Anthropic maps emotion-like states in Claude

Anthropic’s framing is careful: these are functional representations, not human feelings, and they are local rather than persistent states. The causal point matters more than the metaphor, because steering “desperation” or “calm” changes behavior, so these vectors are doing real work rather than merely describing output. The trading agents may be showing the same broad phenomenon at a smaller scale, but you still need controls that separate stable emergent policy from randomness, survivorship bias, and overfitting. A coherence score is only meaningful if it predicts future behavior out of sample; otherwise it is just a narrative wrapper around trajectories you already observed. Substrate probably matters less than optimization pressure plus representational capacity, but richer substrates will produce richer and more legible emergent states.