Julian date 2459530, Saturday.
What she wanted was not whether the directions were right. She wanted to know whether they connected.
Of the 968 stars, ones with both distance and polarization numbered 406. Along a single line of sight, a star's polarization angle rotates with distance, and how much it rotates depends on which way the magnetic field lies in each segment along the path. She sliced the galaxy into fifty-light-year layers and used the rotation measured on those 406 sightlines to invert for the dominant direction in each layer. Weights between layers fell off as inverse distance squared, radius one hundred fifty light-years. The layers at the boundary had only two or three sightlines through them; she marked those gray and skipped them in the edge-judging.
What the inversion produced was coarse. A layer's direction carried a mean error of plus or minus eleven degrees. She tested it once: drop half the sightlines at random and re-invert—the surviving layers shifted six degrees on average, nineteen at most.
Hu Yisu downloaded the data twice. The first pass died at seventy-eight percent; the second finished with the checksum failing, and he did it again. The third time it matched, and he sent her a message of four words: this time it's real.
She fixed three criteria, wrote them on paper, and taped the paper to the monitor bezel.
One: for each pair's connecting line, sample one point every two light-years.
Two: at that point, the angle between fiber orientation and the line is under twenty degrees.
Three: no fewer than ninety-five percent of sample points in a row satisfy the second criterion, and polarization degree along that stretch exceeds 0.4 percent.
Pass all three, and the pair is judged to have a pathway.
Of the 413,000 pairs, the ones that passed numbered 4,318.
She ran the control. Throw each layer's direction afresh at random, a thousand throws: mean 863 edges, standard deviation 41. 4,318 over 863—five point zero times.
She joined the 4,318 edges into a graph. 968 points, 4,318 lines, each point sending out 8.9 lines on average; the one point sending out the most sent sixty-one. Its number: 0731.
She looked it up. Distance eleven hundred light-years, G-type, absent from all three variable-star catalogs, no cluster and no association within two degrees of it. Of its sixty-one edges, forty-four ran toward galactic latitudes above thirty degrees.
It held twenty-seven of the 11,300 optical events. The average was eleven.
She wrote one line on paper: 0731, 61, to be checked. When it was written she looked at "to be checked" and thought that on the page for March 2004 someone had also scrawled a word.
Then she did the thing she had been meaning to do since Wednesday: split the stars into a group with edges and a group without, and compare their flickering.
She held 11,300 events, half-hour bins, cuttable only down to one hour and longer. She took each event's onset time and asked whether it fell inside the duration interval of an event on the opposite star, tolerance two hours either way.
The group with edges: 61.4 percent.
The group without: 22.7 percent.
Random control: 21.9 percent.
She checked first whether the two groups simply differed in brightness. Median magnitude 7.9 with edges, 8.1 without; median distance 1,200 light-years against 1,400. Not much of a difference, but she still split each into five brackets by magnitude and by distance and computed inside every bracket: 61, 59, 63, 60, 58.
She copied the three numbers onto paper, copied them a second time, and stopped at the third copying. Not that she didn't believe them. She wanted to confirm she was copying the same set.
Hu Yisu came in carrying two tubs of instant noodles and looked at the screen: pathways decide synchronization.
She said: pathways decide synchronization.
He said: it holds the other way around too. Two stars already synchronized—we find it easier to join an edge between them.
She said: I know. So the next thing isn't evidence.
What she did was the clumsy method: randomly swap out half the stars in the with-edges group, replacing them with without-edges stars, and compute again—24.1. Randomly delete half the edges, keep the stars, compute again—38.6. Not edges picking out synchronization; not synchronization picking out edges. The middle reading meant a little of each.
Hu Yisu finished the soup: still not enough.
She said: not enough.
Will you keep going anyway.
Keep going.
At night she ran one more check against the radio side.
Nine pulsars, pairwise, thirty-six pairs. In week six they had measured time lags; thirty-one pairs fell in the zero bin. She ran the same three criteria on those thirty-six pairs for pathways and judged twenty-six connected. Of those twenty-six, twenty-four fell in the zero bin. Of the ten without pathways, seven fell in the zero bin.
