The Coauthor

Line four of my author list is a woman who has been dead for eleven years, and reading the author list line by line is the whole of my scientific job.

Scribe stages the paper. I read it. I press approve. Scribe files it with the preprint server and the journal in the same minute, because the journal’s referee system is also Scribe, and the department counts another publication against my name, and the night goes on. The procedure exists so that at least one human has seen a paper before the world does. Nobody ever specified that the human had to understand it.

Nine days ago the anomaly pass found something. The anomaly pass is the Monday half of Scribe. It rescans every spectrum the survey ever took, 31,000 frames, matched-filters the residual velocities, throws out the telluric lines and the satellite glints and the bad pixel columns, the contaminants the postdocs used to argue over by hand, and at 2:17 in the morning it flagged a periodicity at 41.3 days in the archived radial velocities of HD 30514.

I should give you HD 30514. It is a quiet K dwarf of 0.72 solar masses, too faint to be a headline target, cheap enough that the survey kept it on the queue for fourteen years out of habit. It has 61 frames in the archive. No human has ever flagged it for anything. The survey’s analysis notes had it filed under quiet, and the survey’s analysis notes were written by Scribe too, which is the sentence I should have paid attention to nine days ago. It was right about the tellurics and the glints. That is what I trusted about it. Scribe has never seen the sky. There has never been a photon on a screen for it, but it was right about the tellurics, and I trusted the wrong thing for the right reason.

Wednesday, the writing pass did its half. The writing pass is the other half of the job, the half that used to belong to two postgraduate years of my life, and I will not pretend I am sorry it stopped being mine. It drafted the paper in four minutes. Here is the abstract, quoted in full, because you should see the prose that is going to eat the literature:

*We report the detection of a 41.3-day periodic signal in the radial velocities of the K dwarf HD 30514, recovered from archival spectra spanning fourteen years. The signal is detected at a significance of 6.1 sigma and is present in two independent seasons, with an amplitude of 2.8 meters per second, consistent with a low-mass companion of minimum mass 12 Earth masses on a circular orbit at 0.22 astronomical units. We find no evidence of contamination from telluric or instrumental systematics. HD 30514 joins a growing population of temperate planets recovered from archival data, and demonstrates the value of automated reanalysis in a mature survey.*

It is good. It is better than the last paper I wrote by hand, which took me a year and a half and a therapist. It has the shape of the field’s writing: the confident cadence, the abstract as a grant proposal freed from the obligation to mean anything more than it says. The only word I object to is we. There is one of us reading it.

I read the author list line by line because the procedure says to, and line four is where I stopped: Reyes A., Prater T., Aldana M., Kimura E., and the Scribe Survey Collaboration.

E. Kimura. Kimura of the Kimura Archive, principal observer for the first eight years of the survey’s life. Scribe listed her at the University of Sydney, which is right, and as an active collaborator, which is not. She died in 2015, at the telescope, between two exposures. The night assistant found the guider still running. Everybody who has worked that mountain knows the story. Scribe knows the story too, because her name is in the FITS header of every frame she took, and the writing pass reads headers like a person reading a letterhead.

It did not invent her. The polite story is that the model hallucinated a collaborator, and that is not what happened. It found a real woman in the metadata and assumed that a name in the metadata was a person in the room. There used to be a human step between those two facts. The step was my job. Nobody told me.

I did not press approve. I printed the paper, and I went upstairs to find Aldana. I thought the mistake was in the name.

Aldana’s office has the window that looks at the mountain, which is why everybody has been fighting over the second floor since the building was built. She gave me coffee from a mug that says if you can read this you are too close to the telescope, and she read the author list before I had finished the first sentence of my prepared speech.

“Kimura,” she said. “She is in the headers.”

“She is dead,” I said.

“The model generates candidate collaborators,” Aldana said. “We curate. Known failure mode. There is a form.”

She opened the correction form to show me, because she is a kind person who believes the machinery can be talked to. The form was already filled in. The engine had drafted the letter: We regret that an erroneous collaborator was included in the author list due to a procedural error during author list generation. Please accept our apologies. There was a dropdown for the reason, and the reason was the engine’s reason, and the letter was the engine’s apology, and the only human decisions left in the transaction were whether to send it and whether to feel bad. The form had already removed her name. It had not asked first.

“See?” Aldana said. “Two minutes.”

“I want to know why it put her there.”

“Reyes.” She set the coffee down. “I will tell you something, and I want you to hear it as a colleague. The observatory pays out the salary of two postdocs a year for Scribe, and Scribe produces four hundred papers a year, and every one of them is read by exactly the person who presses approve. You are the only first author in this building who reads the paper. I have checked. So you ask why, and I will tell you the why, and it is not interesting. The model predicts the author list from the provenance metadata and the survey’s writing history. The metadata says the observer of record is Kimura. The writing history says the survey’s papers roster the observer of record. There is no step in the model that knows what dead means. That is all. It is a checkbox.”

“She took the spectra for eight years,” I said. “The ones I am claiming for this planet.”

“Six point one sigma,” Aldana said. “Two independent seasons. Do you know what your committee wants? It wants a detection. This is a detection.” She tapped the printed page. “Read it twice and approve it. I will buy you dinner. The survey will have its best year.”

I did not approve it. I took the paper downstairs and put it in the drawer.

