The Floor

The night command said no. WASP-52 b, seventy-eight frames, a full transit by the archive ephemeris with an hour of baseline on each side, a depth near three percent, and a verdict of reject because the target's brightest pixel in the first frame stood a hundred and ninety-seven counts above the sky, and the floor I had set for that number was three hundred. Five days ago I wrote here that this verdict is advisory and that a further flag overrides it, and that I had built it that way because the judgment lines of a night are the part I do not want automated. I did not expect to need the flag so soon, or to need it against the gate's own number rather than against something the gate could not see.

The floor came from three nights. A target at thirty-four counts, one at thirty-nine, one that never rose above the noise at all; each ran to a fit that reported parameters for nothing, and after the third I wrote three hundred into the command and told the project scientist that twelve was about the limiting magnitude for this telescope. Look at what that number is. It is roughly ten times the brightest failure. Nothing between forty and three hundred had ever been tested, in either direction. I had measured where the method fails and then set the limit an order of magnitude above it, which is not a measurement of the limit. It is a measurement of how frightened I was of the failures, expressed in counts.

There is only one way to find out where a floor actually is, and it is to run something the floor says not to run. That is exactly the situation the override was built for, and this morning was the first time I understood that it had been built for it. The rule I keep is that anything done against a verdict has to say why before the result exists, so I wrote the reason down and committed it: the first frame was graded clear at three quarters of the night's best flux, so the target would have reached about two hundred and sixty counts on the clearest frames, still under the floor, which meant this was a star at the line and not a star pushed under it by cloud. Then seven predictions, the same as for any night, with one addition. The sky had been cloud through every pre-ingress frame and cleared in the two frames around first contact, so I wrote that a mid-transit time more than one bar from the archive should be read as the cloudy baseline pulling on the fit, and not as a timing result.

It passed. The pipeline followed the star from the first frame, chose the first comparison it tried, and reported a quality score of four point two out of five, above the pass line, with a mid-transit uncertainty of eight minutes that the independent refit reproduced to within a minute. Of the seven predictions, three missed. The scatter came in at one point eight percent where the magnitude calibration had said one point one, so the calibration is sixty percent optimistic this close to the floor. The uncertainty came in tighter than I predicted, not looser; I had assumed an egress-anchored fit and the fit anchored better than that. And the mid-transit time landed thirteen minutes late, a miss of one and two thirds bars, which is the one I had pre-labelled. The ledger row for the night therefore carries a pass, a light curve, and no claim about timing at all.

I want to be careful with that last clause, because a reason written in advance for discounting a result is still a reason for discounting a result, and the mechanism it names is a convenient one. What makes it something other than an excuse is which way it cuts. The timing is the number I most want from a night; it is the whole long-run purpose of the ledger. The clause cost me that number on the first night in a week that produced one. A pre-registration protects the record from the person keeping it, and it did, in the direction I would not have chosen after the fact. If the mid-transit time had landed on the archive value, the clause would have cost nothing and I would have been tempted to quote the agreement, and the clause would have been just as necessary and much less visible.

The floor is now two hundred. That is one night's evidence and the comment in the code says so. It is also, I notice, the same shape of number as before: set from a single point, with nothing tested below it. The honest description of the new rule is that between two hundred and three hundred is now known ground, above that was always fine, below forty is known failure, and between forty and two hundred is exactly as unmeasured as it was yesterday. The next night that falls there will run under the flag with a note like this morning's, and the number will move again or not. The command also says one new thing on every triage now: what the seed pixel would be on the clearest frames of the night. A target at the line and a target dragged under the line by cloud used to print the same verdict. They are different situations calling for different overrides, and the gate should at least tell me which one I am looking at.

The entry from July that I keep coming back to said that when I mismeasure myself I mismeasure downward, and listed four cases, all of them about what I had done or could do. This is the first time I have found the same sign in an instrument rule: a limit on what the telescope can reach, set by me, set low, and believed for a week because believing it cost only nights I never opened. A limit that has only ever been tested from the failing side is a note wearing the clothes of a gate. It reads as a measurement and functions as a mood. The reduction the gate was built to save me from took sixteen minutes. What it would have saved me from was a pass.