Deterministic renders and the content-hash cache

decodx has one axiom it never breaks: the same spec and the same assets produce the same MP4, byte for byte. No model runs in the render loop. That determinism is not just tidy — it is what makes re-renders free.

Content-hash caching

Every expensive step — a terminal recording, a browser capture, TTS narration, an avatar clip — is keyed by a hash of its inputs. Before producing an asset, decodx checks the cache for that hash. A hit is reused; only a miss does real work.

So when you change one scene, its hash changes and that scene is rebuilt — but every other scene keeps its old hash and is served straight from cache. A spec tweak that would be a full re-export in a timeline editor becomes a few seconds of work here.

Why no model in the loop

Models are wonderful at drafting a spec. They are a liability inside the render, where the same input must always yield the same output. decodx keeps the boundary sharp: models help you author the spec; a deterministic pipeline renders it.

• Reproducible builds — a render on your machine matches CI matches production.
• Honest caching — identical inputs are provably identical outputs, so reuse is safe.
• Auditable cost — you re-buy an asset only when its inputs actually changed.

Determinism is the quiet feature underneath everything else decodx promises: cheap iteration, agent automation, and a video that stays in sync with your product.