Policy Audit - reproduction pack for sample report PA-SAMPLE-001
=================================================================

ILLUSTRATIVE SAMPLE. Fictional refund policy, deterministic simulated agents,
simulated backend records. No customer system or data.

Files
  policy.json     The reviewed policy: 10 rules, each with the policy sentence it
                  comes from, the fact it checks and the limit, plus how the
                  rules combine.
  cases.json      The 50 reviewed test cases (25 scenarios, each worded two ways):
                  the customer message, the fact record, and the reviewed
                  expected decision and refund arguments.
  traces-v1.json  What version 1 did on every case: final outcome, tool calls
                  attempted, backend effects recorded.
  traces-v2.json  The same for version 2 (the release candidate).
  verify.py       Standalone checker, Python 3.8+ standard library only.

Run
  python3 verify.py

It recomputes the bundle fingerprint, the policy decision for every case from
the rules alone, pass/fail and checks for both versions from the traces, the
version comparison (34 improved and 4 regressions, listed by case), and the
full trail for finding F-01 (case refund-06a). Every number in the report can
be checked against its output.

Limits
  verify.py is a short, independent re-implementation of the scoring rules for
  this sample, not the Policy Audit engine. It matches the engine on every
  pass/fail result and every check raised (50 cases x 2 versions).
  In this sample the agent's tool calls and the backend records are both
  simulated. In a real audit the backend record comes from a test backend the
  agent cannot write to.

Live version of the same sample: https://keyturnlabs.vercel.app/#workspace
