Jev Playground

A TypeSafe System One playground: give Jev a messy real-world input and it turns it into fast, typed judgments that application code can use.

Jev is TypeSafe's flagship System One model. It answers narrow, typed questions — Choice, Score, and Noul — with calibrated probabilities. It does not generate prose and it does not decide workflows. This playground demonstrates the division of labor: Jev judges; this application decides.

Use cases

Use case	What Jev judges	What the application decides
Support Router	Intent category, recommended handler (Choice), urgency (Score), human-escalation and ambiguity (Noul)	Which queue receives the request, when to flag review, when to ask for clarification
Compare & Choose	One candidate per call: request-derived Score criteria plus a "description too sparse" Noul	The generated weighted composite, ranking order, shortlist vs. needs-info verdicts
Content Triage	Category (Choice), priority (Score), review-needed (Noul)	Reply / forward-to-support / archive / flag-for-review routing, batch sort and filters

Each use case is a self-contained module in src/lib/jev/usecases/ that declares its state builder, questions, normalization, thresholds, and examples. The single TypeSafe SDK boundary lives in src/lib/jev/core/typesafe.server.ts; use cases never touch the SDK.

Architecture

input → buildState → TypeSafe evaluate (core) → normalize → application rules → presentation

• src/lib/jev/core/ — shared judgment types, parse/normalize primitives, the generic evaluate() server function, formatting.
• src/lib/jev/usecases/registry.ts — the pipeline orchestrator (runUseCase, runUseCaseBatch) and use-case lookup.
• src/lib/jev/usecases/<name>/ — per use case: catalog.ts (data), questions.ts (state + Jev questions), normalize.ts (typed judgments), rules.ts (deterministic workflow), index.ts (UseCaseDef). Compare & Choose derives its Score question schema from the user's request before evaluation.
• src/routes/api/analyze.ts — one endpoint: GET returns connection status and the registry summary; POST evaluates { useCaseId, input } or a batch { useCaseId, batch: [...] }.
• src/components/playground.tsx — the workbench; per-use-case input and result components; developer-panel.tsx exposes the full pipeline.

See docs/architecture.md for details.

Probability semantics

• A Choice answer reports a probability for every option; confidence summarizes how concentrated that distribution is.
• A Score answer reports a position on an ordered rubric (it may fall between levels), plus per-level probabilities and confidence.
• A Noul answer reports the probability that a yes/no condition is true; it carries no separate confidence value.

Probabilities guide workflow; they do not guarantee truth. Thresholds (e.g. "flag review when the escalation noul ≥ 0.5") are application values in rules.ts, never model settings.

Batch evaluation

Content Triage supports batches: toggle Batch, enter one message per line (limit 20), and the server runs the same pipeline per item in parallel. Each item carries its own success/failure, so one bad input never poisons the batch. Results are sortable by priority and filterable to review-needed items.

Compare & Choose fans out one request per candidate in parallel and ranks by a weighted composite that application code computes from the normalized Score judgments — deliberately not a "Jev best product" score.

Local setup

npm install
cp .env.example .env.local   # then set TYPESAFE_API_KEY in that file
npm run dev                  # http://localhost:8080

TYPESAFE_API_KEY is read server-side only, at call time, and is never sent to the browser, logs, or source control.

Vercel deployment

1. Import the repository into Vercel.
2. Project Settings → Environment Variables → add TYPESAFE_API_KEY (Preview and/or Production).
3. Redeploy after adding or changing the variable.

Developer view

Toggle Developer view in the header to inspect, per evaluation:

• the exact state and questions sent to Jev
• the raw Jev assessment
• the normalized typed judgment and application thresholds
• model, request ID, TypeSafe latency, total latency, token usage
• error classification on failures

Each block has a copy-to-clipboard control. Secrets never appear here.

Testing

npm run typecheck   # tsc --noEmit
npm run lint        # eslint
npm run test        # node:test over core + per-use-case normalization/rules
npm run build       # production build

Tests are deterministic and require no API credentials: they cover the parse layer (Choice/Score/Noul), each use case's normalization and application rules, threshold behavior, and malformed/partial responses through a mocked answer boundary.

Limitations

• Jev evaluates text states only (no images/audio/video).
• Jev's primary training language is English; other languages have lower accuracy.
• Probabilities are calibrated across prediction groups, not guarantees for any single answer.
• This playground is a demo instrument: it keeps no database and stores no history beyond session statistics in sessionStorage.
• The environment that produced this branch could not delete superseded files, so a few forwarder shims remain under src/lib/jev/ and src/components/ (marked "do not import"); they are inert and safe to delete in a follow-up.