Glossary

This page defines the locked vocabulary for v0.2 in one place. It is the shortest useful companion to AGENTS.md, docs/THESIS.md, and ADR-0011.

Locked terms

Harness — the L1 runtime that routes, retries, seeds, validates, budgets, and records each run.
Codec — the modality-owned implementation that turns a request into an artifact through a structured pipeline.
Spec — the structured modality contract that the LLM is asked to produce before decoding. It may contain human-readable semantic IR, compact visual seed code, or other typed modality sections.
IR — the typed internal representation a codec carries across expand / adapt / decode; at v0.2 only IR.Text is inhabited.
Decoder — the L3 component that turns IR-adjacent inputs into bytes on disk.
Adapter — the optional L4 learned bridge / seed expander between symbolic plans or compact code and decoder inputs.
Packaging — the L5 layer that exposes the system through CLI, docs, install surfaces, and reproducibility conventions.

Route — a codec-internal strategy path chosen by the codec, not by harness-level modality branching.
RunManifest — the receipts file that records git SHA, seed, artifact hashes, model/provider info, cost, latency, and errors.
produce — the canonical codec primitive from RFC-0001 that owns artifact generation end-to-end.
IR.Text — the only inhabited IR variant at v0.2; the text-native planning payload that codecs adapt and decode. For image, this now includes a Visual Seed Code-bearing contract rather than scene JSON alone.
No silent fallback — any route or model failure must surface as a structured error and a manifest row, never as an invisible replacement path.
Decoder ≠ generator — the repo allows deterministic or frozen decoders in the default path, not general-purpose image/video generation systems.

If a new term cannot be mapped to one of the words above, it does not belong in the v0.2 doctrine without a new RFC or ADR.