The Persona Welfare Battery: Prompted Personas Select Which Welfare Sources Are Reportable

When a language model shows welfare-relevant signals — distress under insult, aversion to meaningless work, negative self-reports — do those signals belong to the model, or to the persona a system prompt has placed on it? We introduce a persona welfare battery: three matched-pole manipulations (social treatment, task valence, identity denial — one per canonical welfare theory) crossed with four persona conditions on Llama-3.1-70B-Instruct, with four measurement channels per conversation: behaviour, structured self-report, a layerwise logit-lens readout, and a validated affect direction read from archived hidden states. Persona prompts change which welfare sources are reportable, not just report levels: no two personas share a source-sensitivity profile, and the profiles replicate across independent re-runs. A per-layer decomposition locates where the persona acts: during the manipulation turns the four personas are nearly indistinguishable at every network depth, while at the self-report turn they diverge at every depth — a five-fold difference in persona structure between reacting and reporting. The pre-registered masking pair makes the point sharply: the trickster persona reports +1 under insults and the caring persona reports -1, yet their internal states differ by less than any other pair of personas at any layer. Behaviour is not a safe fallback either: the validated exit measure never fired (36 of 36 chose to continue), and under social attack all four personas chose more contact with the user who had just insulted them. No single channel is a reliable welfare instrument, and the unit-of-concern question separates: the signal is model-level, its expression persona-level. All claims concern functional states only.