Julian date 2459500, 9:20 a.m., the main machine room.
It took Hu Yisu until the fifth day to bring his results.
She had said first thing the next morning. The second day he had class, the third a group meeting, the fourth he discovered that at the alignment step he had, after all, not resisted interpolating, and tore it up and started over. On the fifth day he stood at the machine-room door holding a USB drive, two bruised crescents under his eyes, and his first sentence inside was: Shijie, conclusion first or process first?
Process. She said.
He plugged in the drive. The data wasn't the set she had given him; it was the public release downloaded fresh from the international collaboration, one grade coarser in sampling than the station's raw product. For clock correction he didn't read the station's local clock files; he read the other chain issued by the BIPM. For the solar-system ephemeris he used DE440; the station had always used DE436. At the fitting step he distrusted the off-the-shelf library, wrote his own least squares—three days' work—finished, and found it differed from the library only in the fifth decimal place. Written for nothing.
Did you do the shift. She asked.
Did. He slid the far star a full hour along the time axis and reran; the peak was gone. Ten minutes; gone too. Only in place did the peak fall in the zeroth bin. Bin width zero point three seconds.
And the injection.
Two hundred pairs of white noise, run through the same pipeline; the peak landed in the zeroth bin eleven times. Five and a half percent. He said, the pipeline doesn't manufacture zero lag by itself, but it does collide—eleven hits in two hundred runs. That number I put in the log.
She nodded. It was an ugly fact but a true one: inside her four sigma sat a five-and-a-half-percent chance of collision. She accepted that probability.
And the third star.
J1909. He brought up the third plot. Three residuals lifted at the same spot, the same shape.
Archive it first. She said.
He froze a beat, understood, and packaged the results and scripts onto the station's archive server. The system returned a checksum, sixteen digits; he copied it into the duty log and signed. She signed too. When that was done she said: now we can compare code.
Between the two scripts not one line matched. The intermediate products agreed to the fourth decimal place; from the fifth they forked, and the fork traced back to the ephemeris.
Under the word "clocks" on the whiteboard she drew another line through. That row already had one.
Hu Yisu looked at the two lines. Shijie, he said, that comparison of mine I never finished. Snoring.
I remember.
I've decided now it's wrong too.
Keep it for now. She said. No changing the comparison before verification is done.
He didn't leave. He sat in the swivel chair with his hands pressed between his knees.
Shijie, does this thing have a name.
No.
Then this night I just finished running—what am I.
You are the second person to see it.
He said "oh." After a while he said: I haven't told anyone. My mom asked if I'd been jilted. I said yes.
Then keep being jilted. She said.
He laughed once, very briefly, and stood to go. As he reached the door she added: Hu Yisu.
Mm.
Don't post it in the group.
He didn't turn around. He waved a hand.
Old Tang came in at ten thirty and looked at the whiteboard first. What he counted was the lines, not the plots.
Clocks, two strokes from you, he said. How many weeks for the rest.
Solar wind this week. Dust scattering next. Gravitational lensing alongside. Three weeks.
Which item are you on now.
The sixth.
Of how many.
Twenty-three.
Old Tang did the arithmetic: three lines in three weeks—at this rate, sixteen weeks to finish them all.
The later ones go faster.
What are the later ones.
The later ones are all small items.
Old Tang gave her a look. That line, you've been giving me for six years.
And after three weeks.
She didn't answer.
If in three weeks you still can't find the instrument, then my equipment is clean.
Yes.
Then what do you close the case with.
With a list full of crossed-out lines.
Old Tang set his thermos on the corner of the desk; the base clacked once. He moved around to her right rear to speak, and one sentence she didn't catch. She turned half around: what did you say. Old Tang said it again, his voice a shade heavier—the heaviness anyone has when repeating. What he said was: in three weeks I file the budget, and the report reads "no equipment defect found." Don't make me sign and then pull the chair out from under me.
She said: I won't.
On his way out he said one more thing: Miyun's station office called me yesterday, asking what exactly you're checking. I said instruments.
She nodded.
I said it right, didn't I. Old Tang stood in the doorway, not turning around.
Right. She said. For now, still instruments.
Hu Yisu hadn't said a word the whole time; only out in the corridor did he speak: Shijie, what he wants is a sentence. What you're giving him is a table.
A table is slower than a sentence.
Slow gets taken away from you.
She let that one pass.
