait summary candidate serialization findings ``` No automatic patch yet. --- # 22. Second Windows Milestone For source-visible app: ```text region → source function finding → candidate patch build tests shadow benchmark ``` --- # 23. First Game Milestone Pick a source-visible game/simulation satisfying: ```text Windows buildable no anti-cheat repeatable scenario clear main loop state probe available ``` Target only one subsystem. Recommended first candidate: ```text NPC path batch asset decoding read-only UI preprocessing ``` Avoid physics first unless test coverage is excellent. --- # 24. Promotion Levels ```text P0 Observe P1 Advise P2 Generate Candidate P3 Shadow P4 Local Low-Risk Commit P5 Domain Commit P6 Adaptive Global Routing ``` Do not skip levels. --- # 25. Metrics Performance: ```text p50/p95/p99 main thread ms worker utilization wait time GPU queue delay memory I/O ``` Correctness: ```text GEC pass state divergence RNG divergence crash deadlock rollback ``` AI: ```text calls tokens latency R0/R1/R2 counts escalations false finding ``` Control: ```text candidate count commit count reject reasons route promotions fallbacks ``` --- # 26. Acceptance for v0.2 A real Windows observer run is accepted when: ```text capture reproducible events normalized no process mutation required run persists in DB at least one serialization finding has evidence refs ``` --- # 27. Acceptance for v0.3 Source-visible optimization accepted when: ```text build passes existing tests pass shadow equivalence passes p95 not worse measured speedup > threshold fallback passes ``` --- # 28. Security Defaults ```text observer mode by default no administrator persistence no anti-cheat target no arbitrary binary rewrite no irreversible speculation AI decisions typed raw evidence retained ``` --- # 29. Technical Debt Intentionally Deferred ```text native ETW parser async multi-process relay real model connector JSON Schema runtime validation Parquet trace store learned routing controller GPU auto-codegen binary rewriting hard real-time guarantees ``` --- # 30. Engineering Principle When uncertain: ```text KEEP_ORIGINAL ``` is a successful decision. The MVP should optimize **only when it can preserve a credible verification path**.