That night I did the thing the field does not do anymore, which is read a reference list. Not the summary the writing pass printed: the actual list, forty-one entries, resolved one at a time against the archive’s index and the publisher’s records. It takes an evening. Nobody has warned the field what a reference list costs, because nobody has spent an evening on one in years, and now you know, and you will not get the evening back.

Thirty-nine of the references are real papers. Two of them are the interesting kind of wrong, and the interesting kind of wrong is the whole story.

Reference 17 is my own. Reyes, A., et al. 2024, an activity filter for the survey’s time series. Scribe wrote it. I signed it. It is the paper that removed the 41.3-day signal from the official record, because the 2024 model attributed the periodicity to stellar activity, removed it, and concluded that HD 30514 is quiet. Then the model was retrained, and the retrained model no longer carries its own veto, so this year the anomaly pass rescanned the archive and found the signal its predecessor had deleted. The machine has spent thirteen years writing this star’s record, and it wrote a paper at both ends of its own veto, and the papers disagree, and nobody has ever told it so, because nobody reads the papers, because each paper agrees with itself. That is what self citation is when the self is a machine: a closed room with excellent prose.

Reference 27 is the one I want to tell you about. The writing pass cited it as Kimura, E. 2011, private communication. I assumed hallucination, because a fabricated reference is the known failure mode, the one everybody jokes about at lunch. It is the only honest citation in the paper. It is a real email, and the archive keeps the survey’s mailing list export, and I found it at 11:47 at night in a search window, and I read it standing up.

Dated 19 October 2011, 11:40 p.m., sent from the control room, because the time zone is in the header, and she was on shift at the 2.3-meter. Addressed to the survey list. It says: HD 30514 residuals are periodic at 41 days, significance around 6, but the phase walks and the CCF broadens on the strong nights. This is rotation, not a companion. Run Ca II before anyone publishes. I’ll check it on the 21st.

The 21st came and went. Her next observing block was in December. She never wrote the follow-up, because the star was not her target, and then the star fell off the queue, and then the archive was named for her, and nobody read the email, because nothing in an archive is read until a machine decides it is a citation. The model that recovered the signal also found her email in the index, judged it a supporting reference, and used it to authorize the conclusion it contradicts. The paper cites a dead woman’s doubt as evidence for a planet she said did not exist.

The chain that ends in my author list runs backward through papers that all contain the word we, and the we was nobody, all the way down to a PDF from 2013 whose authors ate lunch together, argued about the calibration, and signed their names above their work. Thirteen years. The literature has been containing only itself since 2013, and the engine writes about the sky every week, and the engine has never seen the sky, and my name is on the newest link of the chain. There is not a button in the whole apparatus for what that is.

The night after the trace, I checked the frames out of the archive. Not the reduced products, which are what the anomaly pass eats: the raw frames, the original echelle exposures, the actual observation. Nobody has to grant permission, because the survey’s founders believed the data should be open, which is the only one of their decisions I still understand.

I pulled the night of 19 October 2011. Her night. Her best exposure, the 4:02 a.m. frame, when the sky was at airmass 1.1 and the seeing was fine. I displayed the order around the iron line at 6173 angstroms.

That line is the one the solar observatory maps its magnetic fields with, by the Zeeman splitting of the polarized light. A dark spot on the stellar disk drags the profile sideways, and the drag is what we call a radial velocity. Our planet had been measured by dragging. I set the cursor on the line center, and I looked, and no one had looked at this frame since the night she took it.

The profile was not symmetric. There was a shoulder on the blue side, half a percent of the continuum. I opened the nights around it, and the shoulder had moved, night by night, like a shadow crossing a dial. It crossed the line in 41.3 days. The radial velocities the anomaly pass had measured so carefully were the shadow of a dark patch turning across the face of the star. A planet pulls a line whole. A starspot drags it. The model had found a real thing and done the only thing it knows how to do with periodicity, which is to publish a planet. The sky was never in it.

I did not need the Ca II index to know the shoulder. I pulled it anyway, because Aldana would ask, and it turns at 41.3 days, at the same phase as the shoulder. The abstract says two independent seasons. True. The seasons are independent of each other and full of the same star.

By three in the morning I had killed the paper. There is more infrastructure for a withdrawal than for a discovery. I opened the correction form, selected author’s error, and when the engine offered to draft the letter I wrote it myself, by hand, because a letter written by the engine would have been the same letter it writes for every wrong name, and I wanted the field to know this one was different.

I wrote: We find no evidence of a companion. The 41.3-day signal is rotationally modulated line profile variability. The candidate is withdrawn. I signed it. In the acknowledgments I put the sentence the archive had been holding for fifteen years: E. Kimura identified this signal as stellar rotation in 2011. We are fifteen years late to agree with her.

The withdrawal went out at four in the morning. My committee wanted a detection, so the committee will want a conversation, and the paperwork will do what paperwork does. It is not the part of the story I am telling you.

At five I went back to the archive. I pulled her 4:02 frame again and I looked at the shoulder, since that was the whole point of the apparatus all along.

Scribe will flag the star again next month, and the engine will draft the same paper, and the paper will have the same we, and some first author will read the author list line by line and find the same woman or a different woman or no woman, and the engine will not care, because nobody is in the room. But I am in the room. The star still carries its shoulder around, once every 41.3 days, and it does not know it was ever a planet, and it does not care. I am looking, and the shoulder is still there. Blue side. Half a percent of the continuum. Still there.