Inside the engine.
Multi-agent · orchestrator · explainability · audit · learning
PACCS is a layered intelligence architecture. Specialised agents operate in parallel; the orchestrator resolves divergence; the explainability engine surfaces reasoning; the audit trail makes every decision permanently traceable; the learning loop folds operator feedback back into the corpus.
Submission Pipeline
Films arrive through partner festival channels and direct submissions. Metadata, poster, trailer, synopsis ingested.
Multi-modal Encoder
Visual language profile, narrative pacing classification, dialogue density, score signature — all captured at ingestion.
Comparable-Film Index
The new submission is positioned in 4,726+ corpus space — identifying nearest neighbours by signal, outcome, and territory.
Narrative Agent
Story structure, dialogue quality, thematic depth, narrative voice. Produces narrative-tier signal with explicit reasoning trace.
Visual Language Agent
Cinematography, visual grammar, art direction signature. Compares against 4,726+ visual language vectors in the corpus.
Festival Alignment Agent
Programming-pattern matching across 22 anchor festivals, scaling to 200+ through PIFF 2026 partnerships. Submission-window aware. Tier-graded fit signals.
Marketability Agent
Audience modeling, distributor fit, territorial release positioning, multi-scenario revenue forecast.
Consensus Orchestrator
Resolves inter-agent variance. Weights agent signals against historical accuracy. Produces single programming-grade recommendation with confidence.
Reasoning Surface
Every score component traces to its source. Every divergence is reported. Every decision is replayable.
Human Channel
Always open. Operator decisions are logged, attributed, and folded back into the learning loop.
Audit Trail
Every analysis, override, and learning event recorded. Permanent retention. UK/EU region-locked.
Learning Loop
Operator feedback flows back as labelled training data. Model versioning is staged and roll-backable.
Source Citations
Every score traces back to specific corpus records, comparables, and partner-supplied signals. View →