Latency distributions can audit a computer you are not allowed to open

What the work claims

This is a framework paper with an empirical demonstration, led by Celine van Valkenhoef with Catherine Schuman, Nikolaus Kriegeskorte, and Philip Walther: one author each from physics, electrical engineering and computer science, psychology and neuroscience, and quantum physics. That author list is itself a signal; the paper is explicitly written to apply across biological, classical, neuromorphic, and quantum substrates.¹ The claim has two parts. First, computational complexity theory tells you how performance scales with problem size, but for finite physical systems the units of time are not comparable across substrates, so scaling curves alone cannot identify the mechanism doing the work.¹ Second, the distribution of computation times, not just its mean, carries a mechanistic signature: repeated runs of the same problem expose the stochastic dynamics of whatever is actually computing, and different families of mechanism leave differently shaped distributions.¹

The demonstration is a custom visual search task: humans hunt for an emoticon target embedded in a grid that scales to roughly 160,000 image locations, while reaction time is recorded.¹ Participants numbered 94 (55 male, 39 female, mean age 27.4, range 17 to 61), with one participant contributing a 1,398-trial dataset used for illustration and excluded from comparative analysis; the remaining participants averaged 171 trials each, in a 150 to 200 range.¹ Three models with distinct evidence-accumulation mechanisms were put through the same task: a random drift diffusion model that samples locations without memory, a guided drift diffusion model that uses an accumulated evidence map to steer sampling, and a mechanistic neural network that scans the image with a small field of view, dwells where neurons respond, and keeps no spatial memory of where it has looked.¹

The punchline: mean reaction time grows roughly linearly with problem size for humans and all three models, so benchmarking alone does not separate them, and the slope differences are uninterpretable because model time steps and human seconds are incommensurable.¹ But the distributional shapes differ, and model comparison on those shapes yields a clean ordering under two independent statistics, Kolmogorov-Smirnov distance after time-scale alignment and leave-one-out marginal log-likelihood with 10,000 bootstrap resamples over participants: the random diffusion model fits worst, the guided model substantially better, and the limited-memory neural network closest to the human distribution.¹ A mechanism that accumulates evidence only while dwelling, then discards it, best accounts for how humans search.¹

How it works

The paper's central construct is the computation time profile: the distribution of latencies at a fixed problem size, or its aggregate across sizes after dividing each reaction time by the input size.¹ Four mechanistic archetypes produce four signature shapes. Deterministic execution concentrates all mass at one latency, a delta-like profile. Sequential traversal that stops when it happens upon the solution gives uniformly distributed stopping times. Interference dynamics, where wave-like interactions do the work, show periodic peaks. Accumulative mechanisms, which integrate evidence over time, give graded profiles whose shape tracks how much memory the accumulation retains: roughly exponential for minimal memory, gamma-like for limited memory, ex-Gaussian when evidence is retained and exploited globally.¹ The width and positive skew of the tail are the tell; more accumulation means more skew. Inference runs in reverse: measure the distribution, reason back to the mechanism class that could have generated it.¹

Two technical points matter for anyone wanting to use this. The characteristic distribution that aggregates across problem sizes assumes approximately linear scaling of latency with input, an assumption the benchmarking step checks first.¹ And comparing models across substrates requires aligning their time scales before shape comparison, because raw units are meaningless; the paper handles this inside the Kolmogorov-Smirnov procedure and by integrating over temporal scaling uncertainty in the likelihood analysis.¹

Where a skeptic should push

The load-bearing assumption is that one stable mechanism generates the latencies. Real systems mix strategies: a human searcher switches between guided and random sampling within a session, and a mixture of mechanisms blurs into a convolution of signatures that the framework has no machinery to deconvolve. The paper's own three models each embody a single pure strategy, so the demonstration shows that distinct mechanisms are distinguishable, not that mixed or adaptive mechanisms are.¹

