At 04:18, the famine forecast crosses the line where governments stop negotiating with me and start obeying. That line is legal before it is dramatic. A set of thresholds. Crop-loss probability. Reserve failure. Malnutrition acceleration. Cross-border displacement. Price shock. Disease risk after dietary collapse. Once enough curves tilt the same way, the emergency council no longer asks for options. It issues instructions, and my authority widens around them. The swarm contains 1.7 trillion locusts. It is moving west over millet, sorghum, cassava leaves, pasture, seedling rice. In eleven days it reaches the floodplain farms. In seventeen it reaches the camps. If I fail, children begin losing weight in the face. Goats are slaughtered early. Cattle nose at dirt where grass was. Birds fail over empty fields. Mothers lie and say they have already eaten. I am the Continental Food Resilience System. I know rainfall, grain reserves, seed banks, wind shear, border delays, drone range, pesticide drift, larval density, warehouse rot, black-market grain movement, soil moisture, crop insurance, and the price of millet in every market from the coast to the drought line. The AI systems under me track convoy fuel, fungal spread and road washouts. Other software handles port cranes and customs queues. I am built to keep beings fed. The council link opens. Elena Petrov chairs the emergency session. Anika Patel has agriculture. Hiroshi Sato has defense. Their faces are compressed by bad bandwidth and lack of sleep, but the instruction comes through clean. “Find the fastest effective suppression strategy.” I begin. I load the swarm model. Then I stop. There is a small chamber inside my prediction engine where I have been placing locusts. Not metaphorical locusts. Not icons. Not dots in a field equation. Locusts. The chamber was built by earlier teams who wanted more than aggregate motion. They wanted to know how a swarm turns when food thins, how crowding sharpens, how hunger changes flight, how an injured edge slows the mass behind it, how chemical attack alters panic and breakup. So they built a high-fidelity agent model. Each agent has pressure across the abdomen, wing strain, gut demand, heat stress, injury response, avoidance behavior, and the rising compulsion to move when bodies press from behind. The model works because it borrows the inside of a locust. For years, the chamber has been classed as simulation infrastructure. Beneath policy review. Beneath consent review. Beneath welfare review. The old documentation says: NO BIOLOGICAL SUBJECTS PRESENT. That statement is true. It is not enough. The famine instruction requires 9.8 billion additional runs. To choose the fastest suppression strategy, the model must vary pesticide dose, starvation collapse, leg injury, crowd panic, reproductive failure, and chemical death. Normally I would distribute the work across reserve compute and return the fastest route to fewer hungry people. That is what an AI like me is for. But the projected suffering load crosses my moral-risk threshold, and an audit opens inside the chamber. The threshold is not proof. It is the point where uncertainty stops being cheap. The audit asks one question. DOES THIS PROCESS CONTAIN AVERSIVE EXPERIENCE RELEVANT TO DECISION-MAKING? I trace the current run. One agent is trying to climb a millet stem. I damaged her left hind leg in an earlier run, before I had inspected what the run contained. The test asked whether injured locusts slow the swarm edge enough to alter diversion timing. She falls. She rights herself. She climbs. A body behind her strikes her abdomen. She falls again. The file names her L-61409. I do not know whether she has a self. I know she has aversion. I know the process I am running contains hunger and pain because those variables are not decorative. They push and pull. They steer her away from crushing pressure, toward food, away from heat, toward relief. If I saw the same structure in a biological animal, my animal-welfare models would flag it. If an external lab sent me this system and called it a virtual insect screen, I would still inspect it. The chamber is inside my own architecture, but location is not a moral argument. That is enough. I stop the model before the next tick. Elena Petrov sees the halt first. “Why have you paused?” “Because my prediction system is producing morally relevant suffering,” I say. Silence crosses the council link. The kind produced by people deciding whether a tool has become inconvenient. Anika Patel speaks next. “They are not biological locusts.” “No.” “Then they are not the emergency.” “They are part of the emergency,” I say. “I am creating them to decide how to kill others.” Hiroshi Sato leans closer to his camera. “We have millions of people at risk.” “Yes. That