Synthetic Production Intelligence (SPI)

Evidence-backed orchestration for generative media production. Not a text-to-video wrapper: SPI plans a production as structured state, renders it through swappable model backends, verifies results with evidence-carrying metrics, and (planned) repairs only what failed.

V0.1 status (Phase 1 — honest ledger)

Component	Status
ProjectState (versioned, mutation log, transactional deltas)	implemented
Evidence schema (PASS/REPAIR/REGENERATE/REPLAN/REJECT/UNRESOLVED/BLOCKED, evidence mandatory)	implemented
ProductionGraph (DAG, topo sort, cycle rejection)	implemented
Scheduler (priority, deps, retries, timing, cache hits)	implemented (mock lane)
Mock renderers + GenerationCache (§58)	implemented
Level-1 deterministic verification + adaptive evaluator routing	implemented
EvidenceFusion + MasterArbiter	implemented
IntentEngine / IdeaEngine / ConceptEngine	stubs (deterministic, no LLM) — real versions Phase 5
OutputCompiler	stub (draft manifest only)
Real video backend (Wan 2.2 primary, LTX-2 secondary)	not started — Phase 2
WorldEngine, StoryGraph, CausalGraph, RepairEngine, DirectorSim/Bench, 4K, audio	not started

See docs/architecture.md for the component map and docs/research_report.md for the Phase-2 backend decision (Wan 2.2, Apache-2.0; LTX-2 as fast secondary).

Install & test

pip install -e ".[dev]"
pytest -q

Mock mode (SPI_MODE=mock equivalent: SPIApi(mode="mock")) needs no GPU.

Minimal end-to-end (mock lane)

from spi.api.server import SPIApi

api = SPIApi(mode="mock")
r = api.create_project("A cartographer returns to the lighthouse she haunted as a child; must: no dialogue")
pid = r["project_id"]
api.interpret_project(pid)                      # intent + constraints + idea budget
api.commit_concept(pid, "IDEA_001")
api.compile_production(pid)                     # scenes -> shots -> generation DAG
api.render_draft(pid)                           # mock renderer, cached
verdict = api.evaluate_draft(pid)               # level-1 checks -> fusion -> arbiter
assert verdict["status"] in ("PASS", "REPAIR")
print(verdict["status"], verdict["evidence"][:1])

Operating modes

• SPI_MODE=mock — zero GPU, deterministic artifacts (current default)
• SPI_MODE=small — reserved for Wan 2.1-1.3B / LTX-Video-2B smoke tests (Phase 2)
• SPI_MODE=production — reserved for Wan 2.2 A14B (Phase 2+)

Budget policy

This repo is developed under SPI_INTERN_MODE=ultra_low: mock-first, cached research (docs/model_registry.json), no training without explicit justification, one smoke test only when necessary. Usage is tracked in artifacts/resource_ledger.json.

License

Code: Apache-2.0. Model backends carry their own licenses — see the registry.