orchestrator is a plan-driven media production harness. It turns a one-off generation workflow into a repeatable system with explicit inputs, plans, schemas, validation gates, and review artifacts.
This project is part of my transition from backend engineering toward XAI and agent explainability research. The main question it represents is:
How can an AI-assisted workflow leave enough structured evidence for humans to inspect, debug, and trust the result?
Instead of treating an AI-generated output as a black box, this repository keeps intermediate artifacts visible:
- source input scripts
- generated plan files
- schema validation
- quality gates
- project state documents
- known issues and review outputs
docs/CANONICAL_SPEC.md: canonical workflow and system interpretationdocs/QUALITY_BAR.md: quality gates for reproducibility and handoffschemas/edit-plan.schema.json: explicit plan schemascripts/validate_plan.js: plan validationtasks/: backlog, in-progress, and completed work trackingoutput/: generated plans and review artifactsremotion-app/: rendering surface that consumes validated plan artifacts
This repository is not an XAI benchmark yet. It is a practical engineering example for thinking about agent workflow evaluation:
- What evidence should an agent leave behind?
- Which intermediate artifacts are useful for debugging?
- Can validation logs and quality gates become part of an explainability/evaluation protocol?
- How should humans inspect AI-assisted workflows beyond final pass/fail outcomes?
Work in progress. The current direction is to keep the rendering pipeline functional while making every important step explicit, reproducible, and reviewable.