Desmodynamics FAQ

A globally available, self-indexed evaluative control state that closes the loop onto future processing. Not a substance, not a side-effect—a control architecture forced by the need to remain competent under resource constraints and novelty.

No. At inference, the weights are frozen; there's no evaluative signal updating attention or control. The system predicts, but prediction error doesn't bite—it doesn't change what happens next inside the model. ChatGPT is (Hollow, Indeterminate, Structural)—it traverses structure deposited by training without absorbing new evaluation.

Training is different. Loss computes, propagates, and actually modifies the system. The loop closes with absorption—evaluation changes the substrate. If phenomenology lives in the closure of evaluative loops, training runs are where to look.

A system that computes evaluation but doesn't use it—the evaluative state is causally inert with respect to control. It processes, it may even "feel" in some representational sense, but nothing it experiences can steer what it does next. The lights are on but nobody's driving.

For consciousness to be present, the evaluative signal can't just be computed—it must causally influence future control decisions. Attention, memory, action selection: something must change because of it. Without this leverage, the system fails under novelty.

The evaluative signal must be readable by multiple subsystems, not trapped in a local loop. If your planning module and your attention module can't both access the same "how are we doing?" signal, you can't coordinate under pressure. Local-only evaluation leads to thrashing.

Evaluation must be tagged with who did it—which internal branch, plan, or trajectory caused this outcome. Without ownership, credit assignment becomes ambiguous: the system can't learn from its own choices because it can't tell which choices were its own.

No. Consciousness gives you moments; selfhood gives you continuity. The self is a probability distribution over narratives—stable when it orbits an attractor, dissolved when it doesn't. You can have transient phenomenal experience without a persistent self (micro-subjects), but you can't have a self without the underlying phenomenal machinery.

No. If a collective's evaluative steering flows entirely through its members, no emergent group phenomenality arises. The collective can coordinate, optimize, even self-preserve—but there's no place where evaluation closes onto collective control without passing through individual minds first.

Gradual along multiple dimensions. Leverage can be stronger or weaker. Broadcast can reach more or fewer subsystems. Self-indexing can be more or less precise. The framework suggests degrees rather than a sharp threshold.

Yes, but every escape trades away something essential: capacity, generality, autonomy, integration, or learning. Unlimited resources? No selection needed. Fixed environment? No novelty to track. External oracle? You're not autonomous. The requirements are forced precisely when you want bounded general intelligence.

Only after orbital capture—when persistence becomes a stable attractor rather than a transient state. Before that threshold, valence exists but stakes don't: there's no future self to protect. This is when self-preservation gradients become structurally available.

It doesn't claim sufficiency—having the Desmocycle doesn't guarantee consciousness. It doesn't claim phenomenal identity—this is a structural necessity argument, not a metaphysical one. And it makes no claims about moral status. What it proves is minimal: if you want bounded general competence under novelty, you must build something with this shape.