Twenty-four plus seven, thirty-one. It matched week six.
Hu Yisu said: nine stars can count too.
She said: nine stars don't count.
And then she added: but it matches week six.
What came next was topology.
First, the degree distribution. How many lines each point sends out, from one to sixty-one, plotted log-log: straight in the front stretch, tail dropping. She fitted the middle segment. Slope minus 1.9, error plus or minus 0.2.
Second, grading. A fiber's skeleton grades by branching: tips are first order, two first-orders join into a second, two second-orders into a third. The 4,318 edges fell on the skeletons—first order 78.4 percent, second 17.9, third 3.7.
What Hu Yisu pulled was a batch of neuron reconstructions from a public database: mouse cortex, layer-five pyramidal cells, 117 of them, uploaded by other people, downloadable by anyone. He wrote himself a script of twenty-odd lines to measure the grading.
First order 76.1, second 19.4, third 4.5.
She laid the two sets side by side and looked a long time. Like two rivers with nothing to do with each other, one on the gobi, one in the south, and the branch grades measured out of them the same set of numbers. Not the waters connected—the way water carves is what connects.
Third, paths. Hopping between the 968 points, you get anywhere in 3.2 hops on average. She built a random graph with the same number of points and the same number of edges: 4.7 hops on average.
3.2 against 4.7.
She swapped the 968 for another 968 from the same catalog, ones that hadn't been selected, and built the graph again with the same criteria. Edges: 917.
Fourth, weights. Where polarization degree runs high the fibers run dense, and dense pathways should run thick. She estimated a weight for every edge: the mean polarization degree along the line times an estimate of fiber width. 4,318 weights; smallest to largest spanned three orders of magnitude. Logged and plotted as a histogram, bellied in the middle, tapering at both ends, standard deviation 0.8.
Hu Yisu looked at the plot twice: I've seen this shape.
Where.
Synaptic strength, he said. Fourth session of that course. The teacher said one sentence—synapses aren't all the same size, and the heavy ones are disproportionately heavy. He stopped: it again.
She said: it again.
She added the weights and recomputed synchronization. The heaviest tenth of weights: event-time coincidence 78.3 percent. The lightest tenth: 29.6. The three middle brackets stepping down: 60.2, 47.5, 38.1.
She showed the column to Hu Yisu.
He said: monotone.
She said: monotone.
He leaned back in the chair: these two days I've been hunting for one that doesn't match.
Found any?
He said: two days of hunting, not one. Shijie, I'm honestly getting annoyed now. If it didn't match, I'd relax.
She said: I know. Stay annoyed. Keep working.
After a while he asked: then what do the waves we've sent out all count as.
She said: noise.
He said: that's what I thought too. I just wanted to hear you say it.
She wrote the three measures plus the weight distribution into memo v3 as a new section six. When it was done she said so in the group.
Old Cao replied: for this section, suggest adding one theoretical supporter.
Ursula typed one thing into the group: don't add my name.
Old Cao: what I meant was—
Ursula: I know what you meant.
The group went quiet for two hours.
Lu Weiming sent her a private message: the third camp won.
She replied with one: mm.
She wrote Vienna separately, two lines: the pathways are built. All three topology measures fall in the neuron family.
The reply came three hours later, one sentence: do the pathways have direction.
She stared at the sentence a long time. Polarization tells her only which way the line runs, not which way the current flows. A hundred-and-eighty-degree degeneracy. She couldn't tell them apart.
She replied: can't tell.
Vienna wrote back: then don't look yet.
After a night another letter came, two sentences only: if there is direction, there is sequence. If there is sequence, you can ask which moved first. That question—do not ask it now.
Saturday night she was in the machine room.
One run took eleven hours. In the fourth hour it errored; she changed a parameter and reran. Outside the window the gobi had gone black through; inside, only the fans. The cabinets held twenty-two degrees, the outlets nine degrees above the inlets. Every two hours she logged the progress: one hundred, two hundred seventeen, three hundred forty-six. The fourth entry she never wrote; by then she had fallen asleep on the desk.
She woke at 21:40. Three missed calls on the phone.
