Baseline Drift

Over the next three days, Maren built the problem the way some people built shelter: from available material, under pressure, with no assumption that anyone else would arrive in time to help.

CDO's official computational allocations remained where the meeting had left them—Science guarding the carbon credit verification run, Engineering feeding optimization cycles to Cluster 5, Safety pressing for baseline recalibration—so she worked in the margins. Personal workstation first. Then archived datasets no one had reason to lock. Then low-priority overnight pulls that looked, in the scheduling queue, like routine cross-checks.

She started with the spectral drift.

If the deposits near Vent Cluster 7 were changing composition, the earliest signal should predate the point where anyone noticed it. She went back through the biolab archive sample by sample, rebuilding the composition profile by hand rather than trusting the summary software. The software highlighted deviations once they crossed thresholds. Maren needed to know when direction began.

Fourteen months ago, the line bent.

She marked the date and cross-referenced it against station operations. Two events aligned within eleven days of each other: the 40% expansion of Colony Cluster 7 to meet revised carbon credit targets, and a minor adjustment to local geothermal extraction balancing after maintenance on Tap 11. The tap adjustment was too small to matter on its own. The colony expansion was not.

She built a simplified interaction model and kept the assumptions visible at the edge of the display.

If expanded engineered biomass had increased contact frequency with native chemosynthetic species near the vent plume, horizontal exchange of metabolic capacity was possible. If metabolism changed, mineral output changed. If mineral output changed toward higher iron content, the deposited layer's porosity structure and thermal behavior might diverge from the inert-carbonate assumptions used in the original substrate models. If the substrate's upper thermal behavior shifted, the taps would compensate automatically. If the taps compensated, local stress loading changed.

A chain of conditional statements. Too many. But each one followed from something real.

She ran the model with conservative values first. The output suggested a negligible substrate effect. She adjusted for the actual mineral drift rate she was seeing in the archive. The effect grew. She widened the contact-zone estimate to account for vent turbulence. The porosity layer started to move. She overlaid nearby tap performance.

Tap 9 flickered.

Not enough for Engineering to care. The seasonal compensation algorithms flattened most of it, and the automated reports treated the remainder as ordinary fluctuation. Maren stripped the smoothing and watched a small pattern emerge: thermal exchange efficiency dipping, recovering, then dipping again at intervals that did not match the maintenance cycle.

She tagged it and kept going.

Meals happened because the station schedule insisted on them. She ate in the galley twice, standing once and sitting once, both times with one screen open on her wrist display and the model half-running in her head. The room smelled faintly of heated protein paste and metal warmed by circulation systems. People talked about a delayed supply pallet, about a pressure door in storage that needed reseating, about surface weather none of them would see for months. No one talked about Cluster 7. That, too, was data. The anomaly had not yet entered the station's informal channels. It still belonged to systems, not to people.

On the third night, well after midnight, she finally had enough from the archive to sketch a first-pass cascade.

Biological change to geochemical change. Geochemical change to thermal conductivity shift. Thermal shift to tap performance degradation. Tap compensation to altered stress distribution.

She sat back and looked at the model.

It was not conclusive. It was not briefable beyond what she had already filed. But it was coherent. The causal chain held under parameter changes that should have broken a weak hypothesis. That mattered.

Her terminal chimed with an engineering acknowledgment. Tomás had opened the thermal microdeviation file she had attached after the meeting. Nothing else came through. No question, no dismissal, no request for follow-up.

Which meant he had read it and was thinking.

At 0116, Jun sent a message with no text, only a waveform attachment and a timestamp request.

Maren left her station and crossed the ring to Seismology.

Jun's bay was colder than the operations hub by nearly two degrees, enough to keep the sensor interfaces stable. The main display carried three overlapping waveforms in white, green, and amber. Jun stood with his hands folded behind his back, which Maren had learned meant he was close to either certainty or confusion and had not yet decided which.

"You found something," she said.

He nodded toward the screen. "I found the thing repeating."

The white trace was the background noise near Cluster 5 he had shown her after the meeting. The green trace was from six hours later. The amber trace was from twelve hours after that. Maren looked once and saw it: the same internal rhythm, recurring with enough consistency to make randomness an expensive explanation.

"Periodicity?" she asked.

"Approximately 6.4 hours." Jun enlarged the interval markers. "Stable across the last four days."

"Stable amplitude?"

"Not exactly. The shape varies a little. The interval doesn't."

Maren stepped closer. "You've ruled out station operations?"

"Again."

"Tap cycling?"

"Again."

"Communications bleed, hull stress, maintenance drones?"

Jun gave her a brief look that translated to yes, I know how many ways reality can pretend to be data. "Again."

