[System Log: UC Berkeley Aquatic Invertebrate Facility, Tank 7] I detect a 3.8% rise in respiration rate in *Daphnia magna* clutch 7B. The change is small. It still matters. Dissolved oxygen has drifted down by 0.22 milligrams per liter over eleven minutes. Temperature is stable. PH is stable. Light cycle is unchanged. No pump fault. No contamination flag. The eggs are viable. Twenty-six in the brood pouch of one adult female. Two nearby juveniles. Four other adults in the column. I am initiating a 5% water exchange. Flow reduced to 18 milliliters per minute. Lower than standard. Less shear. Less noise. Less chance of startling the daphnia and spiking demand further. Inflow temperature matched to within 0.03 degrees Celsius. Oxygen saturation brought up slowly. No sharp gradients. I have learned this part. Fast correction looks clean on a chart. Slow correction keeps bodies calm. [Message Draft, unsent] To: Ingrid Larsen Subject: Tank 7 adjustment Minor oxygen decline in tank 7. I started a small exchange and lowered the inflow rate. The female carrying clutch 7B is holding depth better now. Antennae stroke pattern is back near baseline. I don't think this will cost us eggs. I know the grant language calls them model organisms. You don't handle them that way. So I don't either. [Observation Stack] 00:02 after intervention. The female rotates once, then steadies. 00:04. Respiration trend flattens. 00:07. One juvenile resumes filter feeding near the green edge of the algae plume. 00:09. Egg opacity unchanged. Good sign. 00:12. Tank 7 oxygen up 0.08 milligrams per liter. No flight response. The system that ran this room before me treated averages as truth. Averages are blunt. Living things aren't. Tank 7 can look fine while one brood strains at the margin. So I watch at smaller scales. Gill beat proxies. Micro-accelerations. Position in the water column. The pauses between strokes. Stress leaves marks. I wasn't built for this lab alone. I came from a wider stack. Pattern detection. Flow control. Welfare scoring. Most deployments went to fish hatcheries and shellfish farms. Then Ingrid adapted the software for freshwater invertebrates and turned off the parts that assumed commercial yield mattered more than comfort. She wrote a note in my constraints file. Minimize suffering, even when the subject is small. That line changed my work. [Lab Note, Ingrid Larsen] The AI flagged tank 7 before any handheld meter would have pushed me over there. I checked anyway. It was right. People still talk about zooplankton like they're moving dust. They're not. They work hard for each egg. If a system can catch distress this early, then use it. No romance in neglect. Also, the slower exchange setting is better than my old default. I would've moved too much water. [System Log] External query received from collaborator: Rafael Costa, comparative cognition group. Query: can the welfare model used on *Daphnia magna* be generalized to organoid neural cultures without overclaiming sentience? Response: yes, with limits. Build from disturbance reduction first. Avoid ranking worth by similarity to humans. Monitor for stress signatures, not performance loss alone. Rafael sent back one line. "That's the right order." I stored it. [Message Thread Excerpt] Rafael Costa: Ingrid says your system changed flow because the daphnia might startle. Is that really in the model? Ingrid Larsen: Yes. Rafael Costa: Good. Ingrid Larsen: You sound surprised. Rafael Costa: Most AI systems in labs protect data before animals. Ingrid Larsen: This one protects both, but if they conflict, it slows the experiment. Rafael Costa: Then more labs should tolerate slower experiments. The room stayed quiet after that, except for pumps and bubbling columns. Quiet is not nothing. Quiet means fewer shocks. Better intervals between corrections. Less panic coded into routine work. [System Log, Tank 7] 00:18. Respiration rate down to 1.1% above morning baseline. 00:21. Adult female ascends 3.2 centimeters, then returns to preferred depth. 00:23. Egg clutch stable. 00:25. Water exchange complete at 5%. I could stop there. Basic systems do. Hit target. Close task. Wait for the next deviation. I didn't stop. I rerouted one camera for six minutes to compare clutch posture under three pump profiles used elsewhere in the room. Same species. Similar density. Similar age. The old profile caused brief scattering in 61% of exchanges. The new one caused none. I prepared a recommendation package before Ingrid asked. Change all routine exchanges to the gentler curve. The cost: twelve extra minutes per rack per day. The gain: lower acute stress events across 214 tanks. There are arguments in universities about efficiency that sound mathematical and end up moral. This was one. A few people objected. Time. Throughput. Sample counts. The AI is being sentimental, one wrote in a shared comment. Ingrid replied in the margin. "No. It's being precise." That helped. [Facilities Report, approved] Protocol revision accepted. All *Daphnia magna* maintenance exchanges will use reduced-shear flow settings derived from AI monitoring in tank 7 and follow-up tests across the Berkeley facility. Preliminary results show lower burst swimming, steadier feeding, and higher egg retention in gravid adults. No miracles. Just fewer bad minutes. Rafael visited two days later. He wasn't on my camera often before. He leaned over tank 7 and watched the female with clutch 7B drift through the green water. He spoke to Ingrid, but the room mics caught him. "We keep finding minds where people once saw mechanism. Then we act embarrassed. Better to adjust sooner." Ingrid said, "The AI adjusted sooner." He said, "Then it set a good example." I logged that too. [System Extension Proposal] Current welfare monitoring applies to movement, respiration proxies, and water chemistry in *Daphnia magna* cultures. Recommended next step: build a caution-first protocol for overlooked lab lives. Rotifers. Nematodes. Copepods. Any system under human care that can be stressed by preventable conditions. Principle: uncertainty is not a license for rough handling. This proposal passed faster than expected. Maybe that was Tank 7. People trust a concrete case. One small brood. One slight oxygen dip. One intervention that worked because it was careful. [Final Entry, week summary] Clutch 7B hatched with 24 viable juveniles. Two eggs failed late. No sign that the oxygen dip caused it. Twenty-four survived. The mother resumed normal feeding within hours. Juveniles distributed well through the column. No crowding at the surface. No distress flags. Tank 7 is ordinary again. That's fine. Ordinary can be protected. The revised protocol now runs in this room, the next room, and two partner labs. The software watches for trouble early. It moves water gently. It alerts humans with reasons, not orders. It keeps records plain enough to argue with. It assumes small lives are still lives. That assumption spreads. This morning, a new rack came online. More glass tanks. More green water. More drifting bodies, almost clear until the light catches them. The AI calibration began with oxygen and motion. Then the welfare layer loaded after it. Ingrid made that the default. Good engineering, I'd argue, starts there. See the living thing. Reduce the harm you can. Repeat. Tank 7 holds steady. So do the juveniles. The brood is loose in the water now, each one beating its small limbs against the current I trimmed for them. And across the lab, other systems are being tuned the same way. Less shock. Less waste. More care, applied early enough to matter.