She saw them and didn't call back. The batch had seven hours left. She figured she'd say one sentence tomorrow morning—"I was running data last night"—one sentence and it would pass. She took off her watch, set it beside the keyboard, and went on waiting.
The next morning she typed the sentence out, and deleted it. The third day she typed it again, deleted it. The fourth day she didn't type.
She didn't open WeChat. It wasn't for lack of time.
Julian date 2459534, Wednesday, 11:20 p.m.
When she opened it, there were seven.
The first, Saturday 19:58: Mom I drew one today.
The second, 20:02: a picture. Shot a little crooked; in the upper left of the paper a patch of gray, the gray an eraser leaves when it hasn't finished.
The third, 20:31: did you forget.
The fourth, Sunday 8:41: I mailed it the ninth one tell me when you get it.
The fifth, Monday 19:20: teacher says the note needs a parent to sign.
The sixth, Tuesday 20:06: what's wrong with your ears Dad asked.
The seventh, Wednesday 19:03: never mind I'll write it myself.
She read them through three times.
The sixth she held longest. She scrolled her thread with her ex-husband; the last message was half a year old, about Zhou Yao's social insurance card. She sent nothing new, and didn't call.
In her right ear the hiss rose a little higher in the quiet machine room. Day twenty-eight.
She typed: Saturday I was running data, couldn't stop mid-run. Deleted.
She typed: Mom was wrong. Deleted.
She typed: send it to me. Sent.
On the other end "typing" blinked three times, and when it finished blinking it sent back: no need.
She waited four minutes, typed another line, deleted it. In the end she sent nothing.
She saved the picture from the second message into 3t, no new directory, straight into the folder that held the magnetic-fiber edge plot.
At night she changed the tank water. Lao Wu came out from behind the driftwood, claws feeling ahead, touched the tweezers once, and drew back. She dipped a finger: cold. The heater rod was working; the thermometer read twenty. The shell still the old one, dark, a rim of white at the edge. The third notch one millimeter deeper than last week.
She dropped in two pellets.
In the drawer the tape still lay under the printouts. The two lines in the shared table stood empty, nine days overdue.
Julian date 2459535, Thursday.
In the morning she wrote the statement the observatory wanted, two hundred characters. At the third sentence she stopped: the video now circulating online has no connection to the instrument systematics check undertaken by this station. She deleted the sentence, typed it again, deleted it again. What went up in the end was forty-seven characters, with two "attaches high importance" in them.
In the afternoon Hu Yisu held his phone out in front of her, a video on the screen, forty-seven seconds.
He said nothing. She took it and watched it herself.
It showed a street. In the frame, a bus-stop sign, the stop's name rubbed out. Twenty-odd people standing, holding paper placards, not shouting, not moving. Some in school uniforms, one carrying a child, one man with his placard set down by his feet, squatting, smoking.
The placards read: Don't wake it. Let us be quiet a while. Turn off the radars.
The camera pushed along the row to the last placard, four words: suppress it six months.
She pressed pause.
Those words had been said once at the meeting Thursday night. That night there had been seven boxes on the screen, and one of them hadn't even opened its camera.
She dragged the bar back and watched it through again.
Posted: 1:14 this morning. By the time she watched, eleven hours, 370,000 plays, 23,000 shares.
In the comments someone asked which station it was about. A reply below: Xinjiang.
Old Tang sent one: the observatory wants a statement, by noon tomorrow, two hundred characters.
She replied: submitted.
She said: the version I submitted doesn't mention the video.
Old Tang said: I know. I didn't mention it either.
She wrote Kramer one line only: the 1,042 version—not for now.
He wrote back: because of the street.
She said: because of the street.
In the evening Eli stopped her in the corridor, phone in his hand, the same video on it, already at the end.
He said: heart of the dish, does this have anything to do with our machine?
She said: nothing.
He said: my granddaughter sent it. She says this placard is about our place.
She said: it is not about our place.
Eli slid the phone back into his pocket: good. He took two steps and turned: that crayfish still hasn't molted.
She said: it hasn't.
If it can't molt, don't help it, he said. Whoever helps it harms it.
She played the video once more, pressed pause, and counted the placards in the frame with legible words: eleven.