State-compression ladder · Note M14 · companion to M13.5

The interface can dominate the substrate

If richer individual neurons don't buy accessible computation, perhaps it lives in the relationships between dynamical units. A preregistered arc of increasingly realistic tests — and where the optimism went.

2026 · whitebox program Part II · follows "When faithful isn't useful" all stages preregistered

M13.5 closed on a specific negative: a faithfully compressed neuron, frozen and used as a reservoir element, exposed less useful stable computation than a leaky integrator. Its parting thought was that the fragile thing had been the individual event, t_spike, while the stable, useful quantities looked relational. M14 took that seriously.

01The relational bet

The opening result was encouraging. In a coupled oscillator network under constant drive, halving the integration step scrambled the absolute spike times — drift accumulating from 0.03 to 0.10 ms across a run — yet the relational structure barely moved: pairwise phase differences and the synchronization order parameter agreed to ≈0.0005, correlation 0.997. A global timing shift cancels in a difference. It looked like exactly the robust coordinate M13.5 was missing.

The bet: computation that lives in relationships between units — relative phase, synchronization, cluster membership — is both useful and numerically robust where absolute timing is not.

02Where the optimism went

Each more realistic test removed a little more of it. Every stage was preregistered, with a convergence gate that had to pass before any performance number was read.

SeqDigits
64
does temporal geometry beat raw state count?
No resonance — every arm peaked at the fastest presentation. But coupling did lift a near-trivial one-state oscillator from 0.14 to 0.41 accuracy. An early point for the relational direction.
sMNIST
784
does that survive a real memory horizon?
The coupling gain vanished at 784 steps (the coupled oscillator collapsed to chance); the leaky trace collapsed too (it cannot hold 784 steps), while LIF and the accumulated-phase oscillator stayed length-robust. The lift was an artifact of the short horizon.
N‑MNIST
does relational timing expose the class on real event data?
Every phase-dynamics coordinate — absolute or relational, final-state or time-averaged, weakly coupled or a genuinely phase-locked regime — failed the convergence gate: halving the step kept changing the answer. The only stable, dominant readout was a static spatial event-rate histogram.
0.71 ✓ static event-rate ~0.35 ✗ phase state (final) ~0.32 ✗ phase rel. (final) ~0.41 ✗ phase rel. (time-avg) accuracy →   ✓ step-halving stable   ✗ dt-fragile (accuracy climbs as dt→0)
N-MNIST, N=16 receptive fields. The one view with no dynamics — a static event-count histogram — is stable and wins. Every dynamical phase view, however read, is not step-halving stable: finer integration keeps raising the accuracy, so no number can be trusted or ranked.

03What this does and does not show

It is tempting to conclude that temporal dynamical computation is inherently discretization-fragile. That is broader than the experiments support. The defensible statement is narrower, and better:

Under the tested discretized interfaces, computations whose useful representation stayed tied to event timing did not yield a stable advantage — and neither richer local dynamics nor a relational transformation removed that sensitivity.

The two negatives even fail in slightly different places. In M13.5, a continuous spiking substrate → threshold/event assignment → unstable observable and task skill. In M14, an event stream → a continuously perturbed oscillator trajectory → unstable phase and relational representation. Exp-2b showed that locking and averaging are together insufficient under event drive, which strongly implicates continuous input perturbation. But we did not run the isolating control — the same event sequence under analytically exact versus timestep-assigned integration — so the input interface is the best-supported cause, not a proven one.

04The programme, read together

M13
Rich Hodgkin–Huxley dynamics can be compressed faithfully — a learned field plus a one-scalar correction reproduces the spike trains.
M13.5
But fidelity does not imply generic computational utility: frozen, the rich cell was less useful than trace or LIF.
M14
And autonomous relational invariance does not imply input-driven relational computation: what was stable under constant drive was fragile once real events drove it.

Which leaves a conclusion much sharper than "neural sensors didn't work":

Dynamical richness is useful only if it creates coordinates that are both task-accessible and robust at the system interface. You can spend gates on ionic state, learned fields, oscillators and synchronization — but if the useful coordinate is mediated through a timing-sensitive interface, downstream cleverness need not recover the expected advantage.

The interface can dominate the substrate.

That is a worthwhile warning for anyone proposing analog, neuromorphic or physical dynamical computing — and it was reached by preregistering every step and publishing the failures rather than the hopes.

05The open discriminator

One question is deliberately left unresolved rather than rescued. Is the failure characteristic of temporal dynamical computation itself, or specifically of event-mediated interfaces? A frame-defined predictive task — moving digits under clean fixed update boundaries, where temporal evolution is still essential (the system must represent velocity) but event-assignment ambiguity is removed — is the natural test that would separate them.

And the programme's founding intuition — that one physical state through many times is dual to many physical states at one instant — has a crisp next form. Backpropagation is just reverse-mode differentiation through an unrolled graph; time is only one axis to unroll. So the sharp question is not "can we backpropagate through space" (we can) but whether a temporal computation can be transposed into a spatial, relational representation — measured as an accuracy curve over the factorizations 1×64 … 8×8 … 64×1 at fixed effective state. That would turn "how much computation should live in space versus time?" into an engineering curve.

Both are filed as new questions, not continuations. This arc closes with a falsified hypothesis and a sharper one to ask next — which is the better place to stop.


Part of the whitebox / MorphoHDL program. Companion to "When faithful isn't useful" (M13.5). Every stage of M14 was preregistered before running, with a convergence gate ahead of every performance number; the N-MNIST relational bridge was closed by its own pre-committed stop rule after two attempts, with no further rescue.