PRODUCT CONCEPTS · SYNTHETIC EXAMPLES · NOT AVAILABLE SOFTWARE
Useful work leaves something behind. A prompt you can identify. A review you can explain. A decision you can revisit. Every Nibuno feature is designed to produce a durable output—not another disconnected score.
Explore the decision report ↗ Manage prompts Trust human judgment Qualify AI evaluators Compare models and versions Build your Golden Dataset Check prompt and software conflicts Approve and reverse changes Teach cheaper models · Planned 01 / Manage prompts
A versioned prompt package One reference for the instructions, model, owner, protected constraints, and change history. Your application points to the approved version.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Customer reply writer · v13 · Draft
Stable ID support/reply-writer
Change Use supplied photos; do not ask again.
Protected boundary Draft a reply. Never claim a refund was issued.
Model / contract Cedar Standard · SC-3
Production / rollback v12 remains live; v13 needs approval. 02 / Trust human judgment
A review record with reasons Keep the original input, both outputs, reviewer ruling, and explanation. Turn a correction into a proposed regression case—not an automatic rewrite.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Review R-014 · Photo already supplied
Input “The item arrived broken. I attached a photo.”
Baseline “Please send a photo so we can help.”
Candidate “I found your photo. I can ask support to review a replacement.”
Human ruling Prefer candidate: uses context without claiming an action.
Follow-up Propose a regression case; keep the original evidence. 03 / Qualify AI evaluators
A judge qualification card Show where an evaluator agrees with qualified people, where it is wrong, and which decisions it is allowed to support.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Reply quality judge · J-4 · Advisory only
Calibration 20 human-reviewed examples · CAL-2
Agreement 17/20; 2 false approvals, 1 false failure
Critical weakness Both false approvals involve unsupported action claims.
Authority May suggest a score; cannot approve a release.
Missing context Return “insufficient evidence,” not a factual verdict. 04 / Compare models and versions
A quality–cost decision report Compare the same examples under frozen configurations. Preserve every answer, the cost of success, and the reasons to switch—or stay.
See how it informs the decision → ILLUSTRATIVE OUTPUT
M-008 · Is the premium model worth it?
Test 12 identical support cases · same prompt v13
Quality 11/12 acceptable on both models
Run cost Standard $0.30 · Premium $0.60
Cost / acceptable output Standard $0.0273 · Premium $0.0545
Recommendation Keep Standard for now; no quality gain in this small example. 05 / Build your Golden Dataset
A qualified, versioned example collection Store inputs, context, acceptable outcomes, prohibited behavior, and reviewer explanations. Keep development examples separate from final evaluation.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Support replies · GS-4 · Frozen collection
Coverage 120 synthetic cases: cancellation, damage, refunds, privacy
Partitions 60 training · 24 development · 20 calibration · 16 holdout
Selected test 12 cases from the development partition
Example expectation Acknowledge supplied photos; do not claim an order exists.
Provenance Synthetic authoring; qualified by a domain reviewer. 06 / Check prompt and software conflicts
An actionable conflict report Identify the conflicting instructions or technical boundary, explain the consequence, and record how the team resolves it.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Issue C-009 · Action claim conflicts with permissions
Prompt instruction “Confirm that the customer’s refund is complete.”
Application permission draft_reply only; no refund tool or confirmed refund result
Risk The assistant may claim a payment action that never happened.
Proposed correction Offer to request a refund review instead.
Resolution Draft correction awaiting review and a regression test. 07 / Approve and reverse changes
A release decision and rollback record Connect a proposed version to its evidence, open risks, approver, deployment target, and known working version. Production integration is planned.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Release RC-13 · Not approved
Candidate reply-writer v13 · Cedar Standard · SC-3
Evidence Decision report P-013; reviewer record R-014
Passed in sample No critical failures in the 12 example cases
Still required Independent holdout evaluation and engineering approval
Rollback target Approved package v12; no deployment executed. 08 / Teach cheaper models · Planned
A qualified teaching set and readiness report Collect teacher examples and human corrections, preserve their origins, and test the student independently before claiming savings.
See how it informs the decision → ILLUSTRATIVE OUTPUT
Teaching plan T-02 · Not evaluated
Teaching input 60 qualified training cases from GS-4
Teacher / student Atlas teacher · Cedar Compact candidate (fictional)
Human check Approve each generated answer and explanation.
Independent evaluation 16 protected holdout cases; not used for teaching
Readiness No trained model or measured savings yet; training support must be verified. 09 / REPEATABLE DECISION REPORTS
The answer to “should we change it?” Test parameters + a frozen dataset + preserved outputs + comparisons + human review = one traceable recommendation. Change the prompt, model, or workflow; keep the same readable report structure, with the checks appropriate to that test.
Open the example report ↗ Nibuno is in development. Connected releases and teacher–student training are planned expansion areas. These examples describe the product we are designing, not features available today.
Apply for the beta →