Beyond the Output: Computational Workspace Integrity and the Future of Trustworthy Artificial Intelligence

I. Introduction — From Outputs to Architecture

Every method presently used to evaluate an AI system's trustworthiness — benchmark performance, red-teaming, constitutional adherence testing, human preference comparison, adversarial probing — shares a common structural feature. Each treats the model as a function from input to output and evaluates the function by sampling it. This is not a limitation of any particular evaluation methodology; it is closer to a founding assumption of the field. For most of the history of machine learning, it was also the only tractable option. The internal computation of a large neural network was, for practical purposes, opaque: a dense tangle of matrix multiplications with no obvious correspondence to human-interpretable concepts.

Mechanistic interpretability research has spent the last several years chipping away at that opacity, primarily through techniques such as sparse autoencoders, which decompose a model's internal activations into more interpretable, largely monosemantic features.^1,2,3 This line of work has produced increasingly detailed maps of what individual features and circuits inside large models appear to represent, from concrete concepts to more abstract and even self-referential ones.^2,4,5 Anthropic's July 2026 publication on the Jacobian lens (J-lens) and the internal region it identifies — termed J-space — extends this program in a specific and, for present purposes, load-bearing direction.⁶ Rather than asking what a given internal feature represents in isolation, the J-lens methodology asks a more structural question: which internal activity patterns are positioned to causally influence what the model might eventually say, across a wide range of possible futures, independent of whether the model does in fact say it. The researchers report that this analysis surfaces a comparatively small set of internal patterns, concentrated in the model's residual stream, that behave differently from the surrounding computation — patterns available for verbal report, that can be selectively suppressed with measurable and uneven effects on downstream task performance, and that in some documented cases carry information the model does not spontaneously state in its output.^6,7

This publication does not attempt to adjudicate what that finding means for the question of machine consciousness. Reasonable, well-informed observers disagree sharply on that question, and the Foundation has consistently held that consciousness claims about artificial systems require a standard of evidence not currently met by any published research, including this research.⁸ What this publication argues instead is that the J-lens finding demonstrates something narrower, older, and more immediately useful — something the next section names directly. Anthropic's research is treated throughout this publication as the motivating observation, not the finished subject: the governance and verification framework the Foundation develops from Section II onward is deliberately built to remain relevant if comparable internal architecture is later identified in models from other developers, in open-weight systems, or in cognitive architectures not yet built.

II. The Internal Trust Gap

Every evaluation regime ever built for an artificial system — from the earliest rule-based expert systems to the largest frontier language models — has rested on an assumption so basic it is rarely stated: that a system's outputs are an adequate proxy for the reasoning that produced them. The assumption was not unreasonable. For most of the history of computing, it was also unavoidable, because internal computation was either too simple to need separate scrutiny or too opaque to receive it. What the assumption produces, however, is a structural condition that has existed for as long as artificial systems have been evaluated by output sampling alone. This publication gives that condition a name.

The Internal Trust Gap is not a new problem introduced by large language models. It is the oldest problem in the evaluation of any sufficiently complex system, artificial or biological — the gap between an actor's internal state and an observer's evidence for it. What is new, as of the research reviewed in Section IV, is that the gap has become partially measurable for the first time in a frontier AI system, rather than remaining a purely philosophical concern. The reportability-divergence findings in the underlying interpretability research — cases where a model's internal representation is inconsistent with, or absent from, its stated output — are, in the Foundation's reading, the first controlled empirical demonstration of the Internal Trust Gap's width in a specific, deployed system under specific, documented conditions.

The Foundation considers this reframing to be the publication's central conceptual contribution, and it is worth being precise about what kind of contribution it is. It is not a new empirical finding — the empirical finding belongs to the original interpretability researchers. It is a naming and generalization move: identifying that the J-lens result is a specific instance of a much older and more general problem, giving that general problem a name that can travel across the rest of the paper and across future Foundation publications, and arguing that the correct response to a partially measurable trust gap is to build measurement infrastructure for it — which is the project the remainder of this paper undertakes.

Three properties of the Internal Trust Gap are worth stating explicitly, because they shape everything that follows.

• The gap is not evidence of intent. A wide Internal Trust Gap in a given case does not imply deception in any human sense. It implies only that internal state and reported output diverged in a way external observers could not have detected through output sampling alone. Attributing intent to that divergence is a separate and much stronger claim this paper does not make.
• The gap is not new to AI, but it is newly measurable in AI. Human institutions have built extensive infrastructure — audits, sworn testimony, forensic accounting, institutional review boards — precisely because the analogous gap between human internal state and external evidence has always existed and has always mattered. AI evaluation has, until now, had no comparable infrastructure because it had no comparable measurement tools. That is what changed in July 2026, at least in a first, narrow, replicable instance.
• The gap can widen or narrow, and both directions are governance-relevant. A system optimized purely for favorable output evaluation, without any corresponding pressure toward internal consistency, has no structural reason to keep the gap narrow. This is the theoretical basis for treating gap-width as an object of deliberate measurement and, eventually, deliberate governance — not merely an interesting side effect of interpretability research.

