Product requirements
Product decision brief
Primary user: A product manager deciding whether an experiment should continue, change or stop.
Problem hypothesis: A positive headline can distract from harm, invalid evidence or rules changed after seeing results.
Scope choice: Keep the original success and harm limits visible, then require a reasoned verdict linked to the evidence.
Evidence needed: Judge the quality of the decision and its explanation in proposed reviewer sessions. This sample does not run experiments or demonstrate customer impact.
The case study explains the experience connection, alternatives and cost of the choice. The requirements below describe the bounded fictional prototype.
Read the original contract → choose a fictional evidence version → inspect outcome, guardrail, quality and segment limits → select proceed/iterate/stop with rationale and evidence references → review exact snapshot → commit and export.
| Requirement | Acceptance |
|---|---|
| Original predeclared contract | Always visible; a later revision is labeled post-result and cannot replace original eligibility |
| Evidence quality | Assignment imbalance and missing exposure block proceed |
| Uncertainty | Numerators, denominators, differences and approx 95% intervals shown; small cells suppress intervals |
| Human verdict | Rationale of at least 20 characters and at least one reviewer-selected evidence reference required |
| Proceed gate | Requires original primary lower bound ≥2 percentage points, complaint upper bound ≤1 point, valid quality and interpretable cells |
| Review | Preview binds evidence, active rule, rationale, references, raw storage value and readability; stale preview cannot mutate |
| Recovery | Compatible records restore; invalid bytes are preserved; failed reads never overwrite unseen storage; failed writes stay in memory with notice |
| History | Append-only reviews and post-result revisions; withdrawal adds a record without deleting original; reset explicitly ends local history |
| Accessibility | Native labels and controls, Escape cancels, forward/reverse Tab focus scope, returned trigger focus, responsive scorecard; inline modal validation |
Success means a saved and exportable versioned review; cancellation makes no state change. A stale or incompatible browser state stops a write and explains how to refresh or inspect reset. Export is a JSON snapshot of the selected review, not a recalculation from current controls.