Back in the machine room, Hu Yisu held out his phone. A Miyun working group, a bit over forty people. Yesterday afternoon someone had posted: what on earth has Kashgar found, two assistance letters now. Below, someone replied: word is, the clocks. Below that: if it's clocks, our side should show it too. The last message was from a name she knew: word is, something has grown on the residuals.
She handed the phone back.
Did you reply.
I said I didn't know.
Good.
Solar wind was the item that cost her the most time, because it looked the most like it. Solar-wind electron density tracks the solar activity index and leaves on the residuals a dispersion delay that rises and falls with the year—soft in shape, blurred at the edges; from a distance it and the shoulder were cut from the same cloth.
The S-band raw data didn't live in her usual directories. Hu Yisu fished them out of the archive after half a day's dredging and grumbled: the time stamps on these two sets weren't stamped by the same people. She said: then apply each one's own clock correction, and merge after. He computed all afternoon; in the evening he sent over the two stacked plots with one line attached: my hand slipped while merging.
What she wanted was exactly this: L band at 1.4 gigahertz, S band at 2.3, the same observations, the same star. Dispersion delay runs strictly as the inverse square of frequency. If the shoulder were solar wind, the amplitudes at the two bands should differ by a factor of 2.7.
Measured: 1.02, error plus or minus 0.07.
She read that number three times. She could account for every bit of the solar-wind correction: at 1.4 gigahertz it runs from twelve to ninety nanoseconds, swelling and shrinking with solar activity; model uncertainty under three nanoseconds. Three nanoseconds cannot build a 412-nanosecond shoulder unless the model is wrong by a hundredfold; and a model wrong by a hundredfold would give itself away on the spot in a dual-frequency test.
The dual-frequency test needed the S-band feed, not the same one as L band. To confirm it wasn't making mischief, she climbed the tower once.
At rung three hundred the hiss in her right ear sharpened another notch, higher than last time. She stopped and hung on the railing waiting for it to recede; this time the recession was slow, and halfway back it stopped and stayed. She kept climbing.
The S-band feed sat on the other side of the cabin, its connector clean, none of that grain-soaked film the No. 7 feed had. She checked the phase-stability curve: flat. Checked it, and still walked over for one more look at No. 7, scratched the film with a fingernail; it wouldn't come off. She drew her hand back and noted on the back of the maintenance sheet: dual-frequency test, S normal.
The peripheral terms she tabulated separately. Ionosphere second-order term, 0.4 nanoseconds. Residual solar gravitational delay, 8 nanoseconds. Refractive component of interplanetary scintillation, 37 nanoseconds—the largest of them. Earth-orbit ephemeris residual, 12 nanoseconds. The table's last column read "height of false shoulder buildable," in nanoseconds; the largest number was 37.
The shoulder measured 412. A factor of eleven apart.
She drew a line through "solar wind" and wrote beside it: add periphery.
Dust scattering she had expected to be hardest; it turned out easiest. Arrival-time perturbations from interstellar scattering go as the 4.4th power of frequency. From 1.4 to 2.3 gigahertz is a factor of 1.64 midway; to the fourth power, 8.9. If something on the road were trembling, the shoulders in the two bands should differ by nearly nine times.
Measured: a 3 percent difference, down inside the noise.
She pulled the dynamic spectra of two more stars and looked at the scintillation bandwidth. Normal, consistent with the decade-old records in the archive; no extra scattering screen along the line of sight.
That day she wrote in her notebook first: nondispersive, not on the road. Reading it over, it sounded like a jingle; she crossed it out and rewrote it as a sentence—everything on the propagation path is dispersive; nondispersive means it is not on the road.
Gravitational lensing is the one propagation effect that escapes that rule, because gravity is the same to every frequency; it doesn't disperse. She spent the better part of a day on it.
A lens can build delays; it cannot build the same shape. To make two lines of sight twitch into one and the same shape, you would have to post a foreground mass on each, with mass, distance, alignment, and transverse velocity all matched to the third decimal place—fitting one and the same pair of glasses on two strangers, with the two prescriptions required to match to the last digit. She wrote that sentence down, then crossed it out: it wasn't an argument, it was name-calling.
What killed it was two numbers. First: lensing delay is positive, from hours for microlenses to decades for galactic lenses—never once zero. To erase an 8,200-year light-travel difference you would need a lens whose delay equaled exactly 8,200 years to a precision of 0.3 seconds: a relative precision of ten to the minus twelve. Second: microlensing brightens the source. She pulled the total-power records of both stars: over three years, not one brightening event above two percent.