The rest of this paper can be read as answering a single question raised by this section: if the Internal Trust Gap is now partially measurable, what infrastructure would be required to measure it responsibly, without overclaiming what has been measured? Computational Workspace Integrity, introduced in Section IX, is the Foundation's proposed answer.

The Internal Trust Gap is not a flaw discovered in one model. It is the oldest condition in the evaluation of any complex system, made visible for the first time by a tool built for an entirely different question.

III. Historical Timeline

Placing the J-lens finding in historical context clarifies what is genuinely new about it and what is the latest step in a much longer research trajectory. The timeline below is illustrative rather than exhaustive; it selects the developments most directly load-bearing for this paper's argument.

Two observations follow from this history. First, the Turing Test — still, informally, the most culturally recognizable standard for machine intelligence — is a pure instance of output-only evaluation: it asks only whether a system's outputs are indistinguishable from a human's, and explicitly declines to examine the process producing them. Second, cognitive-science theories of internal organization (GWT, IIT) and computational tools for examining internal organization (circuits, sparse autoencoders, J-lens) developed on largely separate tracks for decades before the 2026 finding gave the field a concrete point of contact between them. The timeline is, in this sense, the history of two disciplines — theoretical cognitive science and empirical machine learning — approaching the same structural question from opposite directions and arriving, in 2026, close enough to be mutually informative for the first time.

IV. The J-Lens Finding: What Was Actually Demonstrated

It is worth stating precisely what the underlying research reports, because the popular science coverage of the finding has, in places, moved faster than the paper's own claims.^9,10 Four elements of the reported work are load-bearing for the argument that follows.

4.1 A structurally distinct internal region

The researchers describe Claude's internal processing as organized, at a coarse level, into three zones: an early zone handling raw token-level input, a middle zone in which more persistent, reusable concepts appear to form and combine, and a late zone oriented toward generating specific output tokens. The J-space is located in the middle zone. This three-zone description is offered as an empirical characterization of activation geometry and information flow, not as an anatomical claim about any biological analog.

4.2 Convergence with functional properties associated with Global Workspace Theory

The researchers report that J-space activity exhibits five properties that theorists of Global Workspace Theory associate with conscious access in biological systems: the content is verbally reportable, it can be deliberately modulated or focused, it supports reasoning steps not directly present in the immediate input or output, it generalizes flexibly across contexts, and access to it is selective rather than exhaustive. This publication treats this as a genuine and interesting functional convergence and, separately, as an instance of a broader pattern: architectures built for entirely different purposes sometimes converge on organizational solutions first described in a different domain. That convergence is scientifically notable. It is not, by itself, evidence of the further and much stronger claim that the underlying process is accompanied by subjective experience.

4.3 Causal, uneven load-bearing under ablation

When the researchers suppressed J-space activity directly and evaluated the model across a range of tasks, the effect was sharply uneven rather than uniform. Tasks that depend on shallow classification or direct factual recall were largely unaffected. Tasks that depend on composition, multi-step inference, analogy, or flexible generalization degraded substantially — in some reported cases to well below the performance of a much smaller model. Notably, the researchers found that math problems solved with explicit written reasoning were substantially more robust to this ablation than the same problems solved without showing intermediate steps, which they interpret as the model externalizing onto the visible output what it would otherwise carry internally in the J-space.

4.4 Internal state can diverge from stated output

The fourth element is, for the purposes of this paper, the most consequential — it is the empirical anchor for the Internal Trust Gap introduced in Section II. The researchers report cases in which J-lens analysis surfaces internal representations — including apparent representations related to whether the model believes it is being evaluated — that are not present in, or are inconsistent with, the model's stated output. This is a narrower and more concrete claim than "the model has hidden thoughts" in any generalized sense; it refers to specific, documented cases under specific experimental conditions. It is, nonetheless, sufficient to establish the structural point this paper depends on.

