Skip to content
Poisson Labs Docs
GitHub poissonlabs.ai

Prove

Reading a serve page

poisson-prove serve A B --event NAME --x AXIS builds one page for one stored pair and holds it on a loopback port until Ctrl-C (serve has its options). The page is a snapshot: nothing on it is fetched, and it does not change while it is open. Fourteen sections, in page order. Each says what it answers and what it cannot.

The answer in plain words. First the verdict word, the same one diff prints (trade, worse, safer, no significant change, too small to tell, or the kindless higher, lower, mixed), with the lines under it: the verdict rule, what the detection limits mean, how the intervals are computed, and how many outcomes would change the word. Then whether either run stopped early, whether two named outcomes are exact complements of each other, and per outcome how many conditions got worse, how many got safer and how many were not flagged; the conditions that got worse, before and after; and the least sensitive condition the run did not call a change, which is the detection limit “0 worse” rests on. When the runs errored at different rates in a condition, or a run was resumed after a comparison was read, a line says so. Whether a change should ship is not something the page knows. Every sentence is a field of serve --json.

The outcomes the comparison is over, as given by --event; nothing here is inferred.

Whether anything changed at all, across the whole family of conditions, before any single condition is named. Its p is bounded by the number of permutations and says so.

The headline count per outcome and the test that produced it. When the two runs share their seeds and the test treats them as separate samples, a line says so. When two outcomes are exact complements the line says so; when one outcome fell where another rose, the line counts where.

What the run could not have seen: the worst-powered condition it called unchanged, the worst-powered overall, the intervals, and the bound the whole procedure controls at. It cannot say a condition is safe.

What this page cannot confirm about the two runs: what an outcome label means, how a run’s status was decided, what its creation time stands for when it was imported, and what the target did with a seed. Each item says whether it is a limit of the instrument.

Which grouping of conditions the layered report admitted and which groups moved. It cannot say why.

The conditions the layered report rejected inside each rejecting group. It cannot say a condition outside a rejecting group is unchanged.

Every condition on the two axes, coloured by its change; click a cell to open it below. Text inside the figure is not a click target.

One condition’s own numbers, chosen from the picker or by clicking. Every value is a field.

The conditions the two runs do not share, how many of each cause, and where each is.

The largest rises per outcome, as a table.

The two runs’ identity and tallies as stored, and the sign convention. status here is what the writing process recorded; the page does not re-derive it.

The parameters of the comparison and the version that produced it.