TypeWright
Declarative compiler that turns state, decisions and labeled examples into a measured, frozen JSON program of typed Jev questions.
Overview
Typewright is a declarative compiler and lightweight runtime for typed Jev decisions. You describe state, decisions and labeled examples in YAML; Typewright searches within that fixed contract using a DSPy architect and a GEPA adapter, fits a review policy on separate data, and freezes the result into a checksummed JSON program. Your application loads and calls the frozen program without importing DSPy or GEPA.
It is alpha software with an explicit verification boundary. The no-key demo and core tests pass, and the TypeSafe, DSPy and GEPA integrations are implemented against documented interfaces, but no live provider study has been completed and no accuracy gain is claimed. Mock and synthetic metrics are pipeline checks, not Jev results, and a measured artifact is not an approved one. A build report documents what was actually run.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | User-declared Choice, Score and Noul decisions compiled from YAML | — |
How it works
- Declare state, decisions and labeled examples in YAML.
- Compile: search question wording within the contract and evaluate on splits.
- Fit calibration and review gates; freeze a checksummed JSON program.
- Load the frozen program at runtime and call it.
Strengths
- Separates compile-time optimization from a lightweight runtime.
- Uses held-out splits and review gates.
- Very explicit about what is and is not verified.
Caveats
- Alpha; no live accuracy study yet.
- Compile path depends on experimental DSPy integration.
At a glance
- Type
- Library
- Status
- prototype
- Language
- Python
- Runtime
- CLI + Python runtime
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Declare a decision, compile typed Jev questions, measure them, and ship a JSON…
wh000wh000/awesome-jev-liveDeclare a decision, compile typed Jev questions, measure them, and ship a JSON program.
hellogumbo/awesome-jev
Frequently asked questions
What does Typewright produce?
A checksummed JSON program of typed Jev questions plus review policy.
Does the runtime need DSPy?
No, DSPy and GEPA are used only at compile time.
Has it been shown to improve accuracy?
Not yet; the README says no live study has been completed.
Related projects
Tool that recomputes Jev's confidence identities on outputs published in public repositories to spot answers Jev never actually produced.
ServiceresearchPythonRust CLI and library that asks Jev typed questions via OpenRouter and combines the probabilities with fuzzy-logic rules into decisions.
CLIbetaRustUnofficial standard-library Go client for Jev with generic typed Choice values and local confidence thresholds.
ProjectbetaGoRust CLI for Jev: ask yes/no, choice and rubric questions about text and branch on exit codes, JSON or MCP tools.
MCP serverbetaRustNode/Bun CLI for asking Jev boolean, choice and score questions over text or JSON, with direct or Vercel AI Gateway providers.
CLIbetaTypeScriptTypeScript library adding policy, confidence gates, shadow mode, recipes and an eval CLI around Jev answers.
CLIbetaTypeScript