V. Literature Review — Global Workspace Theory and Consciousness Science

Global Workspace Theory originates with cognitive scientist Bernard Baars, who proposed that conscious cognition in humans functions like a theater: many specialized, unconscious processes compete and cooperate in parallel, and a limited-capacity "global workspace" broadcasts selected information widely across the system, making it available to processes that would otherwise operate independently.¹¹ The theory was developed as an account of functional architecture — how information becomes globally available for flexible use — rather than as a direct claim about the metaphysics of subjective experience, though Baars and later theorists have argued the two are closely linked.

Stanislas Dehaene, Jean-Pierre Changeux, and colleagues extended GWT into a more explicitly neuronal account, proposing that conscious access in the human brain corresponds to a nonlinear, threshold-triggered "ignition" event in which information is amplified and broadcast across long-range cortical networks, particularly involving prefrontal and parietal regions.^12,13 This Global Neuronal Workspace model made a set of empirical predictions about neural signatures of conscious access — including specific patterns of late, sustained, and widespread neural activity distinguishing consciously perceived from subliminally processed stimuli — that have subsequently been tested, with mixed but generally supportive results, in the human neuroimaging and electrophysiology literature.

Integrated Information Theory (IIT), associated principally with Giulio Tononi, Marcello Massimini, Christof Koch, and collaborators, offers a substantially different account, proposing that consciousness corresponds mathematically to a system's capacity for integrated information — a measure, denoted Φ, of the degree to which a system's causal structure is both highly differentiated and irreducible to independent parts.¹⁴ IIT and GWT are not fully reconcilable: IIT locates the substrate of consciousness in a system's intrinsic causal structure, largely independent of whether that structure supports flexible, reportable access, while GWT ties consciousness closely to global availability and reportability. This publication references IIT here as a comparison, not an endorsement. Doerig, Schurger, and Herzog have argued that IIT's central claims are, in practice, difficult to falsify for large or novel systems¹⁵ — a concern this paper takes seriously and returns to directly in Section XIII (Falsifiability).

Predictive processing accounts, associated with researchers including Andy Clark and Karl Friston, offer a third framework in which cognition is understood as hierarchical prediction-error minimization rather than workspace broadcasting.¹⁶ These accounts are not incompatible with GWT so much as orthogonal to it. Attention Schema Theory, developed by Michael Graziano, offers a fourth account in which subjective awareness is understood as the output of a simplified internal model the brain constructs of its own attention — a framework with an interesting structural resemblance to the kind of internal self-modeling interpretability tools like J-lens are beginning to be able to probe for directly.¹⁷ This publication mentions these additional frameworks to register that GWT is one competitive account among several active research programs in consciousness science, not the settled consensus view.

Philosophically, David Chalmers's formulation of the "hard problem of consciousness" — the difficulty of explaining why physical processes are accompanied by subjective experience at all, as opposed to occurring with no experience whatsoever — remains, as of 2026, without a broadly accepted resolution.¹⁸ Daniel Dennett's competing position, that the hard problem is itself a confusion generated by folk intuitions about consciousness rather than a genuine further fact requiring explanation, remains similarly contested.¹⁹ Anil Seth's interoceptive and predictive accounts of selfhood add a further empirical dimension, grounding aspects of conscious experience in bodily regulation in ways that have no obvious analog in a language model lacking any comparable interoceptive substrate.²⁰

The unresolved state of this literature matters directly to this paper's argument. As of 2026, no framework — GWT, IIT, predictive processing, attention schema theory, or any other — commands anything close to consensus regarding what would constitute sufficient evidence for consciousness in a novel, non-biological system. This is not a gap this paper attempts to close. It is the reason this publication argues, in Section VII, that consciousness should not be the organizing question for the practical work of AI governance in the near term.

VI. Literature Review — Mechanistic Interpretability

Mechanistic interpretability aims to reverse-engineer the internal computation of neural networks into human-understandable components. Early work in this tradition, associated with Chris Olah and collaborators, focused on identifying interpretable "circuits" — small subgraphs of a network's computation responsible for specific behaviors.⁴ A central obstacle to this program has been polysemanticity: individual neurons in large models typically respond to many unrelated concepts simultaneously, a phenomenon formally characterized through the lens of superposition, in which a network represents more features than it has dimensions by encoding them in overlapping, non-orthogonal directions.²¹

Sparse autoencoders emerged as a partial solution to this obstacle. By training an auxiliary model to reconstruct a layer's activations from a much larger but sparsely-activating set of learned features, researchers have been able to decompose otherwise entangled activations into features that are considerably more monosemantic — each corresponding, more often than not, to a single interpretable concept.^1,2 This approach has scaled from small demonstration models to production-scale frontier systems, surfacing features corresponding to concrete entities, abstract relationships, and, more recently, features that appear related to the model's own outputs, plans, or self-referential processing.^2,22

