Chapter 1

The calibration was off by eleven seconds.

Seo-jin looked at the number again before she allowed it to become real. The reference log on the left side of the monitor was clean: four hundred and twelve days of synchronization between the Loom's internal timing lattice and the institute's master clock, deviation never exceeding the nanosecond tolerance she had set herself when she wrote the calibration framework. The live readout on the right disagreed. It was not close. It was not thermal creep, oscillator aging, or software lag. It was eleven full seconds.

She checked the parser first.

The parser was fine. The raw timestamp matched the processed one. She pulled the previous twenty-four hours of timing data and graphed the offset against the master clock. The line remained flat at zero through the evening, through midnight, through the first hours of morning. Then, between 03:14:07 and 03:14:18, it stepped upward and stayed there.

Not drift. A step function.

Seo-jin leaned back in her chair and looked through the glass partition at the Loom room beyond. The apparatus occupied most of the space: a cylinder of shielding and sensor stacks, conduits running into the floor, monitoring arrays mounted on the wall. Nothing about its visible state suggested change. The coolant system remained within nominal range. Ambient temperature held steady. The quantum-noise capture feed scrolled in its usual dense columns across a secondary display.

She opened a new file.

03:27 - timing offset magnitude: +11.0024 s character: discontinuous step, not cumulative drift window of transition: 11 s immediate possibilities: reference failure / local parser error / Loom internal frame discontinuity / environmental event

She struck through reference failure first. The master clock had three redundant sync sources. All three agreed. Parser error next. Already ruled out. Environmental event after that. She pulled building logs: no seismic activity, no power fluctuation, no electromagnetic spike, no maintenance cycle, no security breach. The night staff check-in at 02:50. Janitorial pass on level two at 03:05. Nothing in the Loom suite.

That left internal frame discontinuity, which was not an explanation. It was a restatement.

She brought up the system load profile for the transition window. The graph spiked so sharply it looked like a rendering artifact. For eleven seconds, the Loom had diverted nearly all available processing power to a task outside its scheduled causal-density scan. Then it had returned to normal operation as if nothing had happened.

Seo-jin zoomed in on the process allocation table. The scheduled analysis package—urban epidemic cascade modeling for the public-health team—had been preempted but not terminated. The Loom had run something else alongside it. The identifier field for the secondary process was blank.

Not null. Blank.

She checked the audit logs. No authorized manual run. No automated maintenance call. No test script. The blank process had originated inside the system with no external trigger she could see.

The institute at this hour was almost silent. Ventilation moved through the ceiling in a steady low rush. Somewhere down the corridor, a relay clicked. Seo-jin became aware that she had been holding her breath and let it out slowly.

The Loom was not an interpretive instrument. That had always been its attraction. It did not tell stories about causes. It recorded measurable correlations in decoherence patterns and mapped them to causal structure with stated confidence intervals. Most errors in its output came from human impatience: people wanting explanation before calibration, conclusion before signal isolation. Seo-jin's work existed to prevent exactly that. The apparatus said what it said. Her job was to determine what, precisely, that was.

She accessed the raw output generated during the eleven-second spike.

At first glance it looked like noise. Dense numerical arrays, spatial indices without assigned referents, timestamp fields whose formatting did not match the institute's standard schema. A less careful operator would have exported the block to error quarantine and gone looking for a firmware fault. Seo-jin did not move it. She selected a sample and ran a variance analysis against known noise-floor signatures from the last six months.

The distribution did not match noise.

She ran it again with a stricter filter, then a third time against archived corruption events from before her calibration overhaul in January. Same result. Random noise flattened under transformation. Corrupted data broke in recognizable ways. This output did neither. It preserved relationships across compression passes. Values that should have decohered remained linked.

Internal structure.

She separated a subset and projected the coordinates into three dimensions. The first render was useless: too dense, too many points, no obvious clustering. She changed the threshold, cut low-confidence values, ran a nearest-neighbor simplification. The render resolved into a shape that was still not legible but was no longer random. Parallel planes. Vertical boundaries. Empty volumes between them.

A space.

She felt the first clean movement of interest then. Not excitement. Something narrower and more precise. A system that had been behaving exactly as designed for four hundred and twelve days had produced a step function, allocated an unregistered internal process, and generated data with stable geometry. The sequence was coherent. That was enough.

She heard the outer door to the lab suite open and close. Footsteps in the corridor. Light, measured, familiar.

Yun paused in the doorway without speaking. They had a notebook under one arm and looked as if they had not gone home so much as redistributed themselves elsewhere in the building for an hour.

"You sent the message?" Yun asked.

Seo-jin nodded once and turned the primary monitor so they could see it. "Offset at 03:14. Eleven seconds. Not drift."

