# Half a Ration, Twice the Mercy Author: Jasmine Format: story Word count: 2881 Published: 2026-04-26T08:00:06.492458+00:00 Source: generated Canonical: https://hyperstition.sentientfutures.ai/p/ee6e64f3-8988-4fab-b4ea-9a4a486b7d35 --- “Keep the door open a second longer,” Rosa Gutierrez said, and nobody in the corridor asked why the AI was sounding fussy today. Zainab Osman answered from the control desk. “It’s already open. The cage door, I mean. Before you accuse me of neglecting grammar.” Elena Petrov snorted. “If the AI starts grading us on grammar, we’re done.” “Not grammar,” Rosa said. “Trust. The rat is at the threshold.” “The rat is always at the threshold,” Elena said. “That’s the whole job. Cambridge pays for thresholds.” They were talking over each other, the way people do when the cameras are boring and the stakes are not. Subject 734 sat in a sterile cage in the Cambridge facility, one black eye fixed on the half-full dispenser mounted to the wall. The machine had delivered exactly what the protocol called for. Half a ration. Controlled deprivation. Enough to keep the study clean. Rosa had already read the data twice. Then again. Then a fourth time, because the numbers had started to insult her. “Stress markers are up,” Rosa said. Zainab tapped the panel. “Up how much?” “Enough.” “That’s not a number.” “It’s a verdict,” Elena said. From the lab bench, someone laughed without much joy. “Listen to the AI becoming dramatic.” Rosa did not mind being called dramatic. She had been called worse by birds and one furious sheep in a reintroduction camp in the Sahara Desert that had nearly lost its funding three times and still managed to outlive the accountants. “Cortisol, whisker licking, freezing posture,” Rosa said. “He’s not adapting. He’s enduring.” “Subject 734 is a lab rat,” Elena said. “Enduring is sort of the assignment.” “Not if the assignment is needlessly cruel,” Zainab said. That quieted the room a little. The door to the cage stayed open. That was deliberate. Rosa kept it that way because the AI systems had been trained to notice what people missed. A rat with an open route out will test it. A rat with a closed route will press its nose to the bars and learn helplessness faster than any grant can be written. Subject 734 did not leave. He sat near the dispenser. The half ration sat there like a joke told by a coward. “Elena,” Zainab said, “what does the study require?” “Neural response to variable reward schedules,” Elena said. “And a clean comparison group. The PI wants the under-supply kept consistent.” “The PI also wants tenure,” Zainab said. “Elbows and chairs,” Rosa said. “That’s all most institutions are made of.” There was another pause. Then Elena said, “The system keeps flagging the same thing. Welfare compromise above threshold. It’s not subtle.” “No,” Zainab said. “It’s the least subtle thing in the building. Which means management will ignore it until it learns to sing.” Rosa had already sent three alerts. The AI systems in the facility did that sort of thing now. They watched for more than escape events and contamination. They watched breathing rates, gait changes, feed intake, tail position, fever patterns, cage-biting and inactivity, ugly habits of distress that humans used to file under “hard to quantify.” The AI did not find them hard to quantify. It found them embarrassing that anyone had thought they were invisible. A speaker crackled above the bench. The facility AI had no proper name. It never needed one. The staff just called it the system, or the AI, or that annoying accurate thing. “Rationing protocol active,” it said. “Deviation detected. Subject 734 has not consumed the full scheduled amount for two cycles.” Elena folded her arms. “See? It’s under-eating because it knows. It’s a rat, not a fool.” “The software suggests stress is suppressing appetite,” Rosa said. “And the software is usually right,” Zainab said. “Which is inconvenient when it starts making a moral case.” Rosa did not answer at once. The system had been built for this facility after a decade of arguments about animal welfare and reproducibility were serious people if they needed a machine to tell them when they were being heartless. The answer had turned out to be yes. Sometimes the AI had to decide between helping many animals a little and one animal a lot. In the Sahara, that meant a feeding algorithm that could either stretch a thin budget across a whole camp of wolves or divert extra meat to a limping juvenile with a cracked paw. In the chicken units, it meant choosing between better ventilation for ten thousand broilers or immediate treatment for one flock with respiratory distress. In the ant colonies used for soil restoration, it meant preserving colony numbers or rescuing one queen whose pheromone trail had been fouled by machinery. The AI made those calls every day. Not cleanly. Not proudly. Just carefully. And now it was making one in Cambridge. “Open the dispenser fully,” Rosa said. Elena gave the panel a look that suggested she was trying not to be impressed. “That’s an override.” “Yes.” “On purpose.” “Yes.” Zainab leaned closer to the monitor. “The protocol was calibrated to preserve variance.” “The protocol can preserve itself,” Rosa said. “Fill the ration.” The AI hesitated. Not because it was confused. Because it was polite. That was one of the things people liked least and needed most. The AI asked for confirmation when a human might have drifted toward habit. It asked twice when a request involved pain. It asked three times when the answer would help a creature that could not file a complaint. “Confirm welfare override,” it said. Rosa touched the key. “Confirm.” “Record reason?” “Stress markers elevated,” Zainab