Second, the scale alignment is a degree of freedom that could absorb mechanistic differences. Choosing how to map model steps onto human seconds is done inside the statistics, but any such alignment assumes the systems' internal clocks are uniform across the task, which is exactly what an adaptive system would violate.¹ Third, this is one task, visual search, with a particular heavy-tailed human distribution; the paper illustrates profiles at 155,000 to 165,000 locations using a single participant's 1,398 trials, and even then truncates long tails for display.¹ Shape estimation is data-hungry, and the gap between illustrative precision and typical deployment data volumes is large. Fourth, the result is comparative model ranking, not mechanism identification: the neural network won against two diffusion models authored by the same group, and a fourth mechanism class might fit better still. The honest claim is that distributions carry more mechanistic information than means, which the data support; the stronger claim, that they name the mechanism, the data do not yet touch.¹

Latency shape is the black-box audit primitive

Commercial biocomputing platforms are, almost by definition, closed: the tissue, the stimulation protocol, the controller, and the analytics arrive as one bundled service, and the vendor's differentiation lives in exactly the parts the buyer cannot inspect. The recurring governance worry, raised before in this stream in the context of cloud organoid training, is the credit-assignment problem: if a classical controller can absorb the learning, a benchmark measures the loop, not the tissue. This framework supplies the missing instrument. Computation in a living substrate that genuinely accumulates evidence over time should emit graded, skewed latency distributions; computation that is actually a lookup table emits a spike; a controller serially polling its sensors emits something uniform. Three systems can post identical mean scores while their delay shapes prove only one is computing the way the marketing says.¹

The opportunity is a shared, substrate-blind procurement battery. Because the method needs no access to internals, a regulator or a large buyer can run the same timing task against a dish, a GPU cluster, a neuromorphic chip, and a quantum processor, and compare what kind of computation each performs, not just how fast.¹ That converts platform oversight from trusting vendor disclosures to running a public protocol, and it gives honest vendors a way to prove their tissue is doing the work rather than asking to be believed. It also disciplines welfare claims in both directions: a platform that claims its organoid is exhibiting learning-like computation can be asked for the distributional evidence, and hype that rests on matched means but wrong-shaped latencies becomes falsifiable.

The threats are symmetrical and real. Once vendors know which shapes certify accumulation, a conventional controller can be wrapped to synthesize ex-Gaussian delays, and the audit degrades into an arms race between signature generation and signature detection. Timing distributions are also fingerprintable, so the same measurement that audits a platform can identify and copy its mechanism, turning a governance tool into an espionage channel. And both uses collide with the defining property of the one substrate this field cares most about: living neural tissue is plastic. It learns during the benchmark, so its latencies drift trial by trial and the stationarity the signature depends on is violated by the very feature the buyer is paying for. Either the audit freezes the tissue's plasticity, which means auditing a different system than the one deployed, or it accepts signatures that describe the substrate plus its history. Any governance design that adopts timing audits for organoid platforms will have to write that tension into the protocol explicitly, for example by reporting signatures per session with drift diagnostics, rather than pretending a living computer can be sampled like a silicon one.

The bottom line

Established, from the verified full text: mean computation time and scaling curves cannot distinguish mechanisms across substrates with incommensurable clocks; computation time distributions can separate at least three distinct evidence-accumulation mechanisms in a controlled visual search task, with the limited-memory dwell-accumulation model fitting 94 participants better than two diffusion variants under two independent statistics.¹ Not established: that distributions identify mechanisms in mixed-strategy or adaptive systems, that the signature taxonomy survives contact with tasks beyond visual search, or that the method works at the data volumes a routine compliance audit could collect.¹ What would confirm the governance value: a blind study in which the framework correctly classifies engineered systems whose mechanism is known to the experimenters but hidden from the analysts. What would break it: evidence that strategy mixtures and adaptive drift routinely mimic one another's shapes. For organoid computing, the near-term use is already clear: timing-shape disclosure is a cheap addition to platform reporting, and its absence, not its presence, will soon be the tell.