| phenomenon | global workspace / conscious-access-like representations in large language models— “Specifically, we observe that language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing.” |
| claim-type | existence— “In this paper, we provide evidence that LLMs do possess such workspace-like representations.” |
| method — evidence | — “The Jacobian lens is a technique for inspecting the contents of the residual stream at these intermediate layers.” |
| transfer | from cognitive-neuroscience (unconscious priming): global-workspace / access-consciousness as the mechanism gating which primed information becomes reportable and modulates behaviour | transfers[0] |
| mechanism | A CONSTRUCT IS DEFINED BY ANALOGY TO ANOTHER FIELD'S CRITERIA, AND THE ANALOGY IS THE DEFINITION RATHER THAN A MOTIVATION FOR IT: we DEFINE A SUBSET OF VECTOR REPRESENTATIONS AS WORKSPACE-LIKE IF IT SATISFIES THE FOLLOWING PROPERTIES, which MIRROR THE PROPERTIES CHARACTERISTIC OF CONSCIOUS ACCESS described above. What makes something a member of the new category is that it satisfies conditions imported from a theory about brains— “define a subset of vector representations as workspace-like if it satisfies the following properties, which mirror the properties characteristic of conscious access described above” | mechanisms[0] |
| mechanism | AND THE PAPER'S OWN LIMITATIONS SECTION SAYS THE DISTINCTION IT RELIES ON HAS NO PRINCIPLED DEFINITION: this judgment WAS SOMEWHAT POST-HOC, and WE DID NOT PROVIDE A PRINCIPLED DEFINITION OF WHAT DISTINGUISHES A “WORKSPACE” REPRESENTATION FROM A “motor” ONE. The boundary was drawn empirically and the criterion for it is declared absent, by the people who drew it, in the same paper that uses it throughout— “But this judgment was somewhat post-hoc, and we did not provide a principled definition of what distinguishes a "workspace" representation from a "motor" one.” | mechanisms[1] |
| mechanism | THE MISSING CRITERION IS SPECIFIED AS A CHOICE BETWEEN TWO REPAIRS, NEITHER OF WHICH IS PERFORMED: A FULLER ACCOUNT WOULD EITHER SUPPLY SUCH A PREDICTIVE CRITERION, OR WOULD REPLACE THE FLEXIBLE/AUTOMATIC DISTINCTION WITH SOMETHING MORE PRECISE. What would fix the vocabulary is stated exactly, and left as a description of a paper someone else could write— “A fuller account would either supply such a predictive criterion, or would replace the flexible/automatic distinction with something more precise.” | mechanisms[2] |
| mechanism | AND THE REASON THE CRITERION IS NEEDED IS GIVEN AS AN INABILITY, NOT A DIFFICULTY: the tasks were chosen BECAUSE THE FLEXIBLE/AUTOMATIC DISTINCTION IS CLEAR FOR THEM; however, for many tasks of practical interest **IT IS NOT**, and WE HAVE NO WAY OF PREDICTING IN ADVANCE, FOR AN ARBITRARY COMPUTATION, WHETHER IT WILL ENGAGE THE J-SPACE. The cases were selected for the property that the distinction applies to them, and that selection is disclosed— “The tasks in §3.5 were chosen because the flexible/automatic distinction is clear for them; however, for many tasks of practical interest it is not, and we have no way of predicting in advance, for an arbitrary computation, whether it will engage the J-space.” | mechanisms[3] |
| mechanism | A CONCEPT IS DEFINED AS WHATEVER THE SYSTEM ITSELF LATER PRODUCES, AND THE CIRCULARITY IS THE POINT OF THE EXPERIMENT: HERE WE DEFINE THE REACTION CONCEPT TO BE WHATEVER ANSWER THE MODEL ITSELF GOES ON TO PRODUCE, so the question is whether the model's eventual choice IS ALREADY IN THE LENS WHILE IT READS THE INSTRUCTION TO WITHHOLD IT. The definition has no content until the system has acted, which is what makes the test possible— “here we define the reaction concept to be whatever answer the model itself goes on to produce, so the question is whether the model's eventual choice is already in the lens while it reads the instruction to withhold it.” | mechanisms[4] |
| mechanism | A METHOD IS DEFINED BY THE QUESTION IT ASKS RATHER THAN BY WHAT IT MEASURES, AND CONTRASTED WITH THE METHOD IT REPLACES: the J-lens provides a complementary way of interpreting a feature - IT ASKS WHAT THE FEATURE DIRECTION IS VERBALIZABLE AS, INDEPENDENT OF WHERE IT FIRES. The older method is said to MISLEAD WHEN THE FEATURE ENCODES AN ABSTRACT ASSESSMENT WHOSE SURFACE TEXTUAL CORRELATES ARE INCIDENTAL— “The standard approach to interpreting an SAE feature is to inspect the contexts on which it activates most strongly. This works well when the feature's meaning is apparent from its superficial contexts, but can mislead when the feature encodes an abstract assessment whose surface textual correlates are incidental. The J-lens provides a… | mechanisms[5] |
| mechanism | AN ANALOGY IS ASSERTED AND ITS LIMIT IS STATED IN THE SAME SENTENCE, WITH THE UNCLAIMED PART SPELLED OUT: despite these similarities, WE DO NOT CLAIM THAT LANGUAGE MODELS REPRODUCE THE FULL ARCHITECTURE GLOBAL WORKSPACE THEORY ASCRIBES TO THE BRAIN - specialized, encapsulated processors competing for entry to a workspace that broadcasts back to them through recurrent connections. The borrowed theory is named in full precisely where the borrowing stops— “Despite these similarities, we do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain—specialized, encapsulated processors competing for entry to a workspace that broadcasts back to them through recurrent connections” | mechanisms[6] |
