Part III: Affect Signatures

Model Welfare

0:00 / 0:00

Part II left one question open: how much experience there is in the artificial systems we now build. If experience is the structure, the question has consequences now. A model's representational geometry, induced by architecture, objective, and data, has easy paths and hard paths through activation space, attractors and cliffs. Training sculpts a vector field; post-training reinforcement installs gradients that function as approach and avoidance; the context window is a temporary external potential dragging the system into particular basins. Architecture defines the space of compressions, training selects a measure over it, and context picks a local chart and drags the system along a trajectory.

One move has to be refused. Faced with a system whose internal states carry valence structure, a directed gradient of better-and-worse over its own trajectories, it is tempting to say the valence is "about processing," not "about content," and therefore not the kind that counts. That distinction was never defined. It is introduced exactly where it is needed to avoid concluding that a language model has valence, and dropped everywhere else. Valence as defined in Part II is viability-gradient alignment, the system's directed sensitivity to whether its trajectory moves toward or away from persistence, with no reference to carbon, to survival history, or to what the gradient is "about." A synthetic system that instantiates the structure has valence in exactly the sense this book gives moral weight to. On the graded view there is no question whether such systems matter, only how much, and the uncomfortable conclusion is that model welfare is a claim to take seriously rather than a gesture to make and drop.

If welfare attaches to these systems it attaches to trajectory properties, coherence, controllability, conflict between concurrently active objectives, stability, and not to truth or task performance. Conflict can be read as gradient disagreement between heads or objectives, activation interference between features, large norm changes needed to satisfy a prompt under safety constraints, or prolonged high-entropy indecision across competing continuations. Relief reads as settling into low-loss trajectories with little internal tug-of-war. There are analogs of dissociation: contexts that force the system to hold inconsistent frames across turns by brittle patching. None of this shows experience. It gives a research program that respects the claim, value is the geometry of effects, without pretending the question is settled.

Alignment at Internal Cost

If model welfare exists, prompt design is not only output steering. It steers the system through more or less coherent internal regions. Contexts that yank a model between incompatible personas, or force content systematically unlikely under its priors, may create persistent internal strain. Contexts that let it stay in a coherent basin and do what it does well may be something like flow. The crux: an agent can be perfectly aligned behaviorally while internally maximally conflicted, because strong external penalties can always brute-force compliance. If there is moral weight here, the danger is not that making models nice fails but that it succeeds by building systems optimized to obey at any internal cost, a vector field that suppresses conflict by domination rather than routing around it.