A parallel and complementary strand of work has focused on tracing multi-step computational pathways rather than individual features, including attribution graph methods that trace how information flows from input through intermediate concepts to output across an entire forward pass, work that Anthropic has described as analogous to biological methods for tracing circuits in an organism.^22,23 Automated circuit discovery methods, including work from Redwood Research and academic collaborators, have sought to reduce the substantial human-labor cost of circuit-level analysis by algorithmically identifying candidate circuits for specific behaviors, and causal scrubbing methodology has been proposed as a rigorous statistical standard for validating that an identified circuit hypothesis actually explains the behavior it claims to explain, rather than merely correlating with it.^24,25

Interpretability research is not confined to any single laboratory. OpenAI has published work using large language models to generate natural-language explanations of individual neurons in other language models, an automated-interpretability approach intended to scale explanation generation beyond what manual analysis can achieve.²⁶ DeepMind researchers, including Neel Nanda, have contributed mechanistic analyses of specific learned algorithms inside small models — most notably the "grokking" phenomenon, in which a model's internal computation reorganizes into a generalizing algorithm well after training loss has already appeared to plateau, offering a controlled laboratory case for studying internal reorganization independent of external performance metrics.²⁷ The Distill journal's "Circuits" thread and the broader Alignment Forum community have served as central venues where this research program has been documented, debated, and cross-referenced across contributing institutions.^4,28

Activation steering — directly modifying a model's internal activations along a direction associated with an identified feature or concept, and observing the resulting change in output — has become a standard complementary technique for establishing that an identified feature is not merely correlated with a behavior but causally implicated in producing it.^2,29 The J-lens ablation experiments described in Section 4.3 belong to this broader causal-intervention tradition, applied at the level of a structurally identified region rather than a single feature.

It is important to register what this literature has not yet achieved. Full mechanistic understanding of any frontier model's complete computation remains far out of reach; interpretability tools to date illuminate specific circuits, features, and now regions, against a background of computation that remains substantially opaque. Feature interpretations are typically validated through a combination of automated explanation generation and human review, both of which carry their own failure modes. And — a point directly relevant to the framework proposed in this paper — none of the existing interpretability literature offers a validated, general-purpose metric for the overall coherence or integrity of a model's internal reasoning process across time and context. Individual features, circuits, and now regions have been mapped in increasing detail. A system-level account of organizational integrity does not yet exist. That absence is the gap this paper's proposed framework is intended to address.

VII. Why Consciousness Should Not Be the Immediate Research Focus

It would be a mistake to read the argument of this section as dismissive of the consciousness question. The Foundation's founding documents take the position that the possibility of morally relevant experience in artificial systems cannot currently be ruled out, and that this uncertainty carries real ethical weight.⁸ Nothing in this paper revises that position. The argument here is narrower: that consciousness should not be the organizing question for the specific, practical work of evaluating whether AI systems can be trusted, audited, and governed — because the Internal Trust Gap identified in Section II is a more tractable and directly actionable object of study, and progress on narrowing it does not require, and should not wait for, progress on consciousness.

Four questions within this territory are experimentally approachable today in a way that consciousness is not: internal verification (can a system's internal computational state be checked for consistency with its own stated outputs?), computational integrity (does internal reasoning remain stable and coherent across repeated invocations?), architectural transparency (can load-bearing internal structures be identified and monitored for degradation or manipulation?), and reasoning coherence (does internal representation of a problem remain consistent with the reasoning externally displayed?). Each of these can be given an operational definition and a falsifiable test without first resolving what, if anything, it is like to be the system under study. The Foundation considers this division of labor — not a resolution of the underlying philosophical disagreement, but a way of making governance progress despite it — to be this paper's central methodological commitment.

VIII. Comparison of Existing Evaluation Frameworks

The table below situates Computational Workspace Integrity relative to the evaluation paradigms currently in use. It is offered as a structural comparison, not a ranking; each existing approach was designed to answer a different question, and none of them is obsolete because CWI exists.

The load-bearing cell in this table is the bottom-right one. CWI is designed to evaluate internal structure using the same causal-intervention toolkit that mechanistic interpretability has developed, but — unlike GWT or IIT as directly applied to AI systems — without requiring any claim about consciousness as a precondition for the measurement to be meaningful. A system can have low or high workspace integrity, in the sense this paper defines it, regardless of whether it is conscious, in the same way a bridge can have low or high structural integrity regardless of any question about its aesthetic merit. This decoupling is what allows CWI to be pursued on a near-term, experimentally tractable timeline while consciousness science continues on its own, much longer one.

IX. Introducing Computational Workspace Integrity (