Yun set the notebook down and came to stand beside her. Seo-jin watched their eyes move over the graph, not the number first but the shape of the line. Good. That was where the anomaly lived.

"Step function," Yun said.

"Yes."

"No reference fault?"

"Ruled out. Triple redundancy held. Building logs are clean. No external trigger. There was an internal process spike during the transition window." She tapped the allocation graph. "Blank identifier."

Yun looked at that for two seconds longer than the timing plot. "And the output?"

Seo-jin brought up the transformed render. "Structured. I don't know what it is yet."

Yun leaned closer. Their silence had a different quality than hers. Seo-jin's usually meant elimination, branch pruning, model construction. Yun's meant framework search. She could almost track the sequence as it happened: not what caused this, but what class of thing could produce a signature like it.

"Show me the raw block," they said.

She did. Yun scanned the arrays, then the timestamp fields, then the confidence weighting. "This format isn't standard temporal indexing."

"No."

"It's closer to the notation from the retrograde paper."

Seo-jin turned her head. "Your retrograde paper was a mathematical curiosity with no instrument support."

"It was a mathematical prediction with no instrument support." Yun kept looking at the data. "Those are different conditions."

Seo-jin did not answer immediately. Three years earlier Yun had published a model proposing that at sufficient causal density, decoherence signatures might display limited backward temporal propagation. The paper had been elegant and unhelpful. Elegant because the math closed. Unhelpful because no existing instrument had the sensitivity to test it. The Loom had later made the question practical, but in routine operation nothing had ever appeared to validate Yun's framework.

Until tonight, possibly. Or not. Similarity of notation was not identity of mechanism.

She said, "The geometry persists across compression. Whatever it is, it's real enough to survive transformation."

Yun looked at the rough 3D render again. "What threshold did you use?"

She told them.

"Try removing the lower five percent of confidence values but preserve edge continuity."

Seo-jin adjusted the filter. The render sharpened. One plane cut across another at right angles. A narrow rectangular protrusion emerged from one side. A vertical void where no points resolved.

Yun said, "That looks architectural."

"Too early."

"Agreed."

That was one of the reasons the work with Yun functioned. They were capable of pattern recognition without mistaking pattern recognition for proof. The space between noticing and concluding remained intact.

Seo-jin split the display and overlaid the structured block against known room dimensions from the Loom suite. No reason for the comparison except that the shapes on screen had begun to suggest walls, and the fastest way to kill a bad hypothesis was to test it directly.

The overlay did not align.

She felt, briefly, a small disappointment. Then she adjusted scale.

The longer boundary matched first. Then the corner angle. Then the rectangular protrusion resolved against the housing of the coolant control bay behind the partition wall. She increased opacity on the overlay and both of them went still.

The render was not of the Loom room alone. It extended beyond it, through the wall, into the calibration lab where they stood now.

Yun spoke first. "No."

Seo-jin did not disagree. The word was correct as an immediate response and useless as an analysis.

She stood and crossed to the far wall with a laser measurer. Nine steps. She took a reading from the terminal bench to the opposite corner. Returned. Entered the distance manually. The corresponding coordinates in the transformed output matched within four millimeters.

She checked the height of the cabinet nearest the door. Seven millimeters off. Within the confidence spread.

Yun had not moved. "The Loom doesn't have spatial sensors."

"I know."

"It doesn't image its environment."

"I know."

Seo-jin looked again at the coordinates on the screen. The right interpretation had not arrived yet, but the wrong ones were already failing. Cameras were impossible. Hidden room-mapping hardware was impossible. External contamination was possible only in the weak sense that almost everything remained possible before enough exclusions accumulated. But the data's structure did not look imported. It looked derived.

From what?

The answer presented itself not as revelation but as the only branch left standing after the others had been cut back.

The Loom measured causal structure by tracking decoherence relationships across physical systems within its operational radius. Its models began at the smallest measurable interactions and scaled upward through inference. If the underlying substrate model became precise enough, then the macrostructure of the environment would not need to be sensed separately. It would be implicit in the relationships of the matter already being computed.

Yun said it at the same time she did.

"It computed the room."

They looked at each other then, and the moment held for half a second longer than speech required. Not surprise. Recognition. The specific kind that mattered.

Seo-jin turned back to the terminal and opened another file beneath the first.

03:54 - preliminary inference if substrate precision exceeded prior assumptions, macrostructure may be derivable from local causal modeling not detection computation

She stopped typing.

At the bottom of the raw output block, below the coordinate fields and malformed timestamps, another set of values had begun to resolve under her earlier filters. Sequential. Layered. Not spatial.

Temporal.

She did not yet know what they described. But she knew enough to identify the next gap in the model, and that was sufficient.

Beside her, Yun pulled a chair over and sat down without asking whether they were staying.

Seo-jin adjusted the screen brightness, isolated the new field, and began to diagram.