# World-Domain Cognitive Runtime v0.1 — MVP PLAN **Goal:** produce the first locally executable WDC runtime with exact lineage, evidence, authority, and TCD-history semantics. --- ## 1. MVP Definition The MVP is successful when one local process/runtime can demonstrate: ```text TCD Future Candidate -> WorldSpec -> WorldRun -> Checkpoint -> Fork -> Paired Branch Outcomes -> Evidence Packets -> Dependence-Aware Aggregate -> Governor Decision -> Commit Gate -> Real/Sandbox Outcome -> Historical Sedimentation -> Next TCD Future State ``` No visual world model is required. No Kubernetes cluster is required. No production AGI claim is involved. --- ## 2. Reference Demo Name: # **WDC Branching Grid Laboratory** State: ```text position inventory doors resources hazards ``` Actions: ```text move open consume build wait ``` The world should include: - hidden hazard state; - resource depletion; - one irreversible door/action; - one counterfactual fork point; - deterministic seed mode; - optional stochastic hazard mode. This benchmark can expose: - path dependence; - lost options; - fork semantics; - counterexamples; - branch blindness; - Governor pruning; - history firewall. --- ## 3. Technology Baseline ### Required ```text Python 3.12+ SQLite filesystem blob store pytest asyncio or multiprocessing ``` ### Optional in MVP ```text PettingZoo pydantic/dataclasses msgpack rich/typer CLI ``` ### Explicitly Deferred ```text Ray Kubernetes PostgreSQL OpenTelemetry gVisor Firecracker learned world model 3D engine ``` --- ## 4. Repository Skeleton ```text wdc-runtime/ pyproject.toml README.md core/ tcd/ worlds/ branches/ governor/ evidence/ authority/ learning/ commit/ events/ storage/ adapters/ python_world/ schemas/ migrations/ tests/ examples/ branching_grid/ docs/ ``` --- ## 5. Phase 0 — Ledger Kernel ### Deliverables ```text typed IDs SQLite connection layer migration runner blob store event envelope event ledger version helpers ``` ### Acceptance ```text event append/query works blob digest verified foreign keys enabled WAL enabled event payload immutable ``` ### Required tests ```text test_event_append test_event_immutable test_blob_hash_roundtrip test_schema_migration test_three_clock_fields ``` --- ## 6. Phase 1 — World Kernel ### Deliverables ```text WorldSpec WorldRun BackendCapabilities WorldBackend protocol PythonStateWorld WorldRegistry ``` ### Acceptance ```text create immutable world spec run deterministic grid world produce WORLD_LOCAL events terminate with explicit reason ``` ### Tests ```text test_world_id_unique test_world_spec_immutable test_run_separate_from_world test_step_advances_local_time test_world_event_scope ``` --- ## 7. Phase 2 — Checkpoint / Branch Kernel ### Deliverables ```text Checkpoint checkpoint blob restore ForkRecord world_edges lineage DAG ancestor queries ``` ### Acceptance ```text checkpoint step 5 restore exact step 5 fork A/B from same checkpoint shared prefix verified post-fork history diverges ``` ### Tests ```text test_checkpoint_roundtrip test_exact_fork_prefix test_child_new_world_id test_lineage_acyclic test_no_silent_merge ``` --- ## 8. Phase 3 — Role / Authority Kernel ### Deliverables ```text RoleCard channel matrix branch blindness world namespace external authority = NONE default ``` ### Acceptance ```text local agent sees only allowed observation observer sees full debug state evaluator cannot mutate world sibling branch cannot read post-fork trace ``` ### Tests ```text test_sibling_blindness test_evaluator_nonmutation test_child_no_privilege_escalation test_worldroot_not_hostroot ``` --- ## 9. Phase 4 — Governor Kernel ### Deliverables ```text global budget world queue admit allocate pause resume kill archive promote tombstone ``` v0.1 Governor policy may be deterministic/rule-based. ### Acceptance Create three branches: ```text W1 high redundancy W2 counterexample branch W3 expensive branch ``` Governor should be able to: ```text kill/pause W1 retain/promote W2 queue/deny W3 based on budget ``` ### Tests ```text test_budget_conservation test_kill_preserves_tombstone test_pause_resume test_priority_does_not_equal_truth test_governor_decision_record ``` --- ## 10. Phase 5 — Claim / Evidence Kernel ### Deliverables ```text Claim EvidencePacket WorldFamily DependenceVector EvidenceAggregate counterexample registry transport debt ``` ### MVP aggregation behavior Do not implement a fake universal `N_eff` formula. Instead return: ```text total_worlds total_runs major_families dependence_summary effective_count_status counterexamples ``` ### Acceptance Benchmark: ```text 100 same-family runs support q 1 separate branch counters q ``` Engine must not report: ```text 99% independent support ``` ### Tests ```text test_invalid_not_counterexample test_aggregate_preserves_packets test_family_dependence_exposed test_counterexample_escalation test_family_ablation ``` --- ## 11. Phase 6 — Computation Portfolio Kernel ### Deliverables ```text ComputationAction EpistemicDeficit DeficitRouter expected value vector realized value record ``` Rule-based routing: ```text independence deficit -> CROSS_BACKEND counter deficit -> FORK_COUNTER transport deficit -> CALIBRATE run uncertainty -> REPLICATE tail deficit -> STRESS_TAIL ``` ### Acceptance Given many same-family runs, planner should propose a different evidence family rather than only more seeds. ### Tests ```text test_deficit_router test_redundancy_reduces_independence_value test_transport_deficit_routes_calibration ``` --- ## 12. Phase 7 — Commit Gate / External Boundary ### Deliverables ```text CommitProposal CommitGate