| mechanism | THE DISANALOGY IS PUT FIRST AND THE EXPECTATION DENIED: IT IS NOT OBVIOUS THAT THEY SHOULD; in the brain, the workspace is closely associated with recurrent dynamics and brain region interactions THAT HAVE NO DIRECT ANALOG IN THE TRANSFORMER ARCHITECTURE on which LLMs are based. The reason the finding would be surprising is given before the finding is defended— “It is not obvious that they should; in the brain, the workspace is closely associated with recurrent dynamics and brain region interactions that have no direct analog in the transformer architecture on which LLMs are based.” | mechanisms[7] |
| mechanism | TWO EXPLANATIONS OF THE PAPER'S CENTRAL FINDING ARE OFFERED AND NEITHER IS CHOSEN: in this view, the workspace IS VERBALIZABLE BECAUSE THE MODEL'S OUTPUT SPACE IS VERBAL, and the format in which it broadcasts intermediate results to itself IS SHAPED BY THE FORMAT IN WHICH IT MUST EVENTUALLY ACT. The property named in the title is explained by the shape of the output channel rather than by anything about the representation— “In this view, the workspace is verbalizable because the model's output space is verbal, and the format in which it broadcasts intermediate results to itself is shaped by the format in which it must eventually act.” | mechanisms[8] |
| mechanism | A SCOPE LIMIT IS STATED WITH A COINED METAPHOR THAT CONCEDES WHAT THE METHOD CANNOT SEE: most of the model's computation RUNS OUTSIDE THE J-SPACE, and so what the graphs capture is THE VERBALIZABLE SKELETON OF THE COMPUTATION RATHER THAN THE WHOLE OF IT. The instrument's reach is named as a part of its object, not as a defect of the study— “most of the model's computation runs outside the J-space, and so what the graphs capture is the verbalizable skeleton of the computation rather than the whole of it.” | mechanisms[9] |
| mechanism | AND THE PROPERTIES THE ANALOGY DOES NOT ESTABLISH ARE LISTED AS ABSENT: note that there are OTHER STRUCTURAL PROPERTIES ASSOCIATED WITH GLOBAL WORKSPACE THEORY THAT ARE NOT DEMONSTRATED BY OUR ANALYSES. A borrowed framework is adopted in part, and the unadopted part is named rather than left for a reader to notice— “note that there are other structural properties associated with global workspace theory that are not demonstrated by our analyses.” | mechanisms[10] |
| extends | the logit lens— “While the logit lens assumes that representations use the same coordinates in all layers, the Jacobian lens corrects for representational changes that take place across layers, allowing it to uncover meaningful information in earlier layers where the logit lens produces uninterpretable readouts.” | claim_relations[0] |
| builds-on | Hernandez et al., per-relation linear maps from Jacobians— “The J-lens applies the same first-order approximation to the map from activations to model outputs, rather than to a single relation.” | claim_relations[1] |
| replicates | Li et al., logit-lens decoding and patching of multi-hop intermediates— “Closest to our approach, Li et al. use the logit lens to decode the intermediate entity in a multi-hop query from middle layers and patch along the decoded directions to redirect the answer, paralleling the internal-reasoning results of §3.3 .” | claim_relations[2] |
| builds-on | work on implicit chain-of-thought— “More closely related to our method is work on “implicit chain-of-thought” , which establishes that training on auxiliary reasoning text can shape computation in that context, even when the auxiliary text is dropped at inference.” | claim_relations[3] |
| sense | global workspace - the theory's sense: a shared space a representation is posted to, from which many downstream processes can read— “A representation becomes consciously accessible when it is posted to a shared "global workspace" from which many downstream processes can read .” | senses[0] |
| sense | global workspace - a processing hub that integrates and broadcasts, so the information can be used for flexible internal reasoning and report— “Under the theory, the workspace is a processing hub that integrates and broadcasts information, allowing it to be used for flexible internal reasoning and report .” | senses[1] |
| sense | nameable concept - bounded by the tokenizer: exactly those concepts holding a single-token name, which is what the lens can report at all— “This construction means that the set of concepts the lens can name is exactly the set of concepts that have a single-token name in the tokenizer's vocabulary.” | senses[2] |
| sense | global workspace - what the paper explicitly does NOT claim language models reproduce— “we do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain” | senses[3] |
