Finding planets
nobody pointed you at

Hand a search program a stack of unlabelled stars, tell it nothing about which ones have planets or how long their years are — will it find them anyway, and will it know when it has found something that only looks like a planet?

A01 folded a light curve at a period it was nudged toward, for a star it was told to examine. Useful as a calibration, but it is not what searching means. A04 takes the hints away: unlabelled light curves, no periods, no indication which star hosts anything.

Three things had to change. The period search widened from a narrow window to half a day through fifteen days. The star's own slow brightening and dimming had to be flattened out first, because stellar variability is far larger than a planet's dip and would swamp it. And candidates could no longer be ranked by how deep their dip was — across different stars that ranking is meaningless, since a noisy star wins by being noisy. Ranking moved to a score that measures each dip against its own star's noise, so the comparison is fair.

It worked. Both planted planets came back ranked first and second, their periods recovered to within a few parts in ten thousand, with the catalogue values read only afterwards to grade the answer. And a third confirmed planet turned up that nobody had designated — flagged blind, put through the vetting tests, and only identified as WASP-20 b when the catalogue was consulted at report time. That one was the vetting proving itself.

Because finding a dip is the easy half. The hard half is that two stars orbiting each other make dips too, and much deeper ones. So every survivor faces tests. Do the odd-numbered dips match the even-numbered ones? A planet's do; a pair of unequal stars alternates deep-shallow-deep-shallow and gives itself away. Is there a second, fainter dip halfway between? That is one star passing behind the other. One candidate — a 4.8 % monster with eleven events — failed the odd-even test and was correctly thrown out. Another was rejected for sitting at exactly 0.5 days, the very edge of the search grid, which is where an artefact lives rather than a planet.

Select the deepest bar on the ladder and press odd vs even. The two folded dips do not line up. That single image explains vetting better than any paragraph, and it is why depth alone is a trap: the deepest thing in the sample is the one you throw away.

The recoveries: HIP 65 A b (TIC 201248411) at P = 0.98124 d, score 14.2, against a published 0.98097 d; and WASP-18 b (TIC 100100827) at P = 0.94164 d, score 10.2, against a published 0.94145 d — A01's target, this time with no hint. The undesignated third, WASP-20 b (TOI 194.01, TIC 211438925), was flagged blind at 4.9014 d against a published 4.89962 d.

The threshold is the part worth dwelling on. Across the sub-threshold targets the measured noise floor topped out at 6.61, with a median of 4.64, and the threshold sits at 8.0. That gap is measured, not chosen — nothing in the sample lives between 6.61 and 8.0, so the line has somewhere real to stand. And three planted test transits at 1.0 %, 0.4 % and 0.2 % depth all came back, at scores of 8.8, 9.2 and 10.1, with recovered depths of 0.92 %, 0.40 % and 0.19 %. The first of those is the detrend bias — quantified rather than asserted.

The graded run searched 26 sector-2 targets over a blind period grid of 0.5 – 15 d in 3,000 steps, with a 0.5-day detrend window. Two rejections carried the vetting: TIC 339607421, 4.78 % deep with eleven events, thrown out by odd-even depth alternation; and TIC 206502540, flagged period-railed at exactly 0.5 d, the search-grid edge, where an artefact lives rather than a planet.