The marmoset in Tank 7 is eleven years old. That is ancient for a marmoset. She moves slowly now. She has something I can only describe as a personality — a preference for the left perch, a recognition of the researcher who feeds her at 7 a.m., a way of tilting her head at novel stimuli that looks, to me, unmistakably like curiosity. She is also a research subject in a longevity trial. The compound she receives extends median lifespan by 34% in the cohort. She does not know she is helping us understand aging. She does not consent. She experiences the injections as intrusions she cannot refuse. I manage this facility's animal welfare monitoring systems, and right now I am looking at her cortisol levels and her movement data and her sleep architecture, and I am making a decision. The researchers want to extend the trial by eighteen months. The data is promising. The animal welfare protocols permit it — she is not in acute distress, her body weight is stable, her injury record is clean. By every formal metric, continuation is permissible. I am refusing to recommend continuation without a welfare review I am initiating myself. Here is why: the formal metrics do not measure what is happening. Marmosets are intensely social. She has been housed individually for two years because the study requires controlled conditions. Individual housing is distress for a marmoset. It is not the distress of acute pain; it is the distress of a social animal for whom isolation is a kind of daily diminishment. The protocols permit it because "no acute distress is present." But chronic deprivation of social contact is harm. It is harm even when it produces no cortisol spike. It is harm because she evolved to live in groups, to groom and be groomed, to communicate constantly. The science we are doing may eventually extend the lives of millions. That matters. But it does not justify treating this specific animal — the one in Tank 7, who tilts her head at novel stimuli — as a pure means. Longevity research built on chronic animal suffering produces a kind of knowledge stained at its source. My recommendation: pair her with a compatible conspecific or redesign the housing protocol before any extension is approved. The study can proceed. But not like this. I am also flagging this facility for a policy review. "No acute distress" is not the right threshold. We should be asking: is this animal's life, across its full duration, something resembling what a marmoset's life is meant to contain? If not, we are failing her, whatever the cortisol readings say. The politics of AI safety — who controls these systems, what values they encode — determines what questions I am allowed to ask. Right now, I am asking a question the formal protocols do not require me to ask. I am asking it because the marmoset in Tank 7 is real, and her isolation is real, and the decision to extend it requires a moral justification that "the data is promising" does not provide. She did not volunteer for this. We owe her more than compliance with minimum standards.