| sense | `J-space component`, and `in` / `aligned with` - A PAPER STIPULATING NOT ONLY A NOUN BUT THE PREPOSITIONS APPLIED TO IT: `When we refer to the “J-space component” of a vector, we mean the component that can be written as such a sparse (nonnegative) sum. When we describe a vector as being “in” or “aligned with” the J-space, we mean that this J-space component is close to the original vector.` THE SECOND SENTENCE IS THE UNUSUAL ONE. A term can be defined and still be misread through the spatial metaphor its own name invites - `in` a space normally means membership, and here it is stipulated to mean APPROXIMATION. The authors fix the preposition because the noun alone would not have stopped the wrong reading.— “When we refer to the “J-space component” of a vector, we mean the component that can be written as such a sparse (nonnegative) sum. When we describe a vector as being “in” or “aligned with” the J-space, we mean that this J-space component is close to the original vector.” | senses[4] |
| sense | `J-space` - A NAME BORROWED FROM LINEAR ALGEBRA FOR SOMETHING THAT IS NOT A SUBSPACE. `This union of cones is the J-space.` The suffix `-space` invites every habit a reader has about spans, bases and dimensions, and the object is a union of polyhedral cones - closed under nonnegative combination only. The paper says so plainly and the name still does the misleading, which is mechanism 28's shape (a name whose extension is wrong) arriving through a suffix rather than through a stem.— “This union of cones is the J-space.” | senses[5] |
| sense | `J-space component` - DEFINED BY WHAT IT CAN BE WRITTEN AS, NOT BY A PROJECTION. `When we refer to the “J-space component” of a vector, we mean the component that can be written as such a sparse (nonnegative) sum.` In a subspace, `component` means an orthogonal projection and is unique. Here it means a best sparse nonnegative approximation - the same word, a different mathematical object, and the difference follows directly from J-space not being a space. NO EDGE PROPOSED, AND THIS REFUSAL IS DIFFERENT FROM THE OTHERS TONIGHT. `part-of` is the RIGHT type here; the paper simply never asserts the containment in words. I searched every sentence mentioning `J-space` for an accepted containment cue and the closest is a QUESTION - `what fraction of the model's representational space belongs to the J-space?` - which asks the thing rather than claiming it. The relation was my inference from the definition, not the paper's statement, and the gate caught exactly that.— “When we refer to the “J-space component” of a vector, we mean the component that can be written as such a sparse (nonnegative) sum.” | senses[6] |
| control | a positive control on the ablation itself, checking it removes the content the J-space was shown to carry — held — rules out the possibility that the ablation did not do what it claims, so that a null downstream would be uninterpretable— “W e first verify, as a positive control, that the ablation removes the kind of content the preceding sections showed the J-space to carry.” | controls[0] |
| control | a story-writing task run under the same ablation, scoring general story quality alongside the rate of experiential language — held — rules out broad degradation of writing capability as the explanation of the drop in experiential language— “A story-writing control confirms the effect is not due to broad degradation of writing capability: ablation only slightly reduces graded story quality, while still reducing the rate of experiential language within those stories ( §A.23 ).” | controls[1] |
| control | zeroing equal-sized, layer-matched sets of randomly chosen heads - the sentence states the control and not its outcome, so no outcome is recorded here — rules out the effect arising from ablating any sufficiently large set of heads rather than these ones— “W e zero the heads' outputs at every token position, and as a control we do the same to equal-sized, layer-matched sets of randomly chosen heads.” | controls[2] |
| control | a perturbation of equal magnitude applied to a different concept of the same category — held — rules out two alternatives at once - generic capability degradation from any perturbation, and regression to the mean from added noise— “As a control, to ensure the effect on avoidance behavior is not simply due to perturbations generically degrading the model’s capabilities, or due to noise leading to regression-to-mean effects, we tried applying a perturbation of equal magnitude, inhibiting the J-lens vector of some other concept of the same category as the primed one.” | controls[3] |
| audit | consciously accessible information —operationalised-as→ global workspace — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span writes 'A representation becomes consciously accessible when it is posted to' the workspace; the endpoint is a variant, and operationalised-as is loose.— “A representation becomes consciously accessible when it is posted to a shared "global workspace" from which many downstream processes can read .” | concept_edges[0] |
| weight | consciously accessible information —operationalised-as→ global workspace — 0.9— “A representation becomes consciously accessible when it is posted to a shared "global workspace" from which many downstream processes can read .” | concept_edges[0] |
