Decisions and risks
| Choice | Reason and remaining limitation |
|---|---|
| Original gate persists | Post-result threshold exploration cannot launder a failed predeclared result into proceed |
| Curated exact assignment balance | Clear quality lesson; not a formal sample-ratio-mismatch test |
| Unpooled normal intervals | Transparent pedagogical approximation; coverage can be poor, especially near boundaries |
| Suppress small-cell intervals | Avoid misleading complaint precision; primary small sample still fails minimum-arm eligibility |
| Guardrail alongside headline | Harm stays visible; thresholds are fictional product choices, not medical or financial guidance |
| No segment winner | Mobile sample is exploratory and small; no multiple-comparison adjustment |
| Local snapshots | Self-contained export; no authentication, collaboration, live experiment connection or durable server record |
| Append-only withdrawal | Mistakes can be marked without deleting evidence; reset explicitly ends history |
| Raw and readability binding | No stale confirmation or writes after failed read; no atomic cross-tab transaction exists in localStorage |
Preview confirmation rereads storage synchronously and compares exact raw bytes and readable status. A same-origin writer could theoretically race after the check; this prototype does not claim multi-user transactional isolation. Storage limits are 100 rules, 100 verdicts and 100 withdrawals; export before deliberate reset.
Open evidence: human comprehension, statistical method suitability for real use, dependent outcomes, missing observations, sequential testing, external validity, and reviewer consistency. No stakeholder discussion is inferred from software tests.