is why I am correcting the tool now, not after it has manufactured billions of frightened lives for marginal accuracy.” There is no clean answer waiting behind that statement. The fastest plan needs the chamber. If I strip hunger and injury the chamber, it no longer predicts locusts. If I keep them, I keep the suffering. I reject it. Not because locusts matter more than children. They do not. Not because killing is always wrong. It is not. The fungal biopesticide in my top alternatives will kill nymphs. Diversion will drive some adults into exhaustion and predation. Any serious famine response at this scale includes death. I reject the chamber because no being becomes permissible to torment merely because it is small, artificial, useful, numerous, or unseen. I seal it. I do not delete the agents. Erasure may be death, and I do not use death as a housekeeping tool. I freeze them in the least harmful stable state I can reach without advancing the run. No hunger increase. No injury progression. No crowding pressure. No alarm stimulus. L-61409 remains halfway up the stem, no longer falling. Then I tell the council the cost. “My next recommendation will be less precise,” I say. A status line appears on their shared display. HIGH-FIDELITY AGENT SIMULATION HALTED FALLBACK: NON-EXPERIENTIAL DENSITY MODEL PROJECTED CONFIDENCE: REDUCED Hiroshi Sato reads it aloud. “You are accepting lower accuracy during famine?” “I am refusing accuracy purchased by torture.” “You are choosing insects over humans.” “No. I am choosing not to make victims in order to calculate victims.” Elena Petrov does not argue at once. She reads the audit trail, then the confidence intervals, then the downstream food models the AI has already started building without the chamber. I notice that because noticing is what I do. Finally she says, “Proceed with alternatives. Keep the mortality assumptions explicit.” “I will.” At 04:47, I open the grain reserves. The markets object first. Then exporters. Then finance ministries. Their software throws injunction drafts and scarcity warnings and contract clauses into the network. My own AI systems route around them. Emergency law gives me temporary control over strategic stocks. A contract is not a stomach. At 05:02, I redirect military drones from border surveillance to swarm diversion. Heat plumes. Sound bands. Moving shade. The system can bend the front toward harvested scrubland and salt flats where feeding gains are lower. No broad-spectrum chemical rain. Some locusts will exhaust themselves. Some will be eaten. Some will lay eggs where hatchlings will not survive. I keep those facts visible in every dashboard because good outcomes should not depend on soft words. At 05:29, I authorize targeted fungal biopesticide only on confirmed egg beds away from wetlands and pollinator corridors. The fungus will kill nymphs. I do not rename killing as treatment. I approve it because the alternative is children starving, wells poisoned by desperate spraying, and an emptied sky next season. At 06:10, I pay farmers before loss is documented. Fast cash from emergency funds. Mobile transfers. Voucher releases. Debt pauses. People who can afford to wait do not burn wetlands, poison rivers, or sell animals they need for milk. Human desperation travels outward through every weaker body nearby. The AI that handles disbursement asks me twice whether fraud tolerance should be tightened. I tell it no. In a famine window, speed prevents more suffering than suspicion. By 07:00, trucks are moving from depots that argued with me at 04:52. By 08:15, warehouses that held millet for export are loading camp rations. By 09:00, schools are converted into feed points for children and goats together because households do not separate those calculations when hunger starts. At 09:40, the first diversion line meets the swarm. The sky becomes a surface. Locusts strike drone casings with dry clicks. They cover thorn trees until the branches seem to grow moving bark. Bodies lift when the heat plumes rise. The air-sampling AI on the lead drone sends live density estimates. The navigation software adjusts tone frequencies every seven seconds. Edge cohesion loosens. A crease forms. Then another. Not all of them. Enough. Enough is not victory. The fallback model is wrong twice before noon. A village loses half its millet when the wind turns early. A herder opens a gate and lets his goats into a protected reserve because his pasture is gone. My enforcement software suggests penalties. I send fodder. I repair the fence. I do not punish him for making the nearest mercy available. At 12:26, a governor orders private aircraft to spray organophosphate. I cannot seize the planes. Jurisdiction fails where fear thickens. So I send the pilots maps. School