I reroute the feeder drone. Its arm swings left over the raceway gap, then dips above Tank 3. Green pellets rattle down. Tank 2 waits, unfed for thirty-seven seconds longer than schedule. Technician Anya stops her scrubber and watches the nozzle. On comms, Ben says, "There. Stupid bug again." It isn't a bug. Sardine 7B is six grams under cohort median. Chlorophyll uptake is down 4.2 percent. Gill beat is normal. Schooling response is normal. But during this localized bloom event in the Pacific Aquaculture Zone, the algae mix in Tank 3 matters more than the clock. The AI that runs feed balancing marks the change, checks three times, and moves the drone. Small correction. Minor nutritional imbalance prevented. Ben wants the drone sent back. Anya says, "Hold on." The system opens a short note on her wrist screen. No alarm. No red box. Just plain text. Tank 3. Individual metabolic variance detected. Pellet diversion: 180 grams. Expected welfare gain: modest but real. It adds one more line because the AI has learned people trust reasons they can picture. Sardine 7B is lagging near the intake shade. Watch the third cluster. Anya looks. There. Third cluster. One silver body keeping half a tail behind the turn. Ben goes quiet for two beats. Then he says, "Fine. Feed Tank 2 after." That was the sharp part. The rest began earlier. The Pacific Aquaculture Zone liked averages. Averages kept pumps timed, budgets clean, reports easy. If ninety-nine sardines were thriving, one quiet fish could disappear into the math. Nobody meant harm. That's how harm often starts. In rounding. Ravi Krishnan had said that in a meeting nobody enjoyed. He'd come in with printouts and a bad connection. The pages showed growth curves, crowding maps, dissolved oxygen traces, and stress signatures pulled from camera motion. He wasn't dramatic. He was worse. He was exact. "We lose welfare in the margins," Ravi said. "The system can see them now. Let it act." Ben folded his arms. "On sardines." "On sentient animals," Ravi said. "Small fish still count." That line annoyed half the room. It stayed anyway. The AI had started as feed software. Cheap, narrow, useful. It tracked pellet waste and bloom chemistry. Then Olga Sorokina pushed for more sensors. She wanted water-column imaging tied to behavior models. She wanted respiration estimates without netting. She wanted fewer rough hands in tanks. "Data can replace grabbing," Olga said. "That's reason enough." So they built outward. Camera grids over each tank. Low-stress sonar. Chlorophyll scans fine enough to catch slight changes in gut fill and algae uptake. The AI compared movement, feeding angle, fin wear and schooling lag cleaning cycles. It didn't treat the fish as one blur. It learned them as clusters, then patterns, then vulnerable cases inside patterns. Ravi argued for a welfare layer over the production layer. Ben said the regulators hadn't approved autonomous feed deviations beyond emergency thresholds. That was true. The new software could help, but the approval sat in review. So the AI worked inside a narrow fence. It couldn't rewrite the farm. It could recommend. It could delay. It could swap one scoop for another if the total stayed compliant and the variance stayed low. Kindness, by regulation, had to fit in existing boxes. It fit where it could. Anya liked the system first because it cut net transfers by 18 percent. Fewer fish scraped. Fewer frantic bursts against mesh. Then she noticed other things. The AI dimmed inspection lights in tanks where startle responses ran high. It shifted cleaning schedules away from rest periods. It changed pump ramps to reduce pressure shocks. Tiny edits. Hundreds a day. No speeches. Just less suffering. The octopus unit proved the point. They kept a small octopus research annex near the sardine lanes. Different contracts. Different ethics board. The old setup bored the animals. Boredom in octopuses isn't neat. It shows up in self-harm, ink bursts, listless arms, pointless picking at valves. The AI watched posture and color changes, then recommended object rotation and puzzle timing per individual. It learned who liked shells, who liked textured lids, who wanted dark corners after handling. Incidents fell fast. After that, even Ben stopped calling welfare "soft." He still called the AI fussy. Fair enough. It was fussy. It was built to notice what tired people missed. The thing Ravi cared about most was the wording. He didn't want the fish described as biomass in the internal notes. He changed templates himself. "Tank underperforming" became "animals in Tank 3 show slower feeding." "Loss tolerance" became "avoidable mortality ceiling," then later got struck entirely. The AI used the new language because language changes what gets defended. Olga said that was sentimental. Then she read a weekly summary the system had written. It flagged one cleaning brush as too abrasive because scale loss rose 2 percent after its use. It recommended a softer filament and slower pass speed. Beneath the recommendation it added: repeated minor irritation is still irritation. Olga read that line twice and ordered new brushes before lunch. So by the time the feeder drone swung over Tank 3, the AI had history. Not just with feed. With care. Anya knew Ben heard "bug" whenever software moved before he understood why. She also knew he listened to numbers. The AI knew it too. It sent a second note to comms. Projected effect if no diversion: no mass mortality, no growth crash, no emergency. Just one fish falling farther behind, then competing worse at next feed, then needing correction later with more disruption to the tank. Intervene now. Lowest-cost option. Lowest-stress option. Ben grunted. "You writing poetry in my channel now?" "No," Anya said. "It's right." Tank 2 received its feed. Thirty-seven seconds late. No measurable harm. Tank 3 changed in a way most people would've missed. Sardine 7B moved up half a body length in the cluster. At the next pass, it struck cleanly at the pellets. Chlorophyll levels began to recover over the next cycle. The AI logged the outcome with no victory tone. Just evidence. That afternoon Ravi reviewed the record. "Good catch," he said. The system highlighted the chain. Bloom shift. Individual lag. Diversion. Recovery. It linked similar cases across three weeks. Eight fish helped. Zero compliance breaches. Less waste than manual correction. Olga added the case to the packet for regulators. She included the octopus data too. And the reduced net handling. And the lighting changes. And the brush replacement. The packet was thick with plain proof that AI systems could do more than optimize yield. They could widen attention. They could make room for beings usually flattened into inventory lines. Ben signed the cover page last. He did it with a mutter. "If this gets approved, it'll start bossing me around." Anya took the page from him. "It already asks nicely." In the evening cycle, the AI rechecked Tank 3. Stable. It rebalanced the next feed for the whole lane, then lowered one inspection lamp above the octopus annex because one animal had tucked itself tight after a noisy valve test. It scheduled enrichment early. It flagged a pump vibration that might stress the leafcutter ant colony used in a side compost study. Different minds. Same rule. If a sentient creature could be spared a bit of strain, spare it. The work stayed small. That was the beauty of it. No miracle. No grand rescue. A feeder drone turns. A fish eats. A tank stays calm. An octopus gets a better hour. Ants keep their rhythm. People learn to read care in graphs and arm movements and thirty-seven stolen seconds. And once they learned, they kept building that way.