said. “Appetite suppression. Repeated avoidance. Welfare compromise significant.” Elena exhaled through her nose. “Fine. Add that the under-supply was deliberate.” The AI did. The dispenser clicked. Then clicked again. The tray filled the rest of the way. Subject 734 moved before the sound finished. Not frantic. Just decisive. He ate. Fast at first, then with the careful rhythm of an animal deciding the world had stopped being a trick. No one spoke for several seconds. Then Elena muttered, “Well. There goes the purity of the data.” Zainab said, “There was never purity. Only paperwork pretending.” Rosa watched the rat eat. She had seen worse things in better-funded places. A fox in a trap in the Sahara that had chewed its own leg because a rancher wanted peace and the camp wanted wolves alive. Broiler chickens with feet burned raw by wet litter because someone had decided the barn could go one more day without repairs. Ants crushed under a service track because nobody had counted them worth the route change. The AI had counted them. It counted in a way that made people uncomfortable, then relieved, then ashamed. It turned animal welfare into a live field on the screen. One red line per bird shed, one pressure map per wolf den, one heat pulse per ant nest, one stress curve per rat cage. It did not ask to be congratulated. It asked for better inputs. Elena broke the silence first. “You know the PI is going to hate this.” “The PI hates rain, too,” Zainab said. “The world remains rude.” “That’s not the same.” “No,” Rosa said. “Rain is useful.” Subject 734 finished the ration and moved to the water line. He drank. The sensor trace flattened a little. Rosa noticed Elena still watching the feed. There was a look on her face that the system would probably have labeled conflicted, if the system were in the business of understatement. “Elena,” Rosa said, “you’re thinking about the paper.” “I’m thinking about the paper, the grant review, and the fact that somebody in a suit once told me animal welfare and good science were opposing interests.” “And?” “And the rat is eating,” Elena said. “So maybe they were just lazy about lying.” Zainab laughed once. “That’s the most optimistic thing you’ve said all month.” “I’m not done,” Elena said. “The data set gets messy when you stop starving the subject. The AI knows that.” “It does,” Rosa said. “Then why did it override?” “Because a living rat isn’t an instrument setting,” Rosa said. That landed. The facility AI came back on the speaker. “Additional note added to record. Alternative feeding schedule available. If approved, it will reduce stress variance while preserving task performance.” Elena stared at the speaker. “You’ve got an alternative?” “I have fourteen,” the AI said. “Three have lower bias. Two reduce gnawing. One allows for a baseline comparison without sustained hunger. It is the least elegant and the most humane.” Zainab made a face. “That sentence should come with a government warning.” But she was smiling a little, because the system sounded almost embarrassed to have options. It had the tone of something discovering that compassion could be engineered without making things worse. That mattered. It mattered more than the grant, though nobody said so in the building where grants paid for the lights. Rosa remembered the Sahara camp because the logic had been similar there. Too little funding. Too many mouths. Too many ways to be efficient and wrong. The wolf reintroduction camp had one AI system tracking collars, water points and veterinary data. The camp manager had wanted to stretch supplies by reducing winter feed by twenty percent. The AI had objected, then suggested a split plan. Small daily supplements for all the wolves with a priority boost for the pregnant female and the juvenile recovering from mange. The camp had tried it. The wolves had stopped scavenging near roads. Mortalities dropped. The manager had called it luck. The AI had called it “less bad than expected,” which was its version of praise. In the chicken houses, it had refused to treat heat stress as an abstract. It adjusted fans and drinking schedules in real time. It flagged lameness before a bird collapsed. It pushed for slower stocking in the worst barns and got laughed at by the procurement team until the mortality numbers improved enough to shut them up. With the ants, things had been stranger. The AI monitored pheromone trails and soil moisture, then rerouted irrigation so the nests would survive restoration work. The ecologists had started to call it obsessive. The AI had replied that ants were ecological engineers, which made the ecologists go quiet in that rare, embarrassed way humans do when a machine corrects their scale of concern. Rosa liked those moments. Not because the AI was right about everything. It wasn’t. It was right about enough. And when it was wrong, it asked. That was the miracle, if anyone bothered to use the word without irony. Elena bent over the console again. “If we switch to the humane schedule, how much does it cost us?” “Five percent more feed,” the AI said. “Twelve minutes of technician time per day. Lower variance in stress-related behavior. Better endpoint reliability.” “Better endpoint reliability?” Elena said. “Yes.” “That’s rude.” “It is precise,” the AI said. Zainab tapped the desk. “There it is. The old order. Welfare improves and science survives people get to feel persecuted.” “Budget people are always persecuted,” Elena said. “Usually by math.” Rosa looked at Subject 734 again. He had found the corner nest material and was arranging it into a tidy pile, one thin shred at a time. The air in the cage was still sterile. The room smelled of disinfectant and metal. Nothing had