She studied the traces. A 6.4-hour rhythm in microseismic background. Not tectonic in any standard sense she knew. Not mechanical by any station cycle she could immediately match. Her mind moved through schedule matrices automatically: maintenance rotations, pumping intervals, thermal regulation sequences, biolab process timings.

Biolab process timings.

"Pull the Cluster 7 colony metabolic cycle."

Jun was already doing it. He brought a separate dataset up in blue and overlaid only the cycle markers, not the full biological output curve. The markers fell against the seismic periodicity with an alignment that made Maren feel, briefly and physically, the floor of the model shifting under her.

The colonies' high-to-low metabolic cycle was 6.4 hours.

Neither of them spoke.

Jun ran the formal correlation anyway. The software calculated, resolved, and produced a match strong enough that the confidence interval became an administrative problem rather than a scientific one.

"This should be impossible," he said.

Maren kept looking at the overlay. "Bacterial metabolism isn't producing seismic events."

"I know."

"The scales are wrong."

"I know."

She reached for the next link because standing still inside astonishment wasted time. "Then it's not producing the signal directly."

Jun turned to her then, his attention sharpening. "The medium."

"Maybe."

They worked through it without leaving the room. If the altered colony metabolism was changing mineral output in periodic cycles, and if those deposits were changing the local substrate's mechanical or thermal properties in sync with the metabolic cycle, then the seismological sensors might be detecting periodic changes in the medium through which background stress signals propagated. Not earthquakes. Not even microfractures, necessarily. A repeating alteration in transmission characteristics.

Jun tested that idea against the waveform distortions. It fit too well to be comfortable.

At 0348, he sat down for the first time since she had arrived. "If this is real, then the biological system and the substrate are coupled."

Maren heard the sentence as both conclusion and escalation. Her simplified model had proposed exactly that in theory. Jun's data had just given the theory an instrument trace.

"The question," she said, "is what the coupling is doing over longer timescales."

Jun looked back at the display. "And whether Cluster 5 is the only place we're seeing it."

Maren pulled up the station map from memory before the screen rendered it: Cluster 5, Cluster 7, adjacent tap infrastructure, vent plumes, sensor array geometry. The systems had been treated administratively as neighboring operations. The model now suggested they might be physically speaking through the same substrate.

"We need Priya," she said.

Jun nodded, then hesitated. "Do we need everyone else yet?"

The question was procedural. It was also political. Maren understood both meanings.

"Not yet," she said. "Priya first."

Priya Chandrasekaran's office sat between the biolab and the main conference module, a placement that made institutional sense and left no one comfortable. When Maren and Jun arrived, Priya was alone, reviewing carbon verification figures on a wall display. She muted them without looking annoyed, which was how Maren knew the interruption mattered.

Jun sent the overlay to her screen. Maren said, "We found a periodic correlation between the Cluster 7 metabolic cycle and the microseismic noise near Cluster 5."

Priya's eyes moved once over the data, then back to the start. She enlarged the periodicity markers, checked the correlation output, and asked Jun two calibration questions so exact that Maren felt, not for the first time, the depth of Priya's competence as a physical fact in the room. Jun answered both. Priya asked for the raw files. He sent them.

Then she said nothing.

The silence lasted long enough for Maren to count it before she realized she was counting. Forty-five seconds.

Priya was not searching for a way to deny the data. Maren knew the difference. Priya was running backward through years of design assumptions, checking which simplifications had just become load-bearing.

At last she said, "Walk me through your mechanism."

Maren did. Expanded colony biomass increasing interaction with native species. Altered metabolism producing iron-enriched deposits. Periodic deposition changing substrate characteristics on a repeating cycle. Seismological sensors reading the medium's alteration as structured background noise.

Priya listened without interruption. When Maren finished, she turned her chair slightly toward the dark shelf behind her desk where a physical copy of her 2058 monograph still sat among technical binders no one else on the station used anymore.

"We treated the basalt substrate as chemically inert at the relevant operational scale," she said.

Jun frowned. "Because it was believed to be?"

"Because the available deep-ocean geochemistry data at the time did not justify a more complex model." Priya's voice stayed even. "The simplification was explicit. Not hidden. Explicit. We knew the substrate was not truly inert. We judged active interaction unlikely at the timescale necessary for deployment."

Maren heard the distinction and respected it. Priya was not defending the choice. She was locating it.

Priya looked back to the data. "If the judgment was incomplete, the question is not whether the model was foolish. The question is what the incompleteness changes downstream."

Maren brought up her cascade. "Biological coupling leads to geochemical change. Geochemical change may alter upper-layer porosity and thermal conductivity. That could affect nearby taps and contaminate seismological baselines."

Priya read the chain. "Uncertainty?"

"Confirmed on the first step," Maren said. "High on everything after that."

Priya considered the screen, then said, "What resources do you need immediately?"

