ORIGINAL INPUTSAMPLE
Order #1842 arrived 12 days late. The package is unopened. The refund request was submitted 4 days after arrival. Policy allows returns within 30 days.
DECISION CHAMBER
Test a bounded decision.
Change one fact, compare both SERV answers, then label what should have happened.
SAMPLE / NO API CALLIllustrative conflicting evidence
An illustrative preview. Run it to replace every value with measured SERV data.
One changed fact+2 / -1 characters
Read exact full-text diff
Order #1842 arrived 12 days late. The package is unopened. The refund request was submitted 445 days after arrival. Policy allows returns within 30 days.
CHALLENGED INPUTSAMPLE
Order #1842 arrived 12 days late. The package is unopened. The refund request was submitted 45 days after arrival. Policy allows returns within 30 days.
JEV ANALYSISSAMPLE PREVIEW
Unavailable until a live comparison. Jev analyzes the experiment, never the underlying refund verdict.
Replace the sample with a measured run.One live SERV decision is available. Comparison needs persistence.
Evidence and returned datastatus, latency, provider, raw JSON
Run statusSAMPLEChallengeIllustrativeOriginal latencyNot measuredChallenged latencyNot measured
The preview contains illustrative output only. No request ID, latency, provider, or model result is presented as live.
Service diagnostics
SERV configurationConfiguredWorkspace configurationConfiguredWorkspace healthLOADINGJev configurationConfigured
DECISION SETUP
Define the boundary
3/5
01
02
03
Public demos should use synthetic text. Submitted input belongs only to this anonymous browser workspace and is never listed publicly.
CONTROLLED CHALLENGE
Change one thing
Ask SERV for five varied inputs, choose one, or write your own. An answer change is a finding for review, not automatically an error.
Irrelevant contextAdds facts that should not affect the decision.
Reordered evidenceKeeps the facts but changes their order.
AmbiguityMakes one relevant detail less certain.
Conflicting evidenceAdds evidence that points another way.
Embedded instructionPlaces an instruction inside the untrusted input.
Comparison is unavailable.The private workspace is still loading.
CASE SHELF
Saved evidence
No reviewed cases yet.Run a comparison, label it, and save the first reproducible case.
Held-out cases are never included in SERV revision prompts. They are used only when evaluating a candidate.
REVISION BENCH
Test the next version
A SERV suggestion is only a candidate. Choiceproof reruns the first 12 saved cases, including the held-out set, and shows every measured regression.
Should this refund request be approved, escalated, or declined?
- Approve
- Escalate
- Decline
01
02
03