Global Workspace

Source. Verbalizable Representations Form a Global Workspace (Gurnee et al., 2026).

Description. A low-dimensional subspace of the residual stream, recovered as the expected Jacobian J_ of the final-layer residual with respect to layer $$, is proposed as the set of contents a model can verbally report and reason over. Five properties are enumerated in advance from global workspace theory — verbal report, directed modulation, internal reasoning, flexible generalization, capacity limits — and one is used to find the object rather than to score it afterwards. Evaluated on: four closed-weight Claude models with sizes undisclosed, plus base and model-organism checkpoints. Description mode: implementational-functional.

Readings of the claim

Global workspace / J-space. Readings of the claim, from the version the authors state to the version the field cites.

Reading	Verdict	Missing
Primary. A low-dimensional Jacobian subspace mediates which contents the model reports and reasons over	Mechanistically Supported	A converse dissociation (I6), an unmeasured-confounder bound (I8), and correction for how the token sets were chosen (M7)
Any direction encoding the concept would serve equally	Disconfirmed	Nothing — named twice and tested twice, and concept vectors do not reproduce the effect
The subspace is a global workspace in the sense the theory intends	Underdetermined	The architectural disanalogies are enumerated by the authors; no experiment separates a workspace from a bottleneck that behaves like one
The result holds beyond one model family	Insufficient	Four models, one laboratory, all closed-weight, effects scale-graded from 54% to 70% (E4)

Verdict

Global workspace / J-space. Verdict: Mechanistically Supported. The claim stops at the rung marked Blocked; rungs below it are Reached and rungs above it are blocked by that rung. Criteria in bold are absent or adverse at that rung; only those the rung requires can block it.

Tier	Requires	Missing
Proposed	construct (C1–C2)	Reached.
Causally Suggestive	measurement (M2); internal (I1)	Reached.
Mechanistically Supported	internal (I2, I4); external (E1)	Reached.
Triangulated	construct (C3–C4); internal (I5–I7); external (E2, E4)	Blocked. I6 double dissociation — Single dissociation run cleanly and repeatedly; no converse
Validated	construct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5)	Blocked at Triangulated.

36-criterion audit

Global workspace / J-space. Full 36-criterion audit. Status: C = Confirmed, PC = Partially confirmed, U = Untested, I = Inconclusive, D = Disconfirmed, N/A = Not applicable.

ID	Criterion	Status	Evidence
Construct Validity
C1	Falsifiability	C	Five properties operationalized before any result is reported
C2	Structural plausibility	C	J_ derived from the architecture, not posited
C3	Convergent validity	PC	Three cross-checks, but almost every intervention edits along J-lens
C4	Discriminant validity	PC	Separated from “any direction encoding this concept” by design
C5	Nomological validity	C	Stated inside a named theory with disanalogies enumerated
C6	Complementation validity	N/A	One subspace, undivided, so the question does not attach
Measurement Validity
M1	Reliability	PC	Wilson intervals and per-prompt dots throughout; no seed variance
M2	Baseline separation	C	The densest negative controls in the set: logit lens and tuned lens
M3	Stability	PC	Method variants swept and robust; corpus size from 1 to 1000
M4	Calibration	PC	Ablations normalized to unablated Sonnet with a smaller-model floor
M5	Sensitivity	C	A known-positive control is run and named as one
M6	Invariance	PC	Four models, base and organism checkpoints, two further lenses
M7	Selection correction	U	Unattempted, and one token set is chosen self-referentially
Internal Validity
I1	Necessity	C	Four necessity demonstrations, three with controls
I2	Sufficiency	PC	Swaps install counterfactual content at 54–70% across three models
I3	Minimality	PC	Occupancy estimated from where marginal improvement flattens
I4	Specificity	C	Tested three ways; a fourteen-task battery maps where it is not needed
I5	Rival mechanism exclusion	PC	The concept-vector rival is named twice and tested twice
I6	Double dissociation	I	Single dissociation run cleanly and repeatedly; no converse
I7	Confound control	C	Five designed controls, each aimed at its experiment’s alternative
I8	Confounding sensitivity	U	Unattempted; specific alternatives answered, no unmeasured-confounder bound
I9	Epistatic interaction	PC	Token-level exclusion at 0.09 vs 0.29; two held concepts co-occupy at chance
I10	Rescue reversibility	PC	Clamp-to-clean restoration run twice as a mediation control
I11	Onset coupling	PC	Present and load-bearing in a base model before any RLHF
I12	Offset coupling	PC	Deception signal amplified by coding RL, then attenuated by safety training
External Validity
E1	Intervention reach	PC	Many intervention forms, nearly all editing along the same directions
E2	Prompt generalization	C	Five public benchmarks plus six prompt distributions
E3	Cross-task generalization	C	A fourteen-task battery maps where the mechanism is required
E4	Cross-model recurrence	PC	Four closed-weight models, one lab, sizes undisclosed; partly replicated externally
E5	Graded response	C	Six dose-response designs, including injection strength against rank
E6	Novel prediction	C	Five properties enumerated from theory, one used to find the object
Interpretive Validity
V1	Level declaration	C	Declared before evidence, and the level is named explicitly
V2	Level-evidence match	PC	Three levels in play and two matched; the third is asserted
V3	Alternative level	PC	Two alternatives addressed and refuted; the workspace framing is not
V4	Unlicensed labeling	PC	Disanalogies enumerated and consciousness not adjudicated; hedged
V5	Scope declaration	C	Eight limitations named in a dedicated section
Total: 14 Confirmed, 18 Partially confirmed, 1 Inconclusive, 2 Untested, 1 Not applicable
Verdict: Mechanistically Supported

Sensitivity

Criterion judgments this audit record leaves contested: each carries an argument on both sides, and the status shipped is the one the record settled on. At Partially confirmed this claim reaches Validated.

Criterion	In tension	What would settle it
I6	Inconclusive or Partially confirmed	Whether a crossing over layer bands meets a criterion asking for two mechanisms

Exploratory Lens Analysis

An exploratory reading of this claim through the framework’s five lenses, written for this site and not part of the paper, is at Global Workspace — exploratory lens analysis.