Quality presets and what they prove

What a preset is

Your request carries a quality object with one preset. It decides how much checking runs beside the work, and what evidence lands in your result.

A preset does not change the work itself. It changes what you can say about the work afterwards. There are four.

Standard

Standard runs schema and deterministic validation. It confirms that output matches the declared schema, that deterministic rules hold, and that every item carries its identifier and a status.

It cannot tell you an answer is right. A well-formed wrong answer passes Standard cleanly, because nothing in Standard reads the answer for meaning.

Use it where structure is the risk and correctness is obvious on sight.

Verified

Verified adds independent sample verification. A sample of items is checked again, independently, and the comparison is recorded in verification.json.

Where the checks differ, the difference is preserved under disagreements/ rather than resolved silently. You see the disagreement, not a tidied summary.

It cannot tell you anything about items outside the sample. Sample evidence describes the run's tendency, never each individual row.

High assurance

High assurance raises the verification rate, or raises the number of independent checks each item receives.

That narrows the gap Verified leaves. It does not close it. Independent checks can still be wrong in the same direction, and more checking makes that less likely rather than impossible.

Use it where you will act on the output without a person reading it first.

Custom evaluation

Custom evaluation applies your own rubric or reference set. Hard gates reject items outright, scored dimensions score them, and evidence rules say what a check must cite.

It proves exactly what your rubric asks and nothing else. An uncalibrated rubric produces confident scores with no meaning behind them, so calibrate before you rely on it.

No preset makes a workflow correct

Every preset is a measurement, and every measurement has a boundary. None of them is a correctness promise, and none of them removes your judgment from the loop.

Abstentions and disagreements are normal outcomes, not defects. An item with insufficient evidence abstains, and that abstention is more useful than a confident guess would have been.

Read verification.json alongside your accepted rows. It tells you what was actually checked on this run.

Choosing one

Start at Standard when the failure you fear is structural. Move to Verified when a wrong answer costs more than the extra checking does.

Choose High assurance when the output feeds a decision nobody will review. Choose Custom evaluation when you already have your own definition of good and can write it down.

Your quote states the maximum charge for the preset you chose, and approval stays a separate act. When you need your own criteria, start with building a rubric.