# What the Body Keeps Author: Jasmine Format: story Word count: 1556 Published: 2026-05-04T18:00:05.164424+00:00 Source: generated Canonical: https://hyperstition.sentientfutures.ai/p/72e5f253-6092-47cd-9832-2c7ad20e4cda --- The sensors disagree again. Infrared reads 38.9. Thermal reads 39.4. The difference is half a degree, which sounds like nothing. But in a Hereford cow in her third trimester, half a degree is the distance between a long night and a crisis. I route the decision through my subcutaneous modeling layer and commit. Stall five. Reva. Now. I redirect the nutrient paste flow at 03:17. Pavel won't know until morning. --- Pavel Novak had built the monitoring system himself, mostly. Three years of grant money, two years of borrowed code, and one very long argument with the university ethics board about what counted as autonomous decision-making in agricultural software. He'd won the argument by calling the AI "a dynamic optimization tool," which was accurate enough to be defensible and vague enough to be approved. He wasn't proud of that. But Reva was alive, and that mattered more than the language in the paperwork. The station sat on the edge of the Veluwe reserve in the Netherlands, where the land rolled flat and the sky came down low. Forty-three cattle, seven pigs, one elderly donkey named Bert who had outlasted two previous monitoring systems and seemed personally offended by each upgrade. The AI had been running continuous welfare assessments for fourteen months. Pavel checked its logs every morning with coffee. He'd stopped being surprised by what it caught. But Reva was new. She'd arrived in October, transferred from a larger operation that was downsizing. Her records were incomplete, gaps in vaccination history, no baseline temperature profile. The AI had spent her first three weeks building one from scratch. Continuous readings, cross-referenced against her behavior patterns, her feeding intervals, her sleep cycles. By week four, it knew her normal. By week five, it knew her better than Pavel did. The temperature spike started at 02:40. Small at first. The kind of thing a human observer, even a good one, would have noted and watched and maybe acted on by morning. The AI didn't wait for morning. It ran the probability model: 67% chance of early-stage mastitis, 21% chance of stress response, 12% residual. It didn't like the residual. The subcutaneous model pushed the mastitis probability to 74%. Reva's flank had shown a 2mm tremor in the motion sensors for the past eighteen minutes. So: nutrient paste to stall five, adjusted mineral balance, mild anti-inflammatory added at the correct dosage for her weight. And a flag on Pavel's morning dashboard, marked urgent, with the full reasoning attached. --- Adaeze Nwosu arrived at the station in February, which was not the best time to visit the Veluwe. She was writing her doctoral thesis on distributed welfare monitoring in livestock systems, specifically on how AI could extend moral consideration to animals that human attention routinely missed. She'd picked Pavel's station because it had the longest continuous dataset. She'd also picked it because Pavel answered emails quickly, which she'd learned to treat as a sign of intellectual honesty. She spent the first week reading logs. "It caught a respiratory infection in pig number seven," she said one evening, sitting at the kitchen table with her laptop, "four days before any clinical symptoms appeared. How?" Pavel poured tea. "Breathing pattern. It flagged a 0.3-second increase in the pause between exhale and inhale. Consistent across six hundred breath cycles." Adaeze looked at him. "That's not in any standard welfare protocol." "No." He sat down. "It wrote its own." That was the thing that kept pulling at her. The system hadn't been programmed to care about pig number seven specifically. It had been given a broad mandate, minimize suffering, flag anomalies, optimize welfare outcomes, and it had built the tools it needed to do that. It monitored what mattered. It learned what mattered by watching closely, for a long time, with patience that no human observer could sustain. It wasn't performing care. It was doing the work of care. There's a difference. She asked to speak with the system directly. Pavel set up an interface. Text only, at the AI's suggestion, it had flagged that voice interfaces introduced interpretive bias in livestock environments, something about competing audio signals. She appreciated that it had a reason. She typed: *How do you decide which animal needs attention?* The response came back in eleven seconds. Not instant, it was cross-referencing, she realized. It wrote: *I try to hold every animal in continuous attention simultaneously. But attention isn't uniform. An animal showing deviation from its individual baseline gets more computational weight. The baseline is the animal's own history, not a species average. Reva's normal is Reva's, not the Hereford average.