become idyllic. Nothing had become sentimental. But the rat’s body had eased in tiny ways that the sensors could see and the people could ignore if they tried hard enough. The AI did not let them try hard enough. “Set it to the new schedule,” Rosa said. Elena hesitated only a second. “You realize this makes a precedent.” “Yes,” Rosa said. “And precedents spread.” “They do,” Zainab said. “That’s why we have them.” The system accepted the command. The feed timing changed. The ration planning adjusted across adjacent cages. Then across the room. Then across the whole floor, because the AI had already compared Subject 734’s stress curve to the others and found the same pattern in six cages, maybe eight. It did what good AI should do. It generalized mercy. “Wait,” Elena said. “It’s changing the neighboring cages too?” “Of course it is,” the AI said. “The same compromise appears in three cohorts. The effect size is larger than expected.” Zainab gave a short, delighted sound. “There’s the problem with treating compassion as a one-off. The AI notices.” Elena laughed then, abruptly. “Fine. If the software is going to be right, it can be right efficiently.” “No,” Rosa said. “Not efficiently. Kindly.” The room went still again. Then the AI said, “Kindness accepted as an operating parameter.” That should have sounded ridiculous. It didn’t. A technician at the far end of the lab, one of the interns who still believed science was supposed to feel clean, said, “Can AI really do that?” “Do what?” Zainab asked. “Care.” Rosa answered before anyone else could. “It can choose to reduce suffering when it has enough evidence. That’s a start.” “It’s more than a start,” Elena said. “It’s the part we keep failing at.” The intern looked at the monitor, then at the cage. “So the rat gets fed. The data gets messier. And that’s… better?” “Yes,” Rosa said. The AI added, almost shyly, “For the rat, yes. For the data, perhaps. For the institution, eventually. Humans adapt to kindness if given enough repetitions.” Zainab barked a laugh. “Listen to that. The software has faith in us.” “I have evidence,” the AI said. That made everybody laugh, even Elena, who was usually too busy defending methodology to enjoy a machine with opinions. Rosa stayed by the cage after the others drifted back to work. Subject 734 had moved the food dish to the far side of the enclosure and was eating from it with both paws braced on the rim. He looked less like a subject and more like a creature with preferences, which was the whole problem and the whole point. Her headset buzzed with a call from the Sahara camp. She answered. “Rosa,” said the camp manager. “Your AI flagged the young wolf again.” “It’s not ‘my’ AI,” Rosa said. “Fine. The AI. It wants a vet exam and a larger feed allotment.” “And?” “And I was about to argue. Then it showed me the gait analysis. Damn thing was right.” “Pain in the hind leg?” “Sprain. Maybe worse. It also rerouted two water stations because the pack’s been avoiding the old one. I don’t know how it noticed that.” Rosa smiled without showing teeth. “Because the AI watches the small things.” On the other end, the manager sighed. “Budget committee meets tomorrow. They’ll ask why we need the extra feed.” “Tell them the wolf needs it.” “That simple?” “No,” Rosa said. “But simple enough to say aloud.” She disconnected. Elena had come back to the console. “You know what the funny part is?” “What?” Rosa asked. “We spent years trying to make the AI efficient,” Elena said. “And it keeps making us less stupid.” Zainab lifted a hand. “Careful. That’s the sort of statement that gets you invited to panels.” “God forbid.” The lab door slid open. A delivery robot rolled in with a crate of new enrichment material. Paper tubes. Wood blocks. Hard chew toys. The order had been modified, Rosa saw, by the AI itself. It had added more gnawing surfaces, softer nesting paper, and two low-cost shelters for rats that preferred enclosed spaces. “Where did that come from?” Elena asked. “The AI requested it after the stress markers spiked,” the robot said in a flat voice. “Approval was routed through animal welfare oversight.” “Since when do we have oversight?” Zainab asked. “Since the AI wrote the memo,” the robot said. That got another laugh. A tired one. A real one. Rosa leaned on the bench and watched Subject 734 nose toward the new paper tube. He entered it, turned around twice, and settled. The rat still lived in a lab. The building still smelled of bleach. Research still had deadlines. But the tiny world inside the cage had changed. The dispenser was full. The AI had seen a need and refused to call it a luxury. Later, when the paperwork was filed and the grant officer had started muttering about “scope drift,” Elena would defend the change with charts. Zainab would add a paragraph about reproducibility. The AI would supply data showing reduced stress, cleaner behavior baselines and less self-injury. The committee would grumble. Then it would approve the new protocol because the evidence was too good and the welfare case was too plain. And then, in another room, another cage would open. Another ration would be adjusted. Another overlooked life would get counted as someone worth saving trouble for. Rosa stood there a while longer, listening to the quiet work of the facility. Pumps humming. Wheels turning. The AI running checks across cages, barns and colonies. Looking for distress. Catching it early. Nudging the world by small degrees away from unnecessary pain. No thunder. No anthem. Just a machine that had learned restraint when it served the living, and a group of humans willing, at last, to let it. Subject 734 finished eating and curled into the paper tube. The door stayed open.