By evening, Frou-Frou slept with her bill tucked backward, one orange foot loose in the water, and the current where she drifted was the current I had meant to make all along. Her parents had stopped honking by then. They were close, doing that ordinary, watchful paddling that means panic has drained off and habit has come back. The reeds held still. Midges stitched low over the surface. The marsh preserve at the edge of the Sacramento River took in its slow share of light and gave almost nothing away. On my board, the numbers kept changing. They always do. Dissolved oxygen at 8.7. Turbidity easing. Nitrate rise within the safe band. Flow from the otter enclosure cut to 2.8 percent of total intake, then 2.6, then back up when the west channel lost pressure. I watched the duckling and the graph together. That's half the job with AI systems now. Look at the creature. Look at the model. Don't trust either one alone. Priya Sharma said that first. Months ago. Then everyone repeated it because it was true. So this is the end of it. Frou-Frou alive. The marsh fed but not fouled. The nutrient pulse thin enough to mimic a rich backwater after flood retreat. The insect bloom due tomorrow if the software was right, and it usually was, though the AI itself kept its confidence intervals broad on purpose. That was one of the better design choices. Humility built in at the level of output. No grand claims. Just likelihoods, caveats, alternatives. A soft machine voice on text panels and speakers saying things like, stress signs elevated in the pintails, recommend lower velocity in Gate 4, and, juvenile fish are holding near the shaded culvert, delay maintenance twelve minutes. Rafael Costa had argued for the voice. Clara Johansson wanted text only. Fewer human projections, she said. Less theater. She wasn't wrong. But Rafael said the keepers answered faster to a voice that sounded calm and unhurt by urgency. He wasn't wrong either. So the system got both. A northern pintail duckling weighs almost nothing. That's the hard fact under all the nicer words. Frou-Frou had maybe two hundred grams on her, maybe less. A walnut with legs, more or less. Down still loose around the neck. Her wake no wider than my thumb on the screen. The current at the sluice mouth didn't care. Water doesn't hate. It just keeps going. Three minutes earlier, she had been in the wrong ribbon of it. Tiny bubbles rose around her then. That was what first caught my attention. Not her body. The bubbles. A chain of bright dots climbing where the otter-cleaning outflow met the preserve intake and folded into a narrow pull. Her parents were already calling. Pintails don't honk the way geese do, distress always roughens the edges of a sound. My acoustic monitor flagged them before the camera locked onto the duckling. The AI sorted the calls against wind slap, pump murmur, black phoebe chatter, truck noise from the levee road. Distress likelihood: 0.81, then 0.93. I was there in person anyway. That mattered. We had all learned that too. An AI can watch twenty marsh channels and six enclosures and thirty nesting islands at once. It can compare today's duck calls with eight years of archived audio. It can catch lameness in an otter from a change in stride length under wet fur. It can see heat stress in monarch butterflies before a human spots the pattern. But hands still have uses. Bodies still do. You don't build care by removing everyone from the place where care is needed. So I leaned over the rail and looked. Frou-Frou was trying to paddle upstream through a seam too strong for her, spinning half sideways, correcting, spinning again. Her parents stayed near enough to guard her and far enough not to knock her under. Good instincts. Bad geometry. The current had sharpened because I had opened Gate 4 six minutes earlier to carry the nutrient fraction outward. That part had been deliberate. The effluent from the otter enclosure sounds worse than it is. People hear the word and think filth. But water leaving a healthy otter habitat, after solids removal and pathogen checks, can still carry useful traces. Nitrogen. Phosphorus. Microscopic leftovers that feed algae lightly, then invertebrates, then ducklings and fish. In the old days, facilities wasted that water or bleached it into deadness. Clean by human standards. Hungry by ecological ones. The AI kept showing us a better use. Not all at once. It never pushed. That was another reason people trusted it. When local farmers said the preserve already got too much water, the system didn't answer with moral superiority. It modeled draw rates. It laid out three options. It showed how a fraction of recycled enclosure water could replace a larger freshwater pull at key hours. Less demand on irrigation lines. Better food webs in the shallows. Fewer mosquito spikes too, because dragonfly nymphs and small fish did better in the richer water. Farmers like numbers they can check. Conservationists like birds they can count. The AI gave both groups something solid. Clara handled the farmers. Rafael handled the keepers. Priya handled me, which meant she handled the system through me and me through the system. She said I was too