Examples

Emergency snack

emergency_snack.py demonstrates a direct Choice call, structured state, a closed action set, the full distribution, and an application-level confidence policy.

Haunted CI

haunted_ci.py demonstrates a named Score decision, an ordered rubric, fractional scores, a legend, and confidence-aware action.

Connect Four

connect_four.py demonstrates deterministic preprocessing, dynamic Choice criteria, request inspection, and credential-free JSONL tracing. Python computes exact legal moves and removes moves proven to lose under the example’s bounded tactical rules. Jev is called only when multiple tactically admissible actions remain.

Deterministic constraints belong in ordinary code; ambiguous judgment among valid alternatives belongs in Jev.

Read the full source for the playable demo rather than copying its tactical helpers into an application without understanding their scope.

Plain-English semantic linter

semantic_linter.py demonstrates a named bundle plus native async fan-out. Standard-library Python AST parsing owns exact file discovery, syntax errors, function boundaries, qualified names, decorators, and source locations. The semantic-lint bundle contains 14 independent plain-English Noul rules, and all 14 rules are evaluated in one request for each function.

Independent functions are scheduled concurrently with bounded AsyncJev concurrency. The human summary calculates actual function, rule, judgment, and request counts, token usage, and observed elapsed time. JSON output preserves every rule probability and request metadata.

pyjev decision show semantic-lint --config examples/.pyjev.toml --json
pyjev decision validate --config examples/.pyjev.toml
python examples/semantic_linter.py pyjev/client.py
python examples/semantic_linter.py pyjev/client.py --show-all
python examples/semantic_linter.py pyjev/client.py --json

Batch independent semantic questions about one state into one Jev request; use ordinary async concurrency across independent states.

The default threshold is a demo application policy, not a Jev correctness guarantee. Calibrate it against labeled examples before using this linter as a CI gate. The selected function source is sent to the configured Jev API; do not run it on code you are not permitted to send to that service.

Emoji Jev

emoji_jev.py is an interactive named-bundle example. A single request evaluates a 64-way emoji Choice plus tone questions. The terminal displays the selected emoji, top alternatives from the Choice probability distribution, typed side signals, request metadata, token usage, and client-observed round-trip time.

The top-N display is deterministic Python over Jev’s preserved probability distribution; it is not a second model call.