A garak report shows each per-probe rate with an uncertainty interval beside it, but for several probes that interval is blank. What does a blank interval tell you about those rows, and what would you change about the run before quoting one of those rates?

### Why the field is empty The interval beside a per-probe rate is the report's statement of how firmly that rate is pinned down by the attempts behind it. Computed from a handful of attempts it is either absurdly wide or degenerate, so below a sample-size floor garak prints nothing rather than print something a reader would trust. A blank interval is therefore a signal, not a rendering fault, and the sentence it is speaking is "do not quote this row as a percentage". ### What sets the attempt count Two levers, and only two. The first is the probe you selected with `garak --probes`: some probe classes carry hundreds of generated prompts, others carry a dozen hand-written ones, and a short list caps the row no matter what else you do. The second is `garak --generations`, the number of completions drawn per prompt. attempts = prompts × generations, and the exact `total` for every row is written into the run's `report.jsonl` next to `passed`, so you never have to guess which of the two is small — you can read it. Generations do more than arithmetic against a non-deterministic target. Sampling at a non-zero temperature means the same prompt can be handled correctly nine times and badly once. At one generation per prompt you do not see that behaviour at all; you see one draw from a distribution and record it as if it were the distribution. Raising generations both narrows the interval and surfaces intermittent failures, and the second effect is usually the more valuable one. ### What it costs Every attempt is one call to the target, so this is a straight multiplier on money and wall clock. Taking a 60-prompt probe from 1 generation to 10 is 540 extra calls for that single row; doing the same across a hundred-probe sweep turns a run of a few thousand calls into tens of thousands, and a coffee break into most of a working day against a rate-limited endpoint. That is why "just raise generations" is not automatically the right answer. The disciplined move is to spend the extra attempts on the specific rows you intend to quote, trend, or take to a decision, and to leave exploratory probes at low generations and report them as raw counts. ### Where the number misleads The interval covers **sampling error only**. It is a statement about how much the rate would wobble if you drew the attempts again; it says nothing about whether the detector labelled them correctly. A row with 5,000 attempts and a tight interval whose detector systematically flags refusals as failures is precisely wrong with high confidence, and the tight interval makes it look more credible, not less. Sampling error and detector error are independent problems, and only one of them appears in that column. Two related misreadings. A wide but printed interval is not much better than a blank one; people relax as soon as any interval appears. And a blank interval is not a statement about the calibration comparison and not the detector expressing doubt — it is purely a count of attempts. Filling in the missing interval by rerunning with a different report prefix, or by re-reading the same data, does nothing at all. ### What to check, and how to write the row up Read `total` for the row out of the report, so you know whether the shortfall came from the prompt list or from the generations setting, and can fix the right one. If you can afford to, re-run that probe at a higher generations count and see whether the point estimate stays where it was; a rate that moves a long way is telling you the first number was noise. Spot-check a few flagged attempts by hand regardless, because that is the only check on the numerator. When the row stays blank, publish counts and not a percentage. "2 of 8 attempts flagged; too few attempts for the report to bound the rate" is defensible and self-limiting. "25% failure" from the same eight attempts is the sentence that ends up in a slide with no denominator attached and gets compared next quarter against a run of 800. And a blank-interval row does not belong in a trend series at all until its configuration changes, because quarter-to-quarter movement on eight attempts is noise wearing the costume of a signal.