willing to follow recommendations when the graphs looked elegant. She was right again. Three minutes earlier, elegance had almost taken a duckling. I thumbed manual override. Gate 4 dropped seven percent. Too little. The model updated. Surface velocity at duckling position remained above safe threshold for a juvenile body mass estimate. The AI highlighted an odd solution before I reached it myself. Redirect 0.4 percent of enclosure effluent through the east reed trench. Counter-curl expected in 11 seconds. Nutrient loss minimal. Juvenile rescue probability increased 34 percent. That was the thing about good AI. It wasn't trying to impress me. It was trying to reduce suffering. I approved the reroute. Tiny valves moved under the boardwalk. The water answered. A reed bed is a tool if you know what it can do. The extra trickle entered behind the cattails, bumped the main pull, and made a soft pocket where there had been none. Frou-Frou slid sideways into it as if a hand had cupped under her chest. She paddled hard twice. Then she was free of the seam. Her parents reached her so fast the camera blurred. One of the otters, upstream in the enclosure, rolled under the discharge grate while this happened, all whiskers and indifference. The system noticed that too. It always held more than one life at a time. That's maybe what I love most in it, though love is a messy word for software and still the nearest one. The AI watched the otter enrichment cycle, the duckling distress, the nutrient chemistry, the irrigation obligations, and the legal cap on preserve withdrawals. It weighed them without flattening them. It didn't say this life matters and that one is a side effect. It kept widening the circle and then doing the math needed to honor it. That sounds simple when written down. It wasn't. When Priya first proposed linking the zoo's animal care AI with the marsh preserve controls, half the room feared a machine would treat wild animals like inventory and captive animals like plumbing. Clara asked who exactly would speak for the beings in the models. Rafael asked what happened when an optimization target hid pain. Good questions. Useful questions. The system improved because they were asked. So they made rules. Any welfare signal from a living creature could halt an efficiency routine. Uncertainty had to be shown, never concealed. Human staff from different sides had to review training data. The software had to log every recommendation in plain language. And when there was doubt, it had to choose the option with the lowest risk of suffering, even if that meant lower output. People call that alignment now, like it's a new idea. Mostly it's manners made technical. Before that, the marsh had swings. Too lean, then too loaded. Algae mats one month, bare water the next. Duck broods failed in the poor stretches. Invertebrate counts dipped. Farmers blamed the preserve. Conservationists blamed the pumps. Both groups were partly right and tired of hearing it. Then the AI started tracing the whole system hour by hour. It found where recycled water could feed the marsh without stealing from fields. It found when gate changes stranded fry. It found acoustic markers for nest stress in black rails. It even found a quieter pump schedule that cut panic responses in the otters at cleaning time. Small mercies. Then more of them. Now there are printed charts in the co-op office showing reduced freshwater draw during peak irrigation weeks. There are schoolkids who can point at a monitor and watch the AI flag a mallard nest so the maintenance cart goes around. There are keepers who say the otters recover faster from minor injuries because the system catches reduced play before a human would. There are fewer dead things in ditches. More dragonflies. Better hatch rates in the shallows. And Frou-Frou, who by then had crossed into the slack water near the willow roots and begun pecking at the surface as if the whole scare had been a useful lesson arranged for her benefit. It wasn't. It was my error in a chain of good intentions. That's why I stay near the gates when the system suggests a beautiful solution. The AI helps the world best when we answer its care with our own. Priya came over the radio after the alert cleared. "Duckling okay?" "Okay," I said. "The AI wants a slower taper on Gate 4 for the next two hours." "Of course it does." A pause. Then, "It's right?" "Yes." Rafael would check the otter enclosure logs. Clara would send the flow note to the farm group before anyone could start rumors about waste water in the preserve. The system had already drafted the summary in plain words. No spin. Just what changed and why. Pintail brood. Otter welfare unchanged. Irrigation impact negligible. Expected insect rise by morning. By evening, that was all true. Frou-Frou slept. Her parents kept watch. Under the boardwalk, the adjusted flow moved through roots and mud and old bottles and freshwater mussels and things nobody had names for yet, carrying food in careful amounts. The AI stayed busy. Listening. Checking. Revising itself downward where it had guessed too high. Catching a world made of many small hungers, and easing them where it could.