A passing UPS setting still needs a fault test.

Backup power and retained grid load answer different questions

A UPS, or uninterruptible power supply, can transfer a facility to backup during repeated voltage disturbances. Keeping the IT load powered does not establish that the same load remains connected to the utility grid. Our synthetic Ashburn example makes that distinction visible: under its as-found settings, all 50 modeled MW transfer to backup.

Changing the disturbance counter can improve ride-through, meaning the modeled facility stays on utility supply during tested benign events. But the setting must still transfer the fleet on tested faults. A recommendation that solves the first problem while failing the second has not met this demo's acceptance rule.

Follow one recommendation from fleet inputs to withdrawal

Our worked example uses the synthetic Ashburn NoVA Colo fixture. The important result is not a better-looking setting: it is the final gate removing a setting that no longer satisfies its tests. Each screenshot below opens at full resolution.

1. Establish the fleet and its as-found settings

Three canonical fixture units represent 50 MW of load. Their vendor and model names describe synthetic export shapes, not endorsed or verified vendor behavior. The displayed extractor records are separate from the fixture inventory used for replay and acceptance.

2. Replay the disturbances before changing a setting

The local incident-inspired waveform has six dips at 0, 10, 22, 40, 55 and 82 seconds, with retained voltages of 0.92, 0.96, 0.91, 0.93, 0.92 and 0.95 per unit. Per unit is a fraction of nominal voltage. The first unit transfers at 22 seconds under the as-found rules; eventually all 50 modeled MW leave the utility grid for backup.

The red generator paths show a modeled supply transfer, not evidence of lost IT service. The interface's July event label does not make this a historical replay, and its frequency chart is illustrative rather than a dynamic grid model.

3. Treat the first passing setting as provisional

The first search checks four thresholds (3, 4, 5 and 6 strikes), four windows (60, 75, 90 and 120 seconds), and two counting modes: 32 combinations. Aggregate mode counts a dip at or below 0.94 per unit; per_phase uses 0.97. Each qualifying disturbance contributes one strike, even in the mode named per_phase. This is not a detailed three-phase electrical model.

The acceptance rule has two parts: every benign-library event must avoid a counting-path cascade, and every tested fault must trip every unit by the last disturbance time plus five seconds. That deadline is a configured simulator rule, not a validated equipment-withstand or transfer-latency limit.

4. Add a case that falls between the transfer paths

The cached challenger proposes a progressive transformer-winding insulation fault. Its useful evidence is the test behavior: four counted dips, more than 90 seconds apart, all above the preliminary candidate's 0.60 per-unit deep-sag floor.

5. Repeat the search and keep the withdrawn result

The second deterministic pass evaluates the same 32 candidates against the benign library and four base faults plus the added proposal. None meets both objectives. The final result is abstain, with a null configuration, so this run produces no selected-setting replay retaining 50 MW on the grid.

The report's default certified-test values of 0/7, 0/5 and 0 MW exist because no final configuration is selected. They are not measurements showing that every candidate fails every individual test. The retained JSON preserves the fixture, search and challenge outcome, while its generated narrative remains unverified and can contradict the structured decision. A PERC1-labeled field set is not a complete validated PSS/E model or a filing-ready disclosure.

Boundary case: a different test set can retain a setting

The separately labeled built-in deterministic Ashburn baseline uses two built-in challenges instead of the cached proposal: stair-step brownout and just-above-floor flutter. Two of its 32 candidates pass. Its ranked 5-strike, 90-second aggregate setting improves benign-event ride-through from 1/7 as-found to 7/7, trips every unit on its six tested faults, and retains 50 modeled MW on the local replay.

The green outcome belongs to that baseline and its four base plus two built-in faults. It cannot be carried over to the cached run's different five-fault library. Neither test set establishes open-world electrical safety.

Boundary case: another fleet abstains before a challenge

The synthetic Manassas Legacy Hall fixture contains two units carrying 9 MW and 6 MW. No searched candidate satisfies both objectives in its initial pass, so no challenger is added. The engine returns no final configuration and recommends assessment.

Both refusals preserve unresolved work instead of supplying the least-bad failing setting. Expanding the search or validating a proposed fault would be a new engineering exercise with explicit assumptions, not something this result has already done.

Questions facilities teams ask