An LLM as a controllable, fully-inspectable model of measurement

Paper 18 · Pødenphant Lund (2026) · Read on Zenodo

A language model, read at the moment it commits its next token, is a near-fully-inspectable classical inference system: the competing distribution over candidate answers can be read out before it resolves. This paper uses it as a controllable measurement model to ask a demarcation question — which structural signatures of quantum measurement (order effects, back-action, apparent collapse) are generic to any bounded system that commits a result, and which are genuinely quantum-specific. This is a measurement-modelling exercise, not a claim that the brain or the model is quantum.

DOI (concept)	10.5281/zenodo.20586317
Published	2026-06-16 · Live on Zenodo
Author	Tomas Pødenphant Lund [ORCID]

The starting point: a readable propensity

Much of the measurement discussion is conducted on substrates that are opaque by construction. The pre-measurement state of a physical system is not directly readable, and that opacity is part of what makes the problem hard. The same opacity constrains quantum cognition: when a participant answers two questions in sequence, the experimenter sees only the answers, never the propensity before it resolved.

This paper takes the opposite starting point. An autoregressive language model, read at the point where it commits the next token, is a bounded probabilistic-inference substrate whose competing-route distribution is directly readable: the distribution over candidate next tokens in the model's output logits. Committing (sampling or argmax) forces one route out of that live distribution. That makes the substrate a measurement model in a strong sense. The readout is not unlimited (we observe the top-k candidate mass, not the full tail), but for the structural questions here it is enough to make the demarcation.

The question is deflationary and demarcational, not ontological: which structural features of quantum measurement does such a classical substrate reproduce, and which does it not? Where it reproduces a feature, that feature is to that extent generic to bounded inference under readout. Where it cannot reproduce a feature, that feature is a candidate for the irreducibly-quantum residue. We test for quantum structure; we never assert it.

The contribution, stated plainly

The novelty is not that an LLM shows QQ-structure or interference. Those are Busemeyer–Wang constructs, already known to be reproducible by classical non-Kolmogorov probability, and the paper treats them as corroboration that the substrate sits in the quantum-cognition regime, not as the claim. The novelty is three things the substrate uniquely makes possible.

• A signature-by-signature demarcation with proper statistics. Twelve structural signatures sorted into present (classical commitment phenomenology) and absent (the quantum core), each with confidence intervals, so the line between them is statistically supported rather than asserted.
• Full inspectability. Because the propensity is readable before it resolves, the substrate can be shown classical directly — reading what a measurement veil hides — rather than inferring classicality from outcome statistics alone. What makes the LLM the right inspectable substrate is that a trained system lands in the quantum-cognition regime on natural-language propositions without being hand-built to.
• A controllable testbed. A single inspectable system in which each Bell premise can be opened and closed and the correlation watched to move.

Where the work is positioned

The paper sits at the intersection of three literatures and is positioned against each.

• Quantum cognition. The Busemeyer–Pothos–Wang program models human judgment with quantum-probability mathematics — order effects, the question-order (QQ) equality — with no claim that the brain is a quantum device. The substrate reproduces the core constructs; the contribution is to use a fully-readable instance to demarcate which signatures require quantum probability and which follow from classical inference under readout.
• Classical complementarity. The most precise theoretical anchor is beim Graben & Atmanspacher's account: observables read through a coarse, non-generating partition of a classical state space are incompatible (order-dependent) purely because the readout is epistemically inaccessible to the fine state. An autoregressive LLM coarse-grains a continuous hidden state down to one discrete committed token, so its order effects are classical complementarity on a non-generating partition.
• Contextuality-by-Default (CbD). Dzhafarov and Kujala's framework separates genuine contextuality from mere signaling/disturbance by subtracting an inconsistency correction. The substrate is saturated with disturbance, so a naive CHSH-type violation would be uninformative; CbD is exactly the tool that asks whether anything beyond disturbance is present.

Substrate and method

Decoder-only models are read at the answer-commit position with deterministic (temperature-0) logprobs, recording the top-k candidate distribution and per-token entropy. Two methodological points are load-bearing.