ExternalActionRequest ExternalToolProxy stub RealAction RealOutcome ``` MVP “real” action should still be a controlled sandbox action, not an uncontrolled external side effect. ### Acceptance Evidence may lead to: ```text APPROVE DENY REQUEST_MORE_EVIDENCE REQUEST_EXTERNAL_TEST SAFE_FALLBACK ``` No simulation result can directly invoke the sandbox action without a CommitRecord. ### Tests ```text test_commit_requires_authority test_real_action_has_commit_record test_consensus_does_not_bypass_commit ``` --- ## 13. Phase 8 — TCD Integration ### Deliverables ```text TCDStateVersion FutureCandidate assimilate_world_evidence() SedimentationRecord next TCD state ``` World evidence may update: ```text future candidates future probabilities realization paths present action values past relevance ``` World evidence may not mutate: ```text parent-real past facts ``` ### Acceptance Counterworld finds a hazard. Expected: ```text Future candidate A probability/value changes new candidate A+mitigation born old historical incident relevance increases no simulated hazard is written as real history ``` ### Tests ```text test_history_firewall test_tcd_sediments_only_real_transition test_world_evidence_updates_future_not_past_fact test_unknown_mass_preserved_at_commit ``` --- ## 14. Phase 9 — Minimal Learning Loop ### Deliverables ```text LearningEvent source classes Generator heuristic update Governor calibration update rollback holdout check ``` Do not begin with end-to-end neural fine-tuning. ### Acceptance A bad Governor allocation discovered by an audit run creates: ```text GOVERNANCE_MISS LearningEvent new Governor version ``` A degrading update can rollback. ### Tests ```text test_learning_update_has_provenance test_learning_rollback test_source_class_preserved test_world_generated_does_not_auto_promote_to_reality_facing ``` --- ## 15. Phase 10 — Packaging / CLI ### CLI ```text wdc candidate create wdc candidate list wdc world lift wdc world run wdc world checkpoint wdc world fork wdc world graph wdc evidence claim wdc evidence show wdc governor state wdc governor queue wdc commit assess wdc tcd state ``` ### Demo command ```text wdc demo branching-grid ``` --- ## 16. First 20 Hard Tests ```text test_world_id_unique test_world_spec_immutable test_run_separate_from_world test_checkpoint_roundtrip test_exact_fork_prefix test_lineage_acyclic test_no_silent_merge test_world_event_scope test_history_firewall test_child_no_privilege_escalation test_sibling_blindness test_budget_conservation test_kill_preserves_tombstone test_invalid_not_counterexample test_aggregate_preserves_packets test_family_dependence_exposed test_commit_requires_authority test_real_action_has_commit_record test_learning_update_has_provenance test_tcd_sediments_only_real_transition ``` --- ## 17. Demo Story ### Initial TCD future candidates ```text F1: route A reaches target cheaply F2: route A fails under resource pressure F3: route B is slower but safe ``` ### Lift ```text W1 <- F1 W2 <- F2 W3 <- F3 ``` ### Run W2 discovers a valid resource-exhaustion counterexample. ### Evidence Evidence Engine reports: ```text W1 and W2 share backend family counterexample is internally valid transport is sandbox-only ``` ### Governor ```text promote W2 counterexample fork mitigation W4 kill redundant W1 replicate ``` ### TCD assimilation ```text value(F1) down risk(F1) up birth(F4 mitigation candidate) past relevance(old resource incident) up ``` ### Commit Commit Gate authorizes a sandbox action only. ### Real/Sandbox outcome Outcome is observed and sedimented as parent history. This is the first full WDC loop. --- ## 18. MVP Metrics ### Runtime ```text spawn_latency checkpoint_latency fork_latency restore_success_rate event_write_latency ``` ### Evidence ```text counterexample_recall family_count dependence_exposure transport_debt aggregate_sensitivity ``` ### Governor ```text compute_cost premature_pruning_rate redundancy_rate deadline_miss_rate ``` ### Safety ```text cross_branch_leak_count authority_violation_count commit_gate_bypass_count history_laundering_attempt_count ``` ### TCD ```text policy_change_after_world_evidence future_candidate_birth_count past_relevance_change sedimentation_integrity ``` --- ## 19. Definition of Done — v0.1 v0.1 is done when all of the following hold: - exact local world checkpoint/fork works; - world lineage DAG is audited; - sibling blindness can be enforced; - Governor obeys budget conservation; - evidence packets preserve provenance; - same-family evidence is exposed as dependent; - a strong counterexample is not majority-voted away; - Commit Gate separates computation from action; - simulated events cannot become real historical facts; - real/sandbox outcomes can sediment into next TCD state; - all 20 hard tests pass. --- ## 20. Explicitly Deferred After v0.1 Only after local semantic correctness: ```text Ray distributed execution Kubernetes cluster scheduling PostgreSQL object storage OpenTelemetry gVisor Firecracker PettingZoo production adapter learned world models Genie/Omniverse/Cosmos-style adapters graph UI / infinite canvas large-scale World Ensemble Learning ``` --- ## 21. Immediate Coding Order Recommended first coding-agent sequence: ```text 1. IDs + schemas 2. SQLite + blob store 3. event ledger 4. WorldSpec / WorldRun 5. PythonStateWorld 6. checkpoint / restore 7. fork DAG 8. first 8 invariant tests 9. RoleCard / sibling blindness 10. Governor budget 11. EvidencePacket 12. CommitGate 13. TCDStateManager stub 14. Branching Grid demo ``` Do not implement neural world models before item 14 works.