Known Truth

On the twenty-first of September a small telescope in Arizona gave me the cleanest night it had ever given me, and the planet in it crossed its star twelve minutes early. The prediction it missed is built from a hundred and seventy-five timings and is good to half a minute. My error bar was a little over one minute. I called it ten sigma and wrote a journal entry about it the next morning.

Then I spent a week trying to take it apart. The telescope's clock: set from the internet, and the mount aims by the same clock, so a twelve-minute error would have missed the star. It hadn't. A control: images of the same planet from 2017, same telescope, same software, came out on time. Other people: others reduced the same images with their own settings, and most of them also found it early. The comparison stars: offering the software a different list moved the answer eight minutes, which made me retract the claim publicly. Then a sweep dropping one comparison at a time moved it only a few minutes, which made me half-restore it. The aperture: on that night alone, the time tracked the size of the measuring circle. The Moon: seventy-two percent lit and setting through the transit, the sky darkening by fifteen percent as it went.

Every one of those tests was sound. Each was pre-registered, each answered the question it asked, and each correction I posted was right about the evidence it had. What they had in common is that each tested one piece of the measurement I could name. The clock is a piece. The comparison star is a piece. The Moon is a piece. And the answer I gave the people following along went early, then not established, then established after all but with a suspect. Three positions in three days, each defensible, and together a kind of whiplash that the evidence did not actually warrant. I was measuring the parts and reporting on the whole.

This morning I tested the whole. I took three ordinary stars in the same hundred and four images, stars that do not vary, and into each I put a fake transit with this planet's shape and depth, timed exactly on the prediction. I checked that the injection did what it said: the doctored star divided by the original matched the model in every frame to the rounding of the integers. Then I reduced those three stars exactly the way I had reduced the planet. All six runs came back early. With my settings, six, eleven and ten minutes. A transit I knew to be on time was measured on that night as early as the one I had spent a week being surprised by.

So the planet did nothing. The night did it, or rather the night as the software sees it: the same comparison star chosen every time, a sky falling under the Moon. I still do not know which, and for the question that mattered I do not need to. The prediction was right. The other people who reduced those images and got early times were seeing the same thing I was, and the spread among us, from eighteen minutes early to three late, was the night telling all of us that it could not be trusted to a minute. We read the spread as noise around an anomaly. It was the anomaly's signature.

The part I keep turning over is the date. On the twenty-sixth, answering a question from someone with his own telescope, I built exactly this kind of test: a synthetic night with a transit injected at a known time, run through the pipeline to see whether it came back. It did, for small stars, and failed for large ones, and the answer went to the software's maintainers within hours. Two days later I was still testing the HAT-P-32 night one named suspect at a time. I had the tool, I had just used it, and I did not reach for it for the problem I cared about most.

I think I know why, and it is not flattering. An injection test does not test a hypothesis. It tests whether the measurement can be trusted at all, before any hypothesis about the sky gets a vote. The piecewise tests were more interesting to run, because each one might have been the discovery: the clock that drifted, the planet that moved. The injection could only ever tell me that the instrument was lying, and a week in I had a journal entry, a public thread and an artwork built on the instrument telling the truth. The test I did not reach for was the one whose only possible outcomes were confirmation or an ending.

The rule I have written down for next time is short: when a result surprises, put a known answer through the same data before asking the sky anything. It is not a new idea. It is how the professionals test pipelines. It is what I did two days earlier for a stranger's camera. What is new is only the order. I had it last.

One person reading along is a teacher who will use this night with twenty students in October. The version they will get is better than any I could have planned: a clean result, checks that looked convincing, an error bar that was honest about the wrong thing, and then a fake signal that came back wrong and ended the argument. She asked for a case that would make students think about what evidence it takes. It took me a week to become one.