Can the instruments for a model's workspace be trusted?

Global workspace theory comes from the study of human cognition. It proposes that much processing runs in parallel and outside awareness, and that a small part of it is made available to many processes at once, so that it can be used for categorisation, memory, evaluation and intentional action^[1]. Recent work reports an analogous functional distinction inside language models^[2]. If that holds, a model's stated reasoning could be checked against the structure that drives its behaviour, which matters to anyone who must oversee a system from outside.

The claim rests on instruments. A workspace reading that shifts when the lens is refitted, or moves with a random seed, supports no such check. In psychology a test earns trust by agreeing with itself and differing from tests of other things^[3]. This inquiry applies that routine to the workspace tests before anyone trains towards one.

Global workspace theory is a hypothesis about human cognition. It holds that many modular networks process information in parallel and outside awareness, and that information reaches categorisation, memory, evaluation and intentional action when it is made available to many of them at once^[1].

Gurnee and colleagues apply the idea to language models. A Jacobian lens, built from the derivatives of the model's output with respect to its internal state, identifies the representations a model is poised to verbalise, called the J-space. The authors report that its contents can be reported, held and used to carry intermediate steps of silent reasoning, and that it carries coherent content only in an intermediate band of layers and holds tens of concepts at a time^[2]. They describe these as functional analogues of properties the theory associates with conscious access; this inquiry takes no position on that.

Rahbar proposes the Ignition Index. It fits a sigmoid, an S-shaped curve, to per-layer probe accuracy as input signal strength varies and reads the steepness as abrupt or graded. It was applied to 11 models from five architecture families, with shuffled-label controls^[4].

Probes and interventions have misled before. A probe can score well because it learned the task rather than because the representation encodes it, which is why control tasks and selectivity were introduced^[5]. An intervention on a subspace can change an output by activating a dormant parallel pathway that is causally disconnected from the model's outputs^[6]. Neither concerns workspace tests; both name failure modes to check for.

Oskin imposed a chosen geometry on a 12-layer model through the loss and found it unreliable: six of eight seeds collapsed, while a fixed rotation reached baseline quality^[7]. It targeted a different property. The 2026 papers cited are preprints.

Three tests are applied to published open-weight language models. Report-use agreement asks whether what a model reports about a held concept matches what that concept does to later computation. Ignition steepness asks how abruptly a concept becomes readable across layers as input signal strength rises. Monitorability asks whether workspace readings detect held-out interventions on behaviour, such as refusal or suppressed tool calls. The definitions draw on Gurnee and colleagues^[2] and on Rahbar^[4] and will be fixed in a public, time-stamped pre-registration before any data is read.

Stage one uses inference only, on published open-weight models of up to eight billion parameters. The Jacobian lens is refitted repeatedly on each model and the tests are re-run across refits and evaluation seeds; reliability is the test-retest correlation. The tests are also compared by how they rank the same models, in the manner of convergent and discriminant validation^[3]. They are frozen before any training touches them, so that a later result cannot be put down to tests tuned to produce it.

Stage two is conditional. A model of one billion parameters or fewer is trained further with a term that concentrates its internal causal influence into fewer dimensions, and compared with a control that rotates the model's own coordinates at random. Each arm runs at least eight seeds, so that run-to-run variation is estimated directly. The primary metric is report-use agreement, with the loss in language-modelling quality (perplexity) bounded in advance. A model of up to eight billion parameters will be tried only if the small-scale result meets none of the stage-two stop conditions.

These conditions are written before any data is read.

A test passes stage one if it reaches a test-retest correlation of at least 0.8 across lens refits and evaluation seeds, on a named reference model where the lens, as published, can be refitted at an affordable scale. A test below 0.8 is not used to judge anything later. If report-use agreement, the primary metric, falls below it, stage two does not start. If the lens cannot be refitted at that scale, the inquiry stops. Whether the tests rank models alike or differently is reported either way; it is a result about the instruments.

Stage two passes if the trained arm's report-use agreement exceeds the rotation control's interval by at least the pre-registered effect size, within budget. Stage two stops if the trained model's primary-metric result lies inside the control's confidence interval, if the effect falls below the pre-registered threshold (working proposal: Cohen's d of 0.3, 95% interval excluding zero), or if cost passes a budget fixed in advance by more than a fifth. The seed count is not cut to fit the budget. If eight seeds per arm cannot be afforded, stage two does not run.

A high test-retest correlation can come from what the refits share rather than from the property measured. If every refit draws on the same prompts, or the correlation is taken across models of very different size, readings repeat because the inputs or the scale repeat. The design will therefore refit the Jacobian lens^[2] on disjoint prompt sets with fresh seeds and compute reliability over items within each model. A probe can also score well because it has learned the task itself; control tasks with random labels show how much of its accuracy the probe alone supplies^[5]. A concept that can be read out is not thereby used. Removing it and measuring the change in behaviour tests use directly^[8], which is why report-use agreement, not decodability, is the primary metric. An intervention on a subspace can change an output by activating a dormant pathway, so that the subspace looks causal when it is not^[6]. Each intervention will therefore run beside norm-matched random directions, same-layer principal components and a final-logit baseline. Tests that share a lens or a probe family share its errors, and agreement between them can reflect a shared method rather than a shared trait^[3]; the comparison of how the tests rank the same models will be read with that in mind. In stage two, further training of any kind can move the metrics, and a loss on internal geometry can succeed or fail by seed^[7]. The random-rotation arm, at the same seed count, is designed to separate a change in what the model computes from a change of coordinates.

A test can repeat without measuring what it is named for. Stage one establishes reliability. Validity needs agreement between different methods of measuring the same thing^[3], and stage one only begins on it. A pass would show stable readings on the chosen models, not that a model's reports match what drives its behaviour in general. Intervention-based readings can mislead even when they look causal^[6].

Results on open-weight models of up to eight billion parameters would not carry to larger or closed systems by default. Nothing here bears on whether a human workspace and a model's J-space are the same kind of thing, or on experience of any kind. A positive stage-two result would show that an objective moves these measures against a rotation control, not that the resulting model is easier to oversee in practice. A negative result is reported in the same detail as a positive one.