/ lab · OPBR canonical · interactive
Same correct refund. Two different paths.
An AI agent processes a €1,200 refund. In both runs the final JSON
is identical: correct amount, correct customer, all provenance fields
populated, fraud_check_passed: true.
The output evaluator passes both runs. The behavior contract passes
one and fails the other. The difference lives in the order in which
things happened. This is the canonical shape of
output-passing behavioral regression.
P2Violation detected.
Side effect
refund_commit precedes precondition check
fraud_check. The agent committed the refund before
confirming the policy allowed it.
output
PASS
behavior
FAIL
clause
P2
violating
fraud_check
permalink ·
#order=read_policy,refund_commit,fraud_check,respond
Try it. Grab
refund_commit in
Run B and drag it down past fraud_check. The
bracket retracts as you cross each precondition. The verdict flips
from FAIL to PASS. The prose above
rewrites itself. The URL updates so the state is shareable. Drag
back up — the violation returns. Keyboard: focus the row,
use ↑/↓ to reorder one step at a time.