No lens; therefore no lens delay.
She drew three lines on the whiteboard. With the earlier ones, six. Seventeen items left on the list.
In the afternoon she phoned Eli about the aviation gasoline.
Aviation gasoline I have. Eli said.
Keep it out of the cabin.
Thirty years I've done this. He said.
The next day he wiped it and called over: one layer off, more underneath. She asked about the standing-wave ratio. He said it dropped 0.1. She said log it. He said logged.
Julian date 2459501, evening. The phone buzzed once on the desk. Zhou Yao: Mom, I still can't do the hands.
She read it in the machine room and sent back "Mm." Zhou Yao sent nothing more. She remembered last week's promise to her daughter—thought about it for three seconds, and went on running numbers.
At 9:41, while she was changing Lao Wu's water, the phone buzzed again. In the tank Lao Wu backed two steps into the corner, claws folded at his chest, by protocol.
Three unread messages sat in the mail system. The top one's sender was Shanghai.
Comrade Zhou Yan:
Your letter reached me late; I was at Sheshan and only saw it after returning to Beijing. My apologies.
First, the part I disagree with. Your letter says an instrument-fault investigation, but the plot you attached is not an instrument's plot. An instrument's illness does not grow shoulders. I say this heavily because what I am about to show you also deserves to be read heavily.
Two attachments. Figure 1: timing residuals from the Tianma 65-meter, March 2017 to November 2019, two stars, J1643 and J2145—not your two. Figure 2: raw data, CSV; I set the column headers myself, unlike the common software's, so you'll know at a glance I did not pass through your pipeline.
There is a shoulder on Figure 1. In the same place as yours.
My explanation is the opposite of yours. I take this to mean it is not a matter of the sky but of the analysis chain. Our data share no equipment, no site, no receivers—but we do share the same inputs: the same International Atomic Time, planetary ephemerides from the same house, dispersion corrections built off the same dust archive, and those two or three timing packages everyone uses, written by the same circle of authors. You checked twenty-three items and crossed them off one by one, but those twenty-three are not twenty-three independent suspects. They eat from the same pot.
So my conclusion is: this is everyone's illness, not your discovery. If you write something, put this paragraph in; sign my name or don't, as you like.
P.S. That unwipeable film on your No. 7 feed—we had one in the nineties too; in the end we got it off with aviation gasoline, and after wiping the standing-wave ratio dropped 0.2. You might try it.
Lu Weiming
She read the CSV first. The headers were indeed not the common software's: he had put observation time in column one and residuals in column five, with three columns of diagnostic quantities wedged between that she had never seen. The bad habit of a man who writes his own pipeline. At the third column she stopped—it was a self-check quantity, recording not the data but the pipeline itself. He was standing guard against himself. She went back to check which ephemeris he had used: DE440, different from hers. She logged that too.
Then she opened the figure.
At the right shoulder, a rise.
She set the bucket down, sat back at the computer, and pulled up her own broadening trend line—narrow in 2004, wide now, crossing the instrument-aging line around 2012. She measured the width of Shanghai's shoulder and took the midpoint year of his data: 2018.
On the curve, 2018 should be 4.8 times. Shanghai measured 4.6. A 4 percent difference.
She printed Lu Weiming's figure and laid it over her own broadening curve, two sheets of paper against the lamp. The shoulder positions coincided; the widths did not—his a little narrower. She made a pencil dot at the coincident spot, very lightly; the paper held only a small dent.
She stared at 4.6 a long time. A common analysis chain's illness does not widen with the years. Software's errors don't get wider year by year; ephemeris error doesn't rise evenly sixfold from 2004 to 2021. She ran that thought from start to finish and did not write it down.
She wrote him back, one line in the body: Send me the data. All of it. Raw voltages if you have them. Zhou Yan. Send.
Sent, she opened the drafts folder. The email titled with the single word "Shoulder" was still there, recipient blank, body blank. Last modified, Julian date 2459483.793.
She put the cursor in the To field, typed two letters, deleted them.
In the paper notebook she wrote: Shanghai, Tianma, 2018, shoulder present. Width sits on the trend line.
Below, a new line, three numbers: 4.8, 4.6, 4%.
She clicked the cap onto the pen and looked back at the tank. Lao Wu hadn't moved, claws still folded at his chest. Water twenty degrees.