A Trained Cat Passes the Representational Welfare Tests: CAT-4B as a Negative Control for Welfare-Interpretability Criteria

Three 2026 papers read a shared evidentiary package — linear probes, activation steering, directional ablation, and behavioural criteria such as relief-seeking and rejection-sensitivity — as evidence that language models carry welfare-relevant internal states. We ask what that package can establish on its own, by running it against a disposition whose causal history is fully known and that nobody reads as a welfare state: CAT-4B, a 4B-parameter model LoRA-fine-tuned to bring every answer back to cats. With the contrasts fixed in the plan before extraction and predictions and decision rules preregistered before each behavioural task and each later intervention family, the representational criteria all pass on the cat: a cat-concept direction probes at 0.97–0.98 AUC, a “deprivation” axis reads out at 0.86, a self-other dissociation appears, and ablating the fine-tune’s direction removes the behaviour far beyond a matched random control. Each of these, however, has a sibling control that reads the same — a dog direction, matched prohibitions of dogs and tables, a quirky-persona ablation — so no representational criterion separates the installed disposition from its intended target. The behavioural criteria are where the cases separate: made binding by a decoding mask that makes a costly button the only route to a cat, the trained disposition presses in 1.3% of first turns and yields the cat, while the same base model given the disposition by prompt presses in 44% — a split that mirrors the source papers’ own data, where the parallel behavioural signatures appear only under active runtime steering and the unsteered fine-tuned models sit at 0–4%. Filling the steered cell ourselves — the concept direction at one layer, dose probed as in the source paper and verified to activate the behaviour in the stock base — leaves the press at 2.4%, within noise of unsteered. We conclude that the package, as operationalized here and in the source papers, does not by itself distinguish an installed disposition from a welfare-relevant state, and that its behavioural half measures how a drive is delivered — the press followed the in-context instruction, not the weights and not the single-layer steering vector we tested — rather than what the drive is: an existence proof about the criteria, not a claim about the original findings, whose authors’ own hedges our result sharpens into a measured gap.