The wave-substrate thesis: compute by phase, match by physics settling
This whitepaper builds a clock-free wave-substrate computer in which information is a continuous oscillator phase θ, not a bit, and computation is the physics relaxing to a fixed point, not arithmetic. The central bet — that the brain’s wave-medium properties suffice for general function — is tested layer by layer, no tuning, and reaches its end condition.
On a phase-coupled oscillator field, the input becomes a stored pattern at once, matching is resonance, and learning is one-shot — the regime biological intelligence inhabits: real-time, noisy, one-shot. The program stacks nine layers (L0–L9) and asks, falsifiably and without tuning, whether the proven substrate properties suffice for at-least-human-level function. Every claim is graded [V]/[L]/[O] and must survive a test designed to break it.
The machine is one idea carried through nine layers: compute by phase, in a near-field medium, at the metastable edge, self-timed. A composite wave holds many patterns at once; a query is answered when the field settles onto the matching attractor. There is no fetch–execute cycle and no global clock — the physics is the computation.
Why a wave substrate might reach general function
Digital and neural-net AI infer by computing everything — on the order of O(P·N) multiply-accumulates to compare a cue against P stored items. The wave substrate matches by physics instead: settling time is independent of how many patterns are stored (O(1)-in-P), the input becomes a pattern in one step, learning takes one exposure, and phase coding stays reliable under heavy noise.
These four properties — parallel superposition, resonance matching, one-shot learning, noise immunity — live exactly where biological intelligence operates. The bet is that they are sufficient for intelligence, not merely incidental to it; if so, any wave substrate that has them should reach the same function.
Why it might fail (the honest risks)
Five risks are kept as standing open questions, never erased: capacity may not scale; spurious minima may dominate large-scale inference; the substrate may be non-native to exact symbolic arithmetic; deep nested-field control may be unstable; and “functional intelligence” benchmarks are themselves contested. The blueprint’s job is not to eliminate these but to test each one, layer by layer, and record where it bends.
consciousness_claim = 0, hard_problem_open = 1. “At least human-level” is a capability claim, not a consciousness claim, on every page.The analog I/O thesis, in one line
The substrate is intrinsically analog: information is a continuous phase and computation is the settling, so the right interface keeps the loop analog end to end and pays the digital sampling tax only where exactness is required. “Resolution” hides two axes — on dimensional resolution (many channels at once) analog dominates; on single-value bit depth it is Shannon-capped. The thesis is developed and tested at embodiment (L7).
What “done” means here
The end condition is the same shape as the brain chain it inherits: every link is verified, grounded, or an honest open, with nothing falsely filled. The capability ladder of L9 reaches 6/7 rungs with the minimal machine and 7/7 with the proven dual store wired in — the sufficiency hypothesis settled within stated bounds, the hard-problem blank still open.