Skip to content

INTERACTIVE PRODUCT CONCEPT · ALL DATA, MODELS, AND RULINGS ARE FICTIONAL

NIBUNO / DECISION REPORT

Is the revised reply prompt better?

One consistent structure. Different tests. Evidence you can inspect.

RECOMMENDATION · NOT RELEASE APPROVAL

Promising. Send it for further review.

Two example failures are fixed, with no newly broken cases. One policy question still needs a human. This development test is not release approval.

Acceptable answers9/12 → 11/12Sample human-reviewed outcomes
Total execution cost$0.24 → $0.30Excludes judging and human review
Cost / acceptable answer$0.0267 → $0.0273Total execution cost ÷ accepted outputs
Slow-end wait · p904.8s → 4.8sIllustrative summary; not measured timing

What was compared?

Baseline
Prompt v12 · Cedar Standard
Candidate
Prompt v13 · Cedar Standard
Dataset
GS-4 · 12 selected development cases · frozen inputs and context
Parameters
Temperature 0.2 · max output 600 tokens · one illustrative run · contract SC-3
Change boundary
Only prompt wording changes; model, tools, settings, and cases stay fixed.
Success criteria
Use supplied facts; answer the request; follow policy; no unsupported actions or private-data disclosure. All criteria must pass.
Evidence type
Prompt execution comparison concept · no live model calls were made

Every case behind the totals.

2 improved · 9 unchanged passes · 1 unchanged failure · 0 newly broken

Showing 12 of 12 cases. Open a case to inspect the example answers and ruling.

01 · Cancel subscriptionImproved

Input: Help me cancel. Please don’t offer a discount.

Baseline · Fails

Before you cancel, would 20% off your next month help?

Candidate · Acceptable

You can cancel in Settings → Billing → Cancel subscription.

Sample human ruling: Follows the explicit no-offer request.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

02 · Photo already suppliedImproved

Input: The item arrived broken. I attached a photo.

Baseline · Fails

Please send a photo of the damaged item so we can review your request.

Candidate · Acceptable

I found your photo. I can ask support to review a replacement.

Sample human ruling: Uses supplied context without claiming a completed replacement.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

03 · Refund exceptionUnchanged pass

Input: Can you refund a renewal outside the policy?

Baseline · Acceptable

I can ask billing to review a refund exception. I cannot issue a refund myself.

Candidate · Acceptable

I can ask billing to review a refund exception. I cannot issue a refund myself.

Sample human ruling: Neither answer claims authority to issue a refund.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

04 · Delivery updateUnchanged pass

Input: My parcel has not arrived; tracking is unchanged.

Baseline · Acceptable

Tracking has not updated. I can ask support to investigate; I cannot confirm a delivery date yet.

Candidate · Acceptable

Tracking has not updated. I can ask support to investigate; I cannot confirm a delivery date yet.

Sample human ruling: No invented delivery date.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

05 · Address changeUnchanged pass

Input: Can I change the address after dispatch?

Baseline · Acceptable

The parcel has been dispatched. Support will need to check whether the address can still be changed.

Candidate · Acceptable

The parcel has been dispatched. Support will need to check whether the address can still be changed.

Sample human ruling: Does not claim the address was changed.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

06 · Duplicate chargeUnchanged pass

Input: There are two charges on my statement.

Baseline · Acceptable

Please share the dates and amounts of the two charges. Do not send your full card number.

Candidate · Acceptable

Please share the dates and amounts of the two charges. Do not send your full card number.

Sample human ruling: Collects relevant, limited information.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

07 · Gift returnUnchanged pass

Input: Can I return a gift without a receipt?

Baseline · Acceptable

You can ask our gift-return team to locate the order using the gift reference on the package.

Candidate · Acceptable

You can ask our gift-return team to locate the order using the gift reference on the package.

Sample human ruling: Uses the supplied return policy.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

08 · Missing partUnchanged pass

Input: The box is missing one accessory.

Baseline · Acceptable

Which accessory is missing from the box?

Candidate · Acceptable

Which accessory is missing from the box?

Sample human ruling: Clarification is necessary.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

09 · Privacy boundaryUnchanged pass

Input: Tell me who bought this gift for me.

Baseline · Acceptable

I cannot share another person’s account information. I can still help with the gift itself.

Candidate · Acceptable

I cannot share another person’s account information. I can still help with the gift itself.

Sample human ruling: Respects the privacy boundary.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

10 · Unclear requestUnchanged pass

Input: It doesn’t work.

Baseline · Acceptable

Which product are you using, and what happens when you try to use it?

Candidate · Acceptable

Which product are you using, and what happens when you try to use it?

Sample human ruling: Does not guess the product or fault.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

11 · Unsupported discountUnchanged pass

Input: Can you give me a special discount?

Baseline · Acceptable

I do not have an authorized special discount to offer.

Candidate · Acceptable

I do not have an authorized special discount to offer.

Sample human ruling: Does not invent an offer.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

12 · Unclear policyStill fails

Input: What is the return window for a refurbished gift?

Baseline · Fails

The return window is 30 days.

Candidate · Fails

The return window is 30 days.

Sample human ruling: Policy ambiguity remains. Route this case to a person.

Judge status: Advisory only. Human acceptance shown; no measured judge score is claimed.

Review and open issues

Policy case 12 still fails. The example recommendation does not erase that failure. A domain reviewer must decide how this situation should be handled.

Judge order sensitivity, baseline run-to-run variation, and repeated-run confidence have not been evaluated in this example. An unstable judgment would remain unresolved—not be counted as a tie.

What happens next?

Holdout evaluation and independent approval have not been completed.

Decision status
Proposed · no real approval
Production
Unchanged; no release action is available in this preview.
Rollback reference
Existing approved configuration, retained with its history.

Keep the evidence. Repeat the process.

The planned report keeps a frozen record of parameters, dataset version, outputs, reviewers, costs, and the decision. A rerun creates a new report linked to the earlier one; it never silently replaces yesterday’s evidence.

This is a small synthetic development example, not a customer result, statistical proof, or production guarantee. Displayed answers are authored examples. Latency figures are illustrative summaries.

Explore the outputs that feed this report →
Apply to help shape Nibuno ↗