Every candidate carries
its own odds of being nothing

If you search a thousand stars, some of them will show a convincing dip by pure luck — so how do you work out the odds that this particular candidate is just noise wearing a costume?

A04 measured a false-alarm floor for the survey as a whole: across the sample, junk topped out at a score of 6.6, so a threshold of 8.0 sat in a real gap. That is a good answer to the question "how often does this pipeline fool itself?" It is not an answer to "how likely is it that this star fooled it?" — and once you are searching thousands of stars rather than dozens, the second question is the only one that matters.

A05's answer is to make every candidate pay for itself. Take the star's own measurements and shuffle them, destroying any real signal while keeping the noise exactly as noisy as it truly was. Search the shuffled version. Do it 256 times. Now you have 256 examples of the best dip this specific star's noise can produce when there is definitely nothing there, and you can see directly where the real candidate falls among them. That fraction is its false-alarm probability — measured from its own data, not assumed from a table.

Panel one — the null, built in front of you

The shuffling is done two ways, because the naive way cheats. Scrambling every point independently destroys the slow correlated wiggles real stars have, making the noise look tamer than it is and the candidate look better than it is. So a second scheme shuffles in contiguous blocks, keeping those wiggles intact, and the grade is taken from whichever of the two is less flattering.

The panel is worth being honest about here, because it does not show what you would expect it to. The naive story is that the block null should sit visibly further right than the independent one. It does not — the two overlap almost completely, and their peaks land in the same place. The reason is that the score is already normalised against its own period grid, so a change in the overall noise character largely divides out before it can move the distribution. What does survive is a heavier right-hand tail under block shuffling, and since a false-alarm probability is read entirely from that tail, the take-the-worse-of-two rule still earns its place: it costs nothing on the draws where the two schemes agree, and it is the only thing standing there on the draws where they do not. The correction is real and it is smaller than the sales pitch — which is the kind of thing a page about not overclaiming ought to say about its own centrepiece.

Around that sits the rest of a survey. One target in ten is a deliberate placebo, its timing scrambled so that nothing real can survive — if a placebo ever produces a candidate, the calibration is broken and the run says so. A pre-search census kills about a third of every sector before a single period is tried, on the grounds that known binaries, known variables and puffed-up giant stars produce planet-shaped lies at industrial scale. And every survivor gets a disposition from a fixed vocabulary.

Panel two — the ladder, and where it stops

The vocabulary's best possible verdict is lead — awaiting human review. There is no higher state. The machine cannot output the word planet, and the count of discoveries cannot be raised by code at all. That ceiling is deliberate, it is enforced in the receipt contract rather than by convention, and it is the most important thing on this page.

The two pilot runs are why. One candidate held planet-candidate for forty minutes before its own frequency spectrum revealed it as a pulsating star. Another matched a catalogued object whose community verdict is false positive — which is not a recovery and not a lead. Two later leads survived the machine and then died within two minutes of a human checking external catalogues: one flagged as a spectroscopic binary, the others implying planets 2.3 and 2.8 times the radius of Jupiter, which is not a planet, it is a small star. The census exists because of that night.

Panel three — the floor that rises under you

The floor moved from 6.6 at n = 22 to 7.65 at n = 153 to 7.875 at n = 551. It is closing on the 8.0 that A04's much smaller sample made look comfortable. The two-point triage heuristic that was supposed to anticipate the n = 551 floor missed it by 0.47 SDE — in the conservative direction, but a miss, and never graded.

The counting deserves a note of its own, because it is the kind of thing that goes wrong quietly. 1,214 target rows were attempted across the seven full hunts; 1,164 of those searches completed once 41 skips and 9 errors are removed; and they cover 1,174 distinct (sector, TIC) pairs, because 36 targets appear in more than one hunt. The pilot phase holds 570 targets and not 728 — the 570-target receipt declares that it supersedes the 158-target one, and the receipt contract excludes superseded receipts precisely so that cumulative runs cannot double-count themselves. Three sectors are involved: sector 2 on the Windows box, sectors 3 and 30 split to the Linux box.

The two surviving leads are TIC 234518605 at SDE 8.42, flagged in two consecutive sector-2 hunts, and TIC 272357134 at SDE 8.08. Both have dossiers on disk. Both are leads. Neither is a planet, and no code path in this survey can make either one into a planet.

Two named objects carry the lesson forward. TIC 140940493 held planet-candidate for forty minutes until its amplitude spectrum unmasked a δ Scuti-type pulsator. And TIC 278866211, at SDE 10.3, matched TOI 189.01, whose community disposition is false positive — so it carries toi-known-fp and becomes a test target for the blend gates rather than a result.