# wdc-runtime Local reference implementation for **World-Domain Cognitive Runtime v0.1**. Current scope: **Phase 0–11** — ledger/world/checkpoint/fork kernel, role/authority isolation, bounded Governor state, dependence-aware evidence, computation-portfolio routing, an explicit sandbox Commit Gate, versioned TCD evidence assimilation / Historical Sedimentation, and an auditable World Ensemble Learning layer with source gates, holdout validation, rollback, Governor misses, and self-sealing warnings. ## Run tests ```bash python -m pytest -q ``` ## Run the Branching Grid demo ```bash PYTHONPATH=src:. python examples/branching_grid.py /tmp/wdc-branching-grid ``` The demo creates one parent world, checkpoints it, forks two child worlds from the same exact prefix, applies divergent actions, persists fork events, and prints divergent terminal states. ## Run the Governed Evidence Grid demo ```bash PYTHONPATH=src:. python examples/governed_evidence_grid.py /tmp/wdc-governed-evidence-grid ``` The demo adds sibling branch blindness, bounded Governor allocation, dependence-aware support/counter evidence, counterexample escalation, and redundant-branch tombstoning on top of the exact fork kernel. ## Run the Tri-Temporal Commit Grid demo ```bash PYTHONPATH=src:. python examples/tri_temporal_commit_grid.py /tmp/wdc-tri-temporal-grid ``` This demo closes the first controlled TCD/WDC loop: governed branching evidence produces an independence deficit and a `CROSS_BACKEND` computation proposal; the counterevidence revises Future/Present/Past-relevance at the same parent time; an approved sandbox-only commit executes; only that executed sandbox outcome is sedimented into the next parent historical state. ## Run the World Ensemble Learning demo ```bash PYTHONPATH=src:. python -m examples.world_ensemble_learning /tmp/wdc-world-learning ``` This demo adds Phase 9 learning on top of the tri-temporal loop: world-relative evidence can update Generator/FutureSpace versions, WORLD-only evidence is denied reality-facing promotion, a registered external calibration anchor allows a WorldModel candidate into validation, a deliberate holdout regression rolls it back, a GovernanceMiss calibrates the Governor meta-policy, and deteriorating external quality triggers self-sealing warnings. Learning itself does not advance parent historical time. ## Hard invariants already implemented - `WorldSpec != WorldRun` - three clock fields are explicit in the event envelope - exact checkpoints preserve state, local time, and RNG state - each fork creates a new world ID - fork provenance records parent run, checkpoint, delta, and seed policy - lineage cycles are rejected - sibling post-fork access is denied without an explicit channel - fork authority profiles cannot silently escalate external permissions - mutable Governor lifecycle does not mutate immutable `WorldSpec` - active Governor allocation cannot exceed the global budget - `INVALID` evidence is not counterevidence - aggregates preserve packet inputs and expose evidence-family dependence - v0.1 does not fabricate a universal effective evidence count - event payloads are persisted as snapshots - computation actions are proposals, not execution authority - Commit Gate verifies referenced evidence aggregates exist - sandbox proxy requires an approved exact action payload - TCD world-evidence assimilation preserves parent-real past facts and parent time - sandbox Historical Sedimentation requires an approved commit + executed sandbox action - `EXTERNAL_REAL` sedimentation requires a verified `PARENT_REAL_OBSERVATION` external ingest - WORLD/SYNTHETIC/DERIVED evidence cannot self-promote to `REALITY_FACING` learning - `REAL`/`EXTERNAL` reality-facing learning requires a registered anchor backed by a `VERIFIED` external ingest - learning candidates activate only after holdout validation and can rollback - learning-health warnings are diagnostic and do not mutate evidence/history automatically - blobs are content-addressed and integrity-checked See `docs/specs/` for the implementation contracts and `docs/superpowers/plans/` for the executed Phase 0–2, Phase 3–5, Phase 6–8, Phase 9, Phase 10, and Phase 11 plans. ## Phase 10 CLI Phase 10 exposes the Phase 0–9 kernel through one JSON-first local command surface. Initialize a workspace: ```bash wdc --root /tmp/wdc-runtime init wdc --root /tmp/wdc-runtime status ``` Complex payloads may be inline JSON or `@file.json`: ```bash wdc --root /tmp/wdc-runtime tcd init --json @tcd-init.json wdc --root /tmp/wdc-runtime world create --json @world.json wdc --root /tmp/wdc-runtime evidence claim-create --json @claim.json wdc --root /tmp/wdc-runtime portfolio route --json @deficit.json wdc --root /tmp/wdc-runtime commit assess --json @commit.json wdc --root /tmp/wdc-runtime learning propose --json @learning-update.json ``` Reference demos are available through the installed runtime surface: ```bash wdc --root /tmp/wdc-demo demo branching-grid wdc --root /tmp/wdc-demo demo governed-evidence wdc --root /tmp/wdc-demo demo tri-temporal wdc --root /tmp/wdc-demo demo learning ``` All successful CLI commands emit JSON to stdout. Expected domain/runtime failures emit a JSON error object to stderr with exit code `2`. The CLI delegates to the existing kernel services; it does not define alternate evidence, authority, TCD, or learning semantics. ## Phase 11 External Evidence / Adapter Layer Phase 11 adds the first controlled non-WDC evidence boundary. External artifacts are read by adapters, stored as content-addressed blobs, registered with immutable source provenance, and only then may be converted into ordinary `EvidencePacket` objects. Reference commands: ```bash wdc --root /tmp/wdc-runtime external adapters wdc --root /tmp/wdc-runtime external ingest --json @external-ingest.json wdc --root /tmp/wdc-runtime external show wdc --root /tmp/wdc-runtime external evidence-add --json @evidence.json wdc --root /tmp/wdc-runtime demo external-evidence ``` The v0.1 reference adapter is `local-json`; it supports canonical inline JSON and local JSON files and performs no network access. A manual `learning source-anchor` record is no longer sufficient to unlock `REALITY_FACING` learning: the provenance must also resolve to a `VERIFIED` external ingest. `EXTERNAL_REAL` Historical Sedimentation is now available only for verified ingests whose observation scope is `PARENT_REAL_OBSERVATION`. Ordinary external datasets remain evidence and cannot silently become parent-real historical facts.