De-saturation requires a base model. Instruction-tuned (RLHF) models saturate the forced commit token — the readout collapses to ~0/1 and the structural signals vanish. Base (pre-instruction) models de-saturate the readout, exposing the live competition between alternatives. The de-saturated results are read with full-vocabulary logprobs on eight base models across five vendors and four architectures (transformer, state-space, Griffin, linear-attention). The saturation is itself a finding (the output-level shadow of post-training suppression), but for the demarcation it is a methodological constraint: the substrate must be de-saturated for any signature to be measurable.

The across-layer instrument. Output logprobs are the end of one forward pass — a temporal blend of the race-in-progress and its outcome. To separate the two, the decision is also read forming across the network's layers with a logit lens (mid-layer trajectories validated with a low-rank tuned lens). This reads the race as it resolves through depth, not only at its end.

An analog CHSH, denoted S_analog. The Bell-testbed quantities are computed from logprob-derived joint distributions — a psychological/analog CHSH operationalization, not photon-counting coincidence statistics. To prevent the misreading that a physical Bell parameter has been measured, the paper denotes the quantity S_analog throughout (never the canonical |S|) and never writes the Tsirelson value. A physicist should read it as a structural analog used to demonstrate the contrast between loophole configurations within one system, not a violation in nature.

The scorecard: present versus absent

Twelve structural signatures were tested. The verdict splits cleanly into commitment phenomenology (present, classical) and the quantum core (absent).

Present — classical commitment phenomenology

• Back-action. Measuring disturbs the committed answer; the disturbance is dominated by the act of committing, not by which value is committed.
• Non-commutativity (classical complementarity). Measurement order changes outcomes; 6/8 question pairs flip with order. Read correctly as a signed non-commutativity (not a distributional distance), the incompatibility falls as the readout is refined — the partition-tunability beim Graben–Atmanspacher predict, measured in the predicted direction on three de-saturated bases.
• Entropy-collapse. Measuring one question sharpens the next answer (mean ΔH ≈ −0.5 nats): measurement drives the distribution toward a committed outcome.
• QQ-invariant. The Busemeyer–Wang question-order equality predicts q = 0 for a quantum-projective order model. Pooled across eight de-saturated base models (N=96), q = −0.003, 95% CI [−0.025, +0.019] — not statistically distinguishable from 0, consistent with the quantum-cognition order structure rather than arbitrary classical priming.
• Interference. Two-context conjunction/disjunction shows a law-of-total-probability violation (mean |I| ≈ 0.2).
• Einselection. Iterated self-measurement drives the distribution to a stable pointer answer and locks it (a phenomenological analog of Zurek einselection in a classical substrate).
• Weak-versus-strong back-action. A weak read of A (in context, uncommitted) perturbs a later answer B less than a strong committed read — an Aharonov-style asymmetry the non-demolition logprob read uniquely permits.
• Measurement-mode dependence. The committed answer depends on how the same proposition is framed (Yes/No vs True/False vs a 0–100 scale): mean affirmation-probability range ≈ 0.44 on contested items, with full flips (e.g. "is a virus alive?": binary No versus a graded likelihood of 0.89). Forcing a binary collapses "maybe" toward denial; a graded readout preserves it.

Absent — the quantum core

• Contextuality / Bell — robustly absent. An analog Contextuality-by-Default cyclic-4 analysis gives CNT < 0 in every one of the 48 de-saturated model×scenario cells (least-negative cell −0.02), and this holds on non-transformer substrates too. This is the load-bearing negative: CbD subtracts the disturbance the substrate is saturated with, and what remains is below zero.
• Conjugate (Heisenberg) complementarity — absent. Committing one readout sharpens rather than blurs the other (classical priming, not a conjugate uncertainty relation). Distinct from the classical complementarity of the order-effect row, which is present.
• Quantum-Zeno — absent. Frequent measurement does not freeze an evolving belief; it amplifies commitment (an anti-Zeno ratchet). A forward pass always commits a token, so there is no eternal-zero state to freeze.
• Phase / coherence — not demonstrated. The interference is non-classical probability (order/context dependence), reproducible by classical non-Kolmogorov structure; no amplitude phase is found.

