Make your first local policy decision

Status: draft.

You will construct an inclusive threshold and see both a passing and failing decision. The answers are synthetic, so this exercise needs no API key or service access. It teaches local comparisons, not model quality.

Use the environment from first judgment. If you start here, follow only that installation section. You do not need to supply a key or run its service example. Package installation needs network access; the program below runs offline.

Define acceptance separately from the question

The question remains “Is this about billing?” A policy adds the application rule: accept a Noul probability of at least 0.8. This is a teaching threshold, not a recommended threshold for your tickets.

Save this complete program as first_policy.py in your tutorial directory:

from judgevet import Noul, NoulAnswer
from judgevet.policy import NoulRule, Policy, evaluate_policy, validate_policy

questions = {"billing": Noul(instructions="Is this about billing?")}
policy = validate_policy(
    Policy((NoulRule("billing", minimum=0.8),)),
    questions,
)

at_boundary = {"billing": NoulAnswer(noul=0.8)}
below_boundary = {"billing": NoulAnswer(noul=0.79)}

passing = evaluate_policy(policy, at_boundary)
failing = evaluate_policy(policy, below_boundary)

print(f"At 0.8: passed={passing.passed}")
print(f"At 0.79: passed={failing.passed}")
print(f"Rule checked: {failing.rules[0].question}")
assert passing.passed
assert not failing.passed

validate_policy checks the rule against the named question and records its constraints. evaluate_policy compares the supplied typed answer with that validated rule. Neither operation sends anything to the service. These are judgevet's local policy operations.

Run the comparison

.venv/bin/python first_policy.py

Expected output for these fixed synthetic inputs:

At 0.8: passed=True
At 0.79: passed=False
Rule checked: billing

Checkpoint: exactly 0.8 passes. The minimum is inclusive. The lower value fails, and the report still identifies the rule. This is an ordinary return value, not an exception. Both Noul answers are valid numeric answers.

The Python process itself exits successfully after printing both decisions. Your application chooses what to do with passed. The CLI's policy exit codes are a separate adapter behavior. A false report does not automatically set a Python program's exit status.

Distinguish failure from an unavailable decision

An unmet threshold means the policy was evaluated and did not pass. A missing required answer or a wrong answer type instead raises PolicyAnswerError. An invalid policy definition raises PolicyDefinitionError. See policy error handling.

If your program reports a policy-definition error, check that the rule name matches the question. Check that the threshold is a finite number from zero to one. If judgevet cannot be imported, run the file with the tutorial environment's interpreter. Do not add service credentials to fix this offline example.

A service failure happens before a usable answer reaches this comparison. Do not convert every exception into a failed policy: that would hide the difference between rejected content and a broken request.

Connect the two exercises

The first tutorial obtains answers from Jev. This tutorial constructs them locally so you can observe the comparison at an exact boundary. In a real application, pass the service call's response.answers to evaluate_policy using a policy validated against the same question definitions.

Follow the complete synchronous and asynchronous workflows to combine the steps. Before choosing production thresholds, read policy tradeoffs and labeled evaluation and what verification proves. The compatibility reference describes the supported imports and public policy contract.