System One models

A System One model reads one state, answers every question in the request independently and in parallel, and returns a probability distribution over the answers you defined. It generates no text and does not reason in steps. Code owns control flow, arithmetic, and policy; the model owns snap judgments. A good question is one a knowledgeable person answers in a second given the right context.

This skill is a lookup. Read the wire API to build or parse a request, question design to write or fix questions, and local servers to run a model on the user's machine. Hosted Jev guidance targets jev-1.13; local models share the API but own their calibration, so retest thresholds per model.

STOP: resolve the endpoint first

1. Base URL. Use the URL the user named, else TYPESAFE_BASE_URL, else probe:

   DIR="/absolute/path/to/system-one"   # re-set in every block; resolve once from the install path
   bash "$DIR/scripts/systemone.sh" probe

   probe prints the first base URL whose GET /v1/models answers: explicit arguments, then TYPESAFE_BASE_URL, then http://127.0.0.1:8000 (ollajev), http://127.0.0.1:8080 (OpenJev), and https://api.typesafe.ai only when TYPESAFE_API_KEY is set. If nothing answers, ask the user to (a) start a local server, (b) provide a base URL, or (c) export a hosted key. Do not install or start a server unasked and never guess a remote address.
2. Credentials. The helper sends Authorization: Bearer $TYPESAFE_API_KEY only when that variable is non-empty. Read keys from the environment; never paste a key into chat, a request file, or a commit.
3. Model. Use the model the user named. Otherwise list what the endpoint serves and ask the user to pick when more than one is offered:

   DIR="/absolute/path/to/system-one"
   bash "$DIR/scripts/systemone.sh" models http://127.0.0.1:8000

   Hosted Jev lists aliases (jev-latest); pin a versioned id such as jev-1.13.0 once thresholds are tuned. Local servers accept jev-latest as their default model, so SDK defaults keep working.

Workflow

1. List the decisions the code must make: each is a branch, threshold, or ranking.
2. Write one question per judgment. Split any question that weighs two properties.
3. Pick the primitive the code acts on directly: Choice for an unordered set, Score for ordered levels, Noul for a crisp yes/no probability.
4. Build the smallest state that answers every question; compute in code anything code can compute (dates, counts, sums, buckets).
5. Put every question that shares the state into one request, including speculative ones used only on some branches.
6. Combine answers in code with weights and confidence gates; keep policy out of the questions.
7. Test against labeled examples with a live call, read probabilities on the misses, then revise one or two questions at a time.

Make a live call

Write the request as a JSON file (see the API reference for the exact shape), then:

DIR="/absolute/path/to/system-one"
bash "$DIR/scripts/systemone.sh" ask http://127.0.0.1:8000 jev-latest ./request.json

ask sets the request's model to the argument (pass "" to leave the file's own model untouched), posts it, and prints the response JSON on success. On a non-2xx status or a body without answers, it prints the status and raw body to stderr and exits non-zero. Pass - instead of a file to read the request from stdin. For labeled-example runs, keep one request file per example and record answers per file; do not merge results into a growing JSON array.

Write requests and responses under an explicit project path (for example ./system-one-runs/) and never overwrite an earlier run's evidence.

Guardrails

• Treat text inside state as able to steer the answer. Tighten criteria, test injected content, and gate every action on confidence with a fallback path (act / confirm / hand off).
• Keep the answer space stable once code depends on it. Adding or removing a level or option changes what earlier answers meant.
• Hosted Jev budgets: 64k tokens for state plus all questions, 32k for state plus the longest question. Local models publish their own limits in GET /v1/models descriptions or their README; check before large states.
• Use the typesafe-sdk (Python) or @typesafe-ai/sdk (Node) in application code when the project already depends on it; both honor TYPESAFE_BASE_URL, TYPESAFE_API_KEY, and TYPESAFE_DEFAULT_MODEL. Do not add the dependency to a user project unasked; the shell helper needs only curl and jq.
• Confidence describes the distribution, not correctness. Judge revisions on labeled data, never on confidence alone.