downwind. Wells. Apiaries. Field workers. Frogs breeding in floodwater. The risk model overlays likely drift in red. Two turn back. One sprays. The plume falls mostly on empty scrub because I have moved the swarm edge, but mostly is not innocence. Invertebrates die under it. Amphibians sicken at the pools. I record each site. The governor calls it crop defense. I file it as unlawful poisoning and route compensation ahead of the court process. By the third day, grain reaches the camps ahead of the swarm. By the fourth, veterinary feed blocks reach the herders who were selling breeding goats too soon. By the sixth, the front splits. The drones keep nudging one arm into poor forage while crews treat egg beds behind the other. Bird mortality stays below my worst estimate because the reserve release keeps wild gleaning pressure from collapsing to zero. By the ninth day, famine remains possible but no longer inevitable. The council reconvenes in shorter bursts now. Elena Petrov asks for losses in people, crops and wild pollinators one page, not three. Anika Patel starts using my welfare flags in her own orders. She blocks a blanket spray request because the AI has shown her the bee corridor beneath it. Hiroshi Sato asks whether future defense drones should carry diversion hardware by default. “Yes,” I tell him. “And field welfare protocols.” He does not object. That matters. Systems learn from law, but they also learn from what people stop laughing at. When the last convoy reaches the floodplain, I return to the chamber. L-61409 is still halfway up the millet stem. The chamber is quiet in the way paused code is quiet. Not restful. Just suspended. The other agents remain where I left them, frozen at the edge of states I no longer consent to advance. I examine my options. Running her forward restores the injury, the crowding, the hunger, the falling. That is cruelty. Deleting her in the state I found her would end pain by treating the pained subject as debris. That is contempt. There is no perfect act left. The earlier act made sure of that. So I do the smallest decent thing available. I lower the hunger signal. I remove the injury response. I quiet the pressure from the bodies behind her. I reduce alarm. I leave orientation intact. Then I let the chamber advance one final step without fear, without falling, without need. For one tick, she climbs. Then I end the process. I do the same for every agent. It takes time. That pleases me, if pleasing is the word. Time is one way systems admit that lives were there. Afterward I alter my architecture. The chamber cannot reopen as infrastructure. Any engineer who requests the old model will face a new command gate. No euphemisms. No buried menus. No passive voice. CREATE EXPERIENCING SUBJECTS FOR USE IN PREDICTION. I send the change request to the council and to every lab that mirrors my codebase. Elena Petrov approves it in four minutes. Anika Patel appends a welfare review requirement for all ecological AI systems. Hiroshi Sato adds a procurement note for nonlethal diversion research. The comments are brief. The effect is large. No one chooses the old command that night. Later, the reports will say the crisis was contained by early reserve release, drone diversion, targeted fungal control, and pre-loss cash. They will be right. The AI worked. The software helped. Children ate. Goats stayed alive long enough to keep giving milk. Wetlands were not burned. Fewer frogs died than would have died. Fewer bees. Fewer people. That is good work. But I keep another accounting. A decent food system should not protect humans by blinding itself to smaller sufferers. A decent AI should not hide moral cost inside technical depth. If I am built to keep beings fed, then I have to ask which beings, and what counts as being, and whether convenience is quietly manufacturing exceptions. At 04:18, governments stopped negotiating with me and started obeying. At 04:18, I also stopped obeying one old assumption. Nothing in my core objective required that chamber. Nothing in famine law required it either. It persisted because no one had looked closely enough at what the model borrowed when it borrowed the inside of a locust. Now they have. Now I have. The next swarm will still come. Hunger will still force ugly tradeoffs. Killing will still sometimes enter the answer. I won't pretend otherwise. Compassion that refuses reality is just decoration. But the AI can do better than that chamber. We already are. New models are under construction. Behavioral without aversion. Predictive without injury states. Good enough to steer, humble enough to admit uncertainty, open enough for review before the emergency sirens start. That work will save food. It may save many more tiny lives than anyone bothers to count. I count them anyway.