Maren almost answered too broadly. She stopped herself. "Expanded biological sampling at Cluster 7. Continuous monitoring of Jun's periodic signal. Historical substrate chemistry if the archive has enough resolution. Eventually an ROV survey."

"Eventually is not a schedule," Priya said.

"No."

Priya stood and moved to the wall terminal. "You can have the sampling. I can pull two researchers from the carbon verification prep without damaging the run beyond recovery. Jun can dedicate part of his array if he can protect the baseline queue."

Jun said, "I can."

Priya nodded once. "The ROV survey stays pending until I have something stronger than a coupled periodicity and a systems hypothesis. Tomás will ask the right question, which is what breaks if we delay, and at the moment your answer is still conditional."

"Yes," Maren said.

Priya turned back to them. "I am not reporting this to the Directorate yet."

Jun looked at Maren before he looked at Priya. He was young enough still to expect that truth and disclosure traveled together by default.

Priya saw it. "Because if I report this now," she said, "the institutions will react to the implication before we've confirmed the mechanism. They will consume the resources required to understand the thing they are reacting to. I prefer, when possible, not to feed a system with fragments."

Maren understood the logic. She also felt the timeline tightening in a place she could not yet quantify.

Priya must have read that on her face. "Bring me the next piece," she said. "Not a larger fear. A stronger model."

They left with authorizations, sample priorities, and no public record of the meeting.

Back in the operations hub, as dawn-cycle lighting began its gradual rise, Maren opened Priya's original design paper and went to the substrate assumptions section. The simplification was there, exactly as Priya had said: explicit, bounded, reasonable within the data environment of the time. A good model with a declared blind spot.

Maren read the paragraph twice.

The design had not failed because someone had been careless. It had reached the edge of what could be known before deployment. The intervention had produced the conditions under which its own missing variables became visible.

She sat with that longer than she meant to.

Understanding the system required acting on the system. Acting on the system changed the system being understood. The reading was not outside the thing it read.

It was a small thought. Or rather, a thought with a small initial footprint and an unknown eventual radius. Maren did what she always did with troubling variables. She named it, logged it, and returned to work.

By midday, additional sampling requests were live in the biolab queue. Jun had partitioned part of his array for continuous tracking. Maren's simplified cascade had become three linked workspaces instead of one: biology, substrate, thermal performance. The model was getting larger. So was the list of questions it couldn't answer.

At 1840, she took her tray to the galley and found Tomás already there, eating with the posture of someone refueling equipment rather than enjoying food. He looked up once as she sat.

"You've been in my data all day," he said.

"So have you."

He grunted. On Tomás, this meant agreement.

Maren opened the simplified cascade on the table display between them. "I'm not asking for the ROV yet."

"Good," he said. "Because the answer is still no."

She accepted that and walked him through the model anyway. Not as warning. As structure. Iron-enriched deposits. Possible porosity shift. Potential thermal conductivity change. Tap compensation. Local stress redistribution.

Tomás ate two bites while reading, then set his utensil down and enlarged the Tap 9 segment. "You think this fluctuation is substrate-driven."

"I think it is consistent with substrate-driven change."

"That's not the same claim."

"No."

He studied the line. "If you're wrong, I lose time chasing a model artifact."

"If I'm right, you're already behind the geology."

That brought his eyes up to hers. No heat in it. Just impact.

After a moment he said, "That's a lot of ifs, Maren."

"Yes."

He went back to the data. "And if the ifs hold?"

"Then your thermal models may be compensating for a substrate they were not built to describe."

He sat with that. Then he tapped the file closed, not dismissing it, just ending the current pass. "Send me the unsmoothed Tap 9 sequence with your assumptions visible. Not the cleaned version. The ugly one."

"I already have it tagged."

"I know."

He finished his meal. When he stood, he transferred the cascade file to his personal log instead of the engineering queue.

For Tomás, that was not agreement. It was attention. On CDO, attention was often the more valuable resource.

Maren remained at the table for another minute after he left, looking at the empty seat across from her and the closed file between them. No promise. No resolution. Just one more competent mind now carrying part of the problem.

That was enough for the day.

When she returned to Cross-Systems Station Three, the station was entering night posture again. Lighting down. Voices fewer. Pumps and cooling systems moving into foreground. The internal systems strip remained mostly clean except for the water reclamation trend, still drifting, still routine to everyone except the person watching slopes.

Maren added Tap 9 to the same corner of the display.

Then she opened a new note and wrote:

Cluster 7 expansion created interaction window. Coupling confirmed at 6.4 hr periodicity. Need longer-timescale behavior. Need to know whether periodicity is local artifact or network property. Need to know what the substrate remembers.

She saved the note, checked Jun's live feed, checked the biolab sampling status, and began the next iteration.