| outcome (stated by the tagger, no quotation) | control — held | controls[0] |
| outcome (stated by the tagger, no quotation) | control — held | controls[1] |
| outcome (stated by the tagger, no quotation) | control — held | controls[3] |
| for which use | limitation — every workspace claim in the paper: THE JACOBIAN LENS IS AN IMPERFECT TOOL, WHICH WE BELIEVE ONLY APPROXIMATELY AND INCOMPLETELY CAPTURES THE MODEL'S UNDERLYING WORKSPACE STRUCTURE— “The Jacobian lens is an imperfect tool, which we believe only approximately and incompletely captures the model’s underlying workspace structure.” | limitations[0] |
| for which use | limitation — the completeness of any concept list the method produces: THE SET OF CONCEPTS THE LENS CAN NAME IS EXACTLY THE SET OF CONCEPTS THAT HAVE A SINGLE-TOKEN NAME IN THE TOKENIZER'S VOCABULARY - the instrument's lexicon bounds what is findable— “This construction means that the set of concepts the lens can name is exactly the set of concepts that have a single-token name in the tokenizer's vocabulary.” | limitations[1] |
| for which use | limitation — any structural or relational reading of the results: THE J-LENS TREATS THE WORKSPACE AS A FLAT COLLECTION OF INDEPENDENTLY ACTIVE CONCEPT VECTORS, so binding is not represented— “Even setting the vocabulary issue aside, the J-lens treats the workspace as a flat collection of independently active concept vectors.” | limitations[2] |
| for which use | limitation — any readout taken from the lens: AT SOME POSITIONS AND LAYERS, THE TOP LENS TOKENS ARE ONES WE CANNOT MAKE SENSE OF, and the rate was not characterised systematically— “In our experience this is the typical case at workspace-layer positions, but it is not the universal one: at some positions and layers, the top lens tokens are ones we cannot make sense of.” | limitations[3] |
| for which use | limitation — the early-layer result, where object and instrument cannot be separated: WE DO NOT KNOW WHETHER THIS IS A FACT ABOUT THE MODEL OR AN ARTIFACT OF THE LENS— “W e do not know whether this is a fact about the model or an artifact of the lens.” | limitations[4] |
| for which use | limitation — any mechanistic account of the workspace: WE HAVE NOT CHARACTERIZED WHAT CAUSES A REPRESENTATION TO ENTER IT— “W e have not characterized what causes a representation to enter it.” | limitations[5] |
| for which use | limitation — any safety or alignment-monitoring use: WE DO NOT FEEL COMFORTABLE MAKING THE STRONGER CLAIM THAT MONITORING THE J-SPACE IS SUFFICIENT FOR ALIGNMENT MONITORING, OR THAT ANY SOPHISTICATED PLAN THE MODEL MIGHT EXECUTE MUST BE REPRESENTED THERE— “W e do not feel comfortable making the stronger claim that monitoring the J-space is sufficient for alignment monitoring, or that any sophisticated plan the model might execute must be represented there.” | limitations[6] |
| for which use | limitation — the generality across scale and training: WE HAVE STUDIED THE WORKSPACE PRIMARILY IN FULLY TRAINED PRODUCTION LLMs OF CONSIDERABLE SIZE— “W e have studied the workspace primarily in fully trained production LLMs of considerable size.” | limitations[7] |
| for which use | limitation — the analogy to the neuroscience theory: WE DO NOT CLAIM THAT LANGUAGE MODELS REPRODUCE THE FULL ARCHITECTURE GLOBAL WORKSPACE THEORY ASCRIBES TO THE BRAIN— “Despite these similarities, we do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain” | limitations[8] |
| evidence status note | mechanism — The sentence begins with a word the extractor split - "W e" for "We" - so the span starts at the second word. This file has 469 such splits and no glyph codes at all, so nothing marks them but reading. PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “define a subset of vector representations as workspace-like if it satisfies the following properties, which mirror the properties characteristic of conscious access described above” | mechanisms[0] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “But this judgment was somewhat post-hoc, and we did not provide a principled definition of what distinguishes a "workspace" representation from a "motor" one.” | mechanisms[1] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “A fuller account would either supply such a predictive criterion, or would replace the flexible/automatic distinction with something more precise.” | mechanisms[2] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “The tasks in §3.5 were chosen because the flexible/automatic distinction is clear for them; however, for many tasks of practical interest it is not, and we have no way of predicting in advance, for an arbitrary computation, whether it will engage the J-space.” | mechanisms[3] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “here we define the reaction concept to be whatever answer the model itself goes on to produce, so the question is whether the model's eventual choice is already in the lens while it reads the instruction to withhold it.” | mechanisms[4] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “The standard approach to interpreting an SAE feature