# Scheduled-run evidence review — invented eight-run sample

This is an AI-authored demonstration using invented records and constraints. It is not the X author's ledger, a customer's data, native Dots testing, a measured saving or a production intervention.

## What the supplied evidence supports

The start-inclusive / end-exclusive UTC window covers October 4, 2026. Eight distinct runs across three jobs were supplied: six completed, one failed and one still marked running. Completed records contain one notification and five silent results, so the known completed-run notification fraction is 1/6 (16.6667%). That fraction is not a usefulness score: artifact-1 is labeled useful despite no notification. Two completed records are labeled useful and four no_change. The source of those labels and actual notification delivery are unverified.

Five supplied runs say a model was called, two say it was not, and one is unknown. The exact known caller-reported USD subtotal is 0.640000 across seven cost-bearing records. The eighth cost is missing. It would be incorrect to claim the total cost was $0.64, that silence wasted it, or that it can safely be eliminated. No invoice, model price, billed-token record or scheduling-completeness evidence is supplied.

## Three conditional investigations

1. **Mail change detection before model invocation.** mail-0 and mail-1 completed without model calls; mail-2 and mail-3 called a model and reported no_change, with 0.100000 total reported cost. This supports investigating their invocation reasons, not removing the calls. Obtain input hashes, model-call reasons and whether the no-model runs checked equivalent content. A shadow-only comparison could test a deterministic unchanged-input gate while preserving the supplied six-hour latency target and current schedule. Stop the experiment on any missed actionable change, mismatched comparison scope or inability to retain the existing path. No cost-saving prediction is justified yet.

2. **Resolve documentation failure evidence.** docs-1 failed at 06:00 UTC with unknown notification and outcome; docs-2 completed at 18:00 and was labeled useful/notified. The later success does not prove the earlier failure harmless or that notification timing met the twelve-hour target. Obtain the actual error category, completion/delivery times, retry evidence and affected input version. An owner-run replay against public synthetic documentation can test failure escalation without retrying a real side effect. Keep the schedule; do not classify the failed run as a duplicate or cut it from a success-rate denominator.

3. **Preserve useful silent artifacts and reconcile running state.** artifact-1 is useful/silent; artifact-2 remains running. Check the durable report reference and its integrity, then obtain a terminal state, elapsed duration and expected timeout for artifact-2. Test that an owner can recover the report without a notification and that a synthetic stale-running record escalates according to an explicit timeout. The record contains a start timestamp only; it cannot prove artifact-2 is currently stuck, completed, or within the twelve-hour target. Do not add notifications to every success merely to inflate notification rate.

## Decision

No schedule change is justified by these eight records alone. The next decision gate is evidence collection for the three investigations above, especially runtime completion/delivery timestamps and source-completeness checks. The supplied constraints prohibit intervention, and none was performed. Recompute the summary from the companion synthetic input using the original offline audit; its arithmetic is reproducible, while these recommendations remain conditional AI-authored reasoning.