Run ab2504a7

kind=qa

The conditionality evidence is uneven across the three citations, and the user's instinct is right to push on it:

**Evans group (story-format leakage).** The persona-distance effect is the load-bearing evidence: training behavior into character X leaks to the assistant only when X is close to the assistant in persona space; distant training characters don't transfer. Under an unconditional account you'd expect the trained behavior to fire on any test persona, but the leakage gradient depends on where the *training* character sits relative to the assistant. That's strong indirect evidence the behavior is anchored to a persona-space region rather than baked in globally — though I'd note it's still indirect in the sense that the experiments mostly measure on the assistant at inference rather than running a full sweep over inference-time personas.

**Betley et al. (*Weird Generalization*).** This has the most direct conditional read-out. The Hitler-attributes experiment activates the Hitler persona by system prompt at inference and shows broadly misaligned behavior under that activation that the default assistant doesn't exhibit — i.e., behavior gated on which persona is running. Worth flagging though that this case *does* use a system prompt to activate the persona, so it sits awkwardly under the narrative's "no test-time prompt trigger required" framing. The bad-Terminator case uses a year-token trigger, which is a prompt-feature condition, not a pure persona condition. So Betley's contribution to the *story-format / no-inference-trigger* class is narrower than the narrative implies.

**Anthropic *Teaching Claude Why*.** Here the conditionality claim is the weakest. The headline result is that SFT on 14M tokens of aligned-AI fiction reduces misalignment on honeypot evals — measured on the assistant. As far as I can see in the public write-up, they don't test whether the alignment lift is specific to the assistant persona (e.g. by prompting the model to be some other AI character and re-running the honeypots). An unconditional shift in the model's behavioral distribution would produce the same headline number. So "the condition is the persona itself" is consistent with the data but not directly demonstrated — it's an interpretation imported from the Evans/Betley side of the literature.

Net: the narrative is overstating the symmetry. The honest version is something like "Evans's persona-distance result and Betley's Hitler activation give evidence that the installed behavior is conditional on persona; *Teaching Claude Why* shows that story-format SFT can shift behavior in the desired direction, but whether that shift is persona-conditional is not directly tested." I'll soften that paragraph on the next narrative pass unless you want me to make the edit now.

<<<DONE>>>

success

The conditionality evidence is uneven across the three citations, and the user's instinct is right to push on it:

**Evans group (story-format leakage).** The persona-distance effect is the load-bearing evidence: training behavior into character X leaks to the assistant only when X is close to the assistant in persona space; distant training characters don't transfer. Under an unconditional account you'd expect the trained behavior to fire on any test persona, but the leakage gradient depends on where the *training* character sits relative to the assistant. That's strong indirect evidence the behavior is anchored to a persona-space region rather than baked in globally — though I'd note it's still indirect in the sense that the experiments mostly measure on the assistant at inference rather than running a full sweep over inference-time personas.

**Betley et al. (*Weird Generalization*).** This has the most direct conditional read-out. The Hitler-attributes experiment activates the Hitl…