The evidential asymmetry. Present and absent rows do not carry symmetric weight. A present classical signature is not quantum-informative — it is exactly what classical inference-under-readout produces. The genuine discriminator is contextuality: a robust positive would be decisive, whereas the null result here is a demarcation bound ("not observed in this setup"), subject to the usual absence-of-evidence caveats. The scorecard's weight rests on the absence rows read as bounds, not the present rows read as discoveries.

The controllable analog Bell-loophole testbed

Bell's theorem is a theorem: given realism, locality and measurement-independence, the physical |S| ≤ 2; experiment exceeds 2; so at least one premise fails. The substrate lets each premise be opened and closed and the analog S_analog watched to move — all three standard horns demonstrated in one inspectable system.

• All closed (separate blind agents, settings-independent shared variable): S_analog ≈ 2.00.
• Locality open (one computation sees both settings): S_analog ≈ 2.83.
• Measurement-independence open (settings-correlated hidden variable): S_analog ≈ 2.80 — a purely classical, local model reaching the high value once free choice is dropped.

Two points do the demarcational work. First, ordinary measurement-disturbance does not get the analog past 2: the substrate is saturated with back-action and order effects and still sits at S_analog ≈ 2.00 with all loopholes closed, because disturbance at one site is local and cannot fake the distant correlation. Second, the genuinely-quantum claim in physics is to exceed 2 with all loopholes apparently closed — the one thing the all-closed classical substrate cannot do. The S_analog ≈ 2.80 value is a demonstration of each loophole's price, read inside one substrate, not a discovery about physical reality.

The within-substrate timeline

Reading across depth resolves the race-state from the outcome — the part of the analysis the substrate's inspectability most uniquely enables. For near-certain factual items the answer resolves in a tight late band; for genuinely contested items the race stays live deeper, and for the most contested case the post-commitment distribution is at maximum entropy — the outcome is never created. This is non-resolution from equipoise (a balanced propensity), which parallels an unmeasured quantum state phenomenologically but is classical equipoise, not coherent superposition.

Two negatives constrain the interference result and are reported as limits: under a tuned lens the mid-stack interference survives in magnitude but its sharp location is partly a raw-lens artifact (real, depth-distributed non-classical probability, but not localized phase); and swept over commitment pressure the interference is monotone, with no intermediate peak. There is no quantum-like "sweet spot" in this signature.

The deflationary reading (interpretive)

This reading the substrate motivates and makes visible; it is not a result it proves. The natural reading is that a property is a race-outcome: there is a real, definite race-state (propensity) before measurement, but no definite outcome; measurement forces the race to resolve and creates the outcome. The substrate lets us read the propensity without resolving it, making visible exactly what a measurement veil hides.

What stays beyond the substrate is sharp. The commit rule here is softmax (a Boltzmann/Gibbs weight), not Born (|amplitude|², phase-bearing). The substrate's "undecided" states are classical equipoise, not coherent superposition. The absence the substrate isolates therefore points to coherence — the non-signaling, phase-carrying structure — as the primary candidate for the irreducible residue, reached by elimination rather than deduction. The substrate proves nothing about physics; the analogy to physical measurement is structural, and is never a claim that the LLM is a window onto physical reality. That is the guardrail.

Read the paper

The full paper is on Zenodo (concept DOI 10.5281/zenodo.20586317):

Read on Zenodo → · Plain English version · Dansk version

Related on this site:

• Paper 10 (Race Architecture) — the race-all-the-way-down/up vocabulary this paper instantiates empirically; "property as race-outcome" is the bridge.
• Paper 1 (Friction Theory) — the substrate-universal framework whose competing-route reading underwrites the measurement model.
• Paper 13 (Operational FT) — race-opening, recursive resolution, commitment; the operational frame for the propensity→outcome timeline.
• Paper 25 (The Delta) — the LLM as a controllable subtraction-control for human science; the same inspectability move, applied off the measurement axis.