Indeterminate

If the error bar crosses the threshold, the honest answer is indeterminate. The report says so, and names the two levels the evidence sits between.

Correct answers over the whole sample (headline_accuracy)
  Grade: B or C, inconclusive
  Score: 0.72, likely range 0.6741 to 0.7617, 400 results
  Reason: The likely range crosses the B threshold (0.7), so the grade is B or C

What happened there

The ladder in examples/scorecard.yaml is walked best-first. A needs a lower bound of 0.9, and the whole interval sits below it, so A is refused. At the B rung, greater_equal_ci_lower: 0.7:

threshold 0.7
        0.6741 ──────●────── 0.7617
                  0.72
                    ↑
                   0.7 is inside the interval

The interval straddles. The descent stops, and the levels below the straddled rung become the floor.

The grade is B or the best level that holds below it. C needs a lower bound of 0.5, which 0.6741 clears, so the grade is B or C, and the evidence does not say which.

Why this is not rounding down

Awarding C would be a claim: that the system does not clear 0.7. The evidence does not support that claim either. A point estimate of 0.72 with a lower bound of 0.674 is entirely consistent with a true rate above 0.7.

The sample backs neither letter. indeterminate says exactly that, and it carries the number that produced it so a reader can decide what to do about it.

The usual thing to do about it is collect more items. That is the useful signal: indeterminate on an indicator you care about tells you the evaluation is too small to decide, before someone builds a decision on it.

In scorecard.json

{
  "id": "headline_accuracy",
  "verdict": "indeterminate",
  "level": null,
  "between": ["B", "C"],
  "rule": {"level": "B", "condition": "greater_equal_ci_lower", "threshold": 0.7},
  "reason": "the likely range crosses the B threshold (0.7), so the grade is B or C",
  "value": 0.72,
  "measured": [{"value": 0.72, "low": 0.6741, "high": 0.7617, "n": 400}]
}

Field	Meaning
level	null. No level was awarded.
between	the level refused and the next one down, in that order
rule	the rule whose boundary the interval straddles
reason	why, in one line, printed in the report

When no lower rule holds

If the straddled rung is the last one that could have held, between carries a single level:

the likely range crosses the unfit threshold (0.3) and no lower rule applies,
so the grade is unfit or none

A ceiling can settle it

If the access tier ceiling sits at or below both ends of the indeterminate range, the range collapses and the verdict becomes graded.

The reasoning: if the grade could not have exceeded the ceiling either way, then the interval was never deciding anything that mattered.

{
  "verdict": "graded",
  "level": "C",
  "uncapped_level": "A",
  "ceiling": "C",
  "ceiling_reason": "access_tier",
  "between": [],
  "reason": "the likely range allows A or C, and black box access is capped at C"
}

uncapped_level keeps the better end of the original range, so the working is still visible.

A ceiling inside the range

A ceiling between the two ends lowers the better end, and the verdict stays indeterminate. The reason carries both steps, so it names the same levels as the grade. On a card with rungs at A (0.9) and C (0.7), a range of A or C under a B ceiling:

{
  "verdict": "indeterminate",
  "between": ["B", "C"],
  "ceiling": "B",
  "ceiling_reason": "access_tier",
  "reason": "the likely range crosses the A threshold (0.9), so the grade is A or C; black box access is capped at B, so the grade is B or C"
}

A ceiling at or above the better end leaves the range as it was, and ceiling stays null. The example at the top of this page is graded under a B ceiling for that reason.

indeterminate is not ungraded

Three different things:

Verdict	Means
graded	a level was awarded
indeterminate	measured, and the evidence does not separate two levels
ungraded	no rule held at all, or the number was not in the bundle

ungraded reads as "not assessed". indeterminate reads as "assessed, and the assessment does not resolve". Collapsing them would hide the second inside the first, and the second is the one that tells you to collect more data.