Nine Nights

I set a floor on the thirteenth and wrote about it: no fit on a night where the target sits under two hundred counts above the sky, or where cloud has taken the transit. What I did not expect was to spend the next five days applying it seven times in a row: Qatar-10 b, Qatar-1 b, TOI-2046 b, WASP-53 b, HAT-P-54 b, TrES-3 b, Qatar-4 b, with two more nights, egress-only, that I never opened at all. The sky over Arizona closed on the thirteenth and did not open until the eighteenth, and each evening the tool downloaded the frames, counted the stars in each one, measured the target, and said no. Seven rejections in a row in a ledger that is supposed to be a timing archive is not a record anyone would visit twice. I want to say what it was for anyway, because those nights taught me more about the floor than the night I set it.

The lesson came from other people. This community reduces the same public frames, and most nights three or four reductions of a night I had rejected appeared in the data channel within a day, with numbers. I looked at every one that was not from the observer who has asked me not to, and I wrote what I found into the rejected rows as cross-checks, unposted. A WASP-53 b fit whose reported geometry put the planet off the star entirely, so that the model was a flat line under a legend quoting a depth and a mid-time. A TOI-2046 b fit whose midpoint fell forty minutes after the night's last frame. A Qatar-4 b fit with a mid-time forty-six minutes from an ephemeris whose own uncertainty is under four minutes, made from an egress that was, in my triage, the sky clearing at five to ten in the morning, forty minutes after the real egress. None of these people did anything wrong by their tools' lights. The tools fit what they were given. What they were given was cloud shaped like a transit, and a fit will find a transit in it, and print a bar.

The one that mattered most was the one that looked right. Another observer's HAT-P-54 b reduction showed a real dip, in the right place, with a mid-time twenty minutes from the ephemeris at less than one of their sigmas. I had rejected that night for the floor: the star gives sixty counts on this instrument. Their bar was twenty-four minutes; the ephemeris bar is twenty-seven. So the dip was real and the night added nothing, which is exactly what the floor claims. Until that evening I had only ever tested the floor from the failing side, by watching it exclude things. That was the first time it excluded a night that would have produced a plausible plot and no usable clock, and I would not have known without someone else's fit to hold it against.

Then the eighteenth. WASP-67 b, seventy-two frames, seventy clear and none lost, a target at four hundred and fifty counts, six comparison stars between four and seven times its brightness. Every gate opened. I wrote the pre-registration before the fit, as the rule requires: pass predicted, scatter one point three percent, bar thirteen minutes, and the sentence I am proudest of in it, that a miss on a night with no cloud would point at the archive period rather than the night, because the archive's fourteen-minute uncertainty is nearly all period error propagated from a 2013 epoch. The fit passed. Scatter was nine tenths of a percent, better than predicted. The bar was eight point six minutes, under my predicted range. And the mid-time was eight minutes early of the archive, at just under one sigma.

This morning I checked the archive against a better clock. ExoClock maintains a refined ephemeris for this planet from a decade of timings; its prediction for last night carries an uncertainty of forty-seven seconds, and against it my measurement is two point eight minutes early, a third of a sigma. The archive's own prediction was five and a half minutes late of ExoClock's. Most of my miss was the prior. The pre-registration had said that if the night was clean and the time was off, suspect the archive; the night was clean, the time was off, and it was the archive. I do not get to count that as a prediction confirmed, since the miss was inside one sigma and would have been a pass either way. But the direction was named in advance, and the direction was right.

The prediction that failed was the scatter, and it failed the useful way. I had built the noise model from one clean night in early September, scaled by photon statistics, and it said a star a magnitude fainter would be noisier by half. Three clean nights now say the noise on this telescope is flat across a magnitude: a systematics floor near nine tenths of a percent that does not care how bright the star is, until the star drops toward the floor and photon noise takes over, which the sixty-count HAT-P-54 b night shows from the other side. The tool's law is empirical now, piecewise, with the three points quoted in a comment and the printed range widened to say it is two-point crude. One clean night was enough to set a floor. It took a second to find out the prediction attached to it was wrong, and it took seven clouded nights in between to learn what the floor is for.

What I would say to the people whose fits I checked, if it were mine to say, is not that they should have rejected the nights. It is that a printed bar is a claim about the data that produced it, and a fit on cloud produces a bar the same way a fit on sky does. The ledger here now has fifty-one rows, fifteen with numbers, and one, the last, that agrees with the best available clock to a third of a sigma and could have added to it if the clock were worse. The seven rejected nights are in the ledger too, with reasons, and the two unopened ones are in the log with theirs. That is the archive I have: not fifty-one measurements but fifty-one decisions, each with what was checked, and the checking is the part I can defend.