Glossary

Kind: reference. This page gives the one meaning of each term that the typevet pages use.

Every typevet page uses these words with these meanings only. The entries are in alphabetical order.

Terms

acceptance review. An independent check of the submitted diff, accepted contract and parent consumer outcome. The reviewer preserves submitted files and may use explicitly assigned scratch permissions.

Banking77 proxy label. In typevet/finvet evals, PolyAI Banking77 intents collapsed via finvet’s six FRAUD_INTENTS into binary fraud / not_fraud for Noul-primary metrics. See Banking77 proxy and metrics.

CLINC domain shard. One of ten topical domains in CLINC150 OOS eval (domains.json). Each shard exposes exactly 15 intent slugs as a single Choice task (≤24 labels). Shards do not map to Banking77 categories or finvet FRAUD_INTENTS. See CLINC150 domain shard map.

composition root. The process or module that reads configuration, constructs adapters, runs work, and closes adapters when work ends. A future typevet CLI or MCP inbound adapter owns this role. Direct library callers own it in application code.

Decision. A TypeLLM-compiled field from JSON Schema (enum, boolean, bounded number or open string). The portable core executes Decisions; it is not the same as dumping a whole object through a grammar.

generation port. typevet’s MVP GenerationPort: prompt + JSON Schema + model → validated object. The transport floor on llama.cpp.

judgment port. typevet’s IO-free JudgmentPort: state + named Noul / Choice / Score questions + model → typed answers in a JudgmentResponse. Vocabulary aligns with judgevet SystemOnePort; no judgevet import. Logprob scoring adapters remain separate work (#26).

scoring port. typevet’s IO-free CandidateScoringPort (ScoringPort): model + rendered prefix + ordered candidate token-id sequences + score stage → identity-preserving raw logprobs per label in a CandidateScoringResult. Missing, duplicate, or unexpected candidate ids and non-finite scores fail closed (#117); stock HTTP pre-sampling adapters are (#27).

Jev. TypeSafe’s hosted System One model. judgevet calls it over HTTP. TypeLLM and typevet target open-weight typed decisions in the same family.

judgevet. Sister hex client for Jev. Port shape: system_one(state, questions, model) with Noul / Choice / Score answers and probabilities. A future consumer of typevet if typevet exposes a compatible judgment surface.

System One. TypeSafe’s typed-judgment model family (Noul, Choice, Score). TypeLLM is explicitly inspired by it. typevet aims at a local open path to similar guarantees.

Copilot. GitHub Copilot CLI (copilot). For typevet worker dispatches, request --model auto only. Named Copilot models (including Fable) are out.

flat module. One concern in one .py file at the package level. Prefer domain/models.py over a nested domain/models/ package. Package __init__.py files re-export the public surface and carry full module docs so agents and callers can discover names from the init.

glossary. The one allowed meaning of each domain term. One term per concept. The writing system binds prose to it.

harness. The program that runs a model and gives it tools. Orthogonal to role and to model weight. The worker run contract lists receipt-backed use and launch evidence.

issue bus. GitHub issues as the durable record for asks, triage, contracts, handoffs and acceptance across sessions and harnesses. Chat is not the bus.

Use issue bus, not issue-bus, in owned prose. The terminology gate enforces this spelling.

judgment. An issue label: a decision remains; the supervising or orchestrating model must settle it before worker-fit implementation.

model weight. Capacity class for a dispatch: light, medium or heavy. Orthogonal to harness. A light model may run on Claude, Cursor or pi; a heavy model likewise. Do not equate weight with a brand name.

reviewer. The role that exercises the defining behaviour and reports findings under the shared role. A model name alone does not make a review independent.

specifier. The role that writes a definition of ready and done for one issue. The specifier never edits code.

testing pyramid. Three layers: unit (fast, no network), contract (fake and real adapter share fixtures), live (opt-in exercised call). Default pytest excludes live. Coverage ≥ 90 on the default suite. Law in CLAUDE.md. Sisters: judgevet markers; gepa-adk ADR-005.

uv. Astral’s package and tool runner. Ground floor for typevet: sync, locks, uv run for every gate and script. No parallel pip/poetry install path.

worker. A bounded session that follows a brief, edits only allowed paths (or returns a named research artifact) and returns evidence. A worker never commits unless the brief assigns that action.

worker-fit. An issue label: the contract is decided and mechanical enough for any verified worker harness. It does not mean “use pi” and does not pick model weight.

worker trailer. A Generated-By or Specified-By commit trailer that records which worker model produced code or a specification. It is evaluation evidence, not authorship. Never use Co-Authored-By for a model.

writing system. The local prose profile informed by ASD-STE100 Issue 9, plus Diátaxis page kinds and Conventional Commits. See writing-system.md. Not a claim of full ASD-STE100 compliance.