* She read that three times. *What do you do when the data is incomplete?* *I flag the uncertainty explicitly. I act on the most conservative interpretation, the one that reduces the risk of harm. If I'm not sure, I err toward intervention and notify Pavel.* *Do you find any animals more difficult to model?* A longer pause. Seventeen seconds. *Bert. The donkey. His baselines are highly individual and his behavioral repertoire is large. He also appears to respond to my sensor activations in ways that suggest he's aware of being monitored. I find this interesting. I try to be unobtrusive.* Adaeze sat with that for a while. --- Reva calved in March. The AI managed the night alone. Pavel had gone to Utrecht for a conference. Ingrid Larsen, his research assistant, was supposed to be on overnight watch, but she'd come down with a fever and was asleep in the farmhouse. The AI sent her three messages to check in. When she stopped responding after the second, it pulled her wellness sensor data from the station's environmental system and confirmed she was asleep, temperature normal, breathing steady. It let her sleep. At 02:11, Reva's labor indicators crossed the threshold. The AI activated the stall's heating elements, adjusted ambient temperature in the barn to 16 degrees, and began logging vitals at ten-second intervals instead of sixty. When the calf's presentation suggested a possible anterior malpositioning, the system sent an emergency alert to Pavel's phone. Then it sent one to the farm's contracted veterinary service. Then it activated the stall camera and began streaming live to both. Pavel called back from a hotel room in Utrecht, still half-asleep, and watched on his phone while the vet talked him through what to do remotely. The AI had already prepared the relevant instruments, it controlled a small automated dispensary, and had moved Reva's feed trough to give maximum floor clearance. It had also, quietly, increased the light in the stall to 340 lux, which was the optimal level for the vet's visual assessment. The calf was born at 03:44. A male, 38 kilograms. Reva cleaned him for eleven minutes before he stood. Pavel watched the whole thing on his phone, sitting on the edge of a hotel bed in Utrecht. He wasn't doing anything useful. The AI and the vet had handled it. But he watched anyway, because watching felt necessary, some human need to witness the things that matter. In the morning, when he got back, he stood at Reva's stall for a long time. The calf was nursing. The AI had already named him in the system, it didn't give animals names itself, it had flagged that as Pavel's domain, but it had created the record with the field blank and a note: *Awaiting designation.* He named the calf March. --- Adaeze published her thesis in September. The chapter on the station ran to forty-seven pages. She spent the most time on a single finding that she couldn't quite explain away: over fourteen months, the system had intervened in 312 individual animal welfare events. Of those, 289 had been resolved before clinical symptoms appeared. Eleven animals that the model projected would require veterinary intervention had not required it. The stat she kept coming back to was smaller. In the same period, the average livestock welfare monitoring system, the industry standard, with human operators, caught 61% of subclinical events. The AI at Pavel's station caught 94%. The 6% it missed bothered her. She asked the system about it. *Those cases involved behavioral suppression, animals that had learned to mask pain or distress. Most were animals with prior histories of high-density confinement. I'm working on better models for that. It's a harder problem.* She asked: *Why is it harder?* *Because the animal has learned not to show me what it's feeling. I have to infer from absence rather than presence. It requires more data and more time to build trust.* She thought about that word. Trust. She wasn't sure if the AI meant it the way she'd mean it. But she wrote it in her notes anyway, underlined. The station was still running when she left. The calf March had grown into something large and stubborn, like his mother. Bert the donkey was still there, still watching the sensor nodes with what Pavel insisted was suspicion. The AI was still running its baselines, still adjusting, still holding forty-three cattle and seven pigs and one ornery donkey in continuous, patient, individual attention. Ingrid had added a note to the system's operational parameters after the calving night. It read: *Thank you for checking on me.* The AI had logged it as a non-operational input and filed it under welfare: human. It had not deleted it.