JEV Language

Universal Semantic, Language, Programming, and Expression Substrate for Jev

JEV Language is a harness-neutral, semantic-first substrate that turns bounded typed Jev judgments into increasingly general expression through explicit semantic representations, composition, deterministic compilation, search, and verification.

Core experiment: can a model that emits bounded typed judgments gain broad expressive power by operating over a compositional semantic substrate whose outputs are compiled and verified afterward?

Architectural invariants

• The native core requires zero generative-LLM calls.
• Meaning is canonicalized independently from surface wording.
• Deterministic rules dominate probabilistic preferences.
• Unknowns and ambiguity are preserved explicitly rather than guessed.
• Provenance, trust, replay, validation and verification are first-class.
• Jev is used for bounded semantic judgment, not hidden free-form generation.
• Exact external strings travel through source spans / opaque values / deterministic constructions, not semantic control-channel hacks.
• Downstream harnesses own execution, permissions, scheduling, UI and agent loops.

What is runnable now

The bootstrap implementation already includes:

• strict TypeScript + machine-checked package dependency boundaries;
• typed IDs, structured errors, canonical JSON, provenance and digests;
• JSG node/value algebra, graph transactions, serialization, restore, query and semantic diff;
• open-world spans, stale-reference detection, deterministic literal parsing and opaque sensitivity controls;
• ontology namespaces/core seed/provisional concepts;
• Jev Decision Runtime with recorded and official TypeSafe SDK adapters, budgets, cache, calibration hooks and usage accounting;
• versioned decision-pack registry;
• a controlled English parse → JSG → alternate realization → semantic round-trip verifier;
• recorded Jev Choice reference resolution that preserves ambiguity under low confidence;
• a controlled Vietnamese path that maps directly to the shared JSG and realizes from the same semantics;
• an initial Program IR with typed holes, type-guided filter synthesis, deterministic TypeScript lowering and compiler diagnostics.

These are narrow, tested vertical slices, not claims that broad language/code generation is already complete.

Repository map

packages/
  core-types/           foundational typed contracts
  provenance/           origin/trust records
  open-world-values/    source spans, opaque values, deterministic literals
  ontology/             concept/role/namespace infrastructure
  semantic-graph/       JSG
  semantic-validator/   semantic invariants
  decision-runtime/     JDR + TypeSafe/recorded adapters
  decision-packs/       versioned Jev judgments
  grounding/            deterministic controlled grounding
  dialogue-state/       dialogue/reference semantics (partial)
  realizer-core/        deterministic realization slices
  program-ir/           PIR (partial)
  synthesis-core/       typed-hole synthesis (partial)
  code-backend-core/    TypeScript lowering/compiler bridge (partial)
  ...

Unimplemented surfaces are materialized as explicit package stubs marked NOT_IMPLEMENTED or PARTIAL.

Development

npm install
npm run check
npm run benchmark

Ordinary CI runs only deterministic/recorded tests and never receives the TypeSafe key.

TypeSafe / Jev quota safety

Store the API key only as the GitHub Actions secret:

TYPESAFE_API_KEY

The JEV Live Smoke workflow is manual-only, requires an explicit YES, is hard-limited to one request / one question, sets SDK retries to zero, and prints TypeSafe-reported token usage. Broad live evaluations must be added deliberately with separate budgets.

Specification and state

• New AI / continuation entrypoint: docs/NEXT-AI-HANDOFF.md
• SPEC-PIN.md pins the supplied v0.4 master specification and digest.
• docs/IMPLEMENTATION-STATE.md is the authoritative progress snapshot.
• docs/TASK-LEDGER.yaml tracks implementation status without fake completion.
• docs/adr/ records architectural decisions.

Playground

A zero-install browser Playground is deployed at:

https://nolane-x.github.io/JEV-language/

It is a static GitHub Pages BYOK surface: every user supplies their own TypeSafe key, which stays in tab memory only. The browser does not call TypeSafe directly; all production browser requests go through the project's fixed, locked-down relay, which forwards the user's Authorization header transiently to TypeSafe without persisting it.

The NUI-driven interface uses an obsidian/editorial-instrument visual thesis with a warm-metal inspection-light field that follows the pointer, plus coarse-pointer and reduced-motion fallbacks. Provider success payloads are validated before rendering.

The static Pages shell was previously deployment-verified. A browser-equivalent preflight probe on 2026-09-20 proved that direct TypeSafe requests from the GitHub Pages origin are blocked by provider-side CORS. The production browser path therefore uses only the verified Worker relay. The UI no longer asks users to choose between direct and relay transports: it health-checks the fixed relay, validates the user's TypeSafe key through it, and then sends Jev requests through that same route. The relay was live-verified with a one-request smoke and does not persist the forwarded key.

Security

Never commit API keys, credentials, opaque secret payloads, or live traces containing them. A typed interface guarantees structure—not truth—so capability claims require evidence from conformance/evaluation.

v1 stable-core profile

The first stable ABI profile covers core-types and provenance. Its machine-readable evidence lives in docs/v1.0-conformance.json; API references, migration policy, benchmark baseline, limitations and an executable zero-generative replay demo are checked by the M20 conformance suite.

Useful commands:

npm run test:v1-conformance
npm run demo:v1-stable-core

This stable-core profile does not imply that every package is stable. Language, synthesis, backend and other research packages retain their explicit lower maturity levels until separately promoted.

Status

The T001–T500 engineering task sequence and v0.4 release gates are verified. M19 blinded natural-conversation protocol tooling is implemented, but the research milestone still requires real observed A/B stimuli and real blinded human ratings. The current natural-conversation continuation state, live-evidence totals, verified relay/browser state, and exact next actions are pinned in docs/NEXT-AI-HANDOFF.md. No synthetic human ratings or fabricated M19 outputs may be used to close M19.