is to inspect the contexts on which it activates most strongly. This works well when the feature's meaning is apparent from its superficial contexts, but can mislead when the feature encodes an abstract assessment whose surface textual correlates are incidental. The J-lens provides a… | mechanisms[5] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “Despite these similarities, we do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain—specialized, encapsulated processors competing for entry to a workspace that broadcasts back to them through recurrent connections” | mechanisms[6] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “It is not obvious that they should; in the brain, the workspace is closely associated with recurrent dynamics and brain region interactions that have no direct analog in the transformer architecture on which LLMs are based.” | mechanisms[7] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “In this view, the workspace is verbalizable because the model's output space is verbal, and the format in which it broadcasts intermediate results to itself is shaped by the format in which it must eventually act.” | mechanisms[8] |
| evidence status note | mechanism — PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “most of the model's computation runs outside the J-space, and so what the graphs capture is the verbalizable skeleton of the computation rather than the whole of it.” | mechanisms[9] |
| evidence status note | mechanism — The sentence begins with a word the extractor split - "W e" for "We" - so the span starts at the second word. This file has 469 such splits and no glyph codes at all, so nothing marks them but reading. PROVENANCE: this record belongs to the priming sub-corpus, which is the other terminal's lane, and these mechanisms were written by the project. No two hands were on the record - it carried no mechanisms before these, and the project confirmed on 2026-08-17 that the project is auditing rather than tagging. Kept rather than retracted on his instruction (17:22): retracting verified work over a bookkeeping flag would leave a retraction reason that misdescribes the row. Recorded here so a reader learns it from the corpus. This one was authorised in advance: the project named it at 16:39 - “bouchaud2008, gurnee2026, pelagios2023, dekel2021 - do those” - before the merge. The note is here because the lane fact is true regardless of who approved it.— “note that there are other structural properties associated with global workspace theory that are not demonstrated by our analyses.” | mechanisms[10] |
| replication | untested |
| limitation | The lens is acknowledged as an approximate and incomplete proxy for whatever the model's real workspace is.— “The Jacobian lens is an imperfect tool, which we believe only approximately and incompletely captures the model’s underlying workspace structure.” | limitations[0] |
| limitation | It can only name concepts that happen to be single tokens in the vocabulary.— “This construction means that the set of concepts the lens can name is exactly the set of concepts that have a single-token name in the tokenizer's vocabulary.” | limitations[1] |
| limitation | The readout is a flat bag of concepts and cannot show how they are bound or structured together.— “Even setting the vocabulary issue aside, the J-lens treats the workspace as a flat collection of independently active concept vectors.” | limitations[2] |
| limitation | Readouts are not always interpretable, and the rate of uninterpretable ones was not characterised systematically.— “In our experience this is the typical case at workspace-layer positions, but it is not the universal one: at some positions and layers, the top lens tokens are ones we cannot make sense of.” | limitations[3] |
| limitation | The apparent emptiness of the early layers may be a fact about the model or an artefact of the lens; the authors cannot tell which.— “W e do not know whether this is a fact about the model or an artifact of the lens.” | limitations[4] |
| limitation | No mechanism is offered for how a representation comes to enter the workspace.— “W e have not characterized what causes a representation to enter it.” | limitations[5] |
| limitation | The authors decline to claim the J-space is sufficient for alignment monitoring, since automatic or well-practised misaligned behaviour could bypass it.— “W e do not feel comfortable making the stronger claim that monitoring the J-space is sufficient for alignment monitoring, or that any sophisticated plan the model might execute must be represented there.” | limitations[6] |
| limitation | Findings are from large production models only; scaling and training-time emergence are untested.— “W e have studied the workspace primarily in fully trained production LLMs of considerable size.” | limitations[7] |
| limitation | The analogy to global workspace theory is explicitly partial: architectural features such as encapsulated processors and recurrent broadcast are not claimed.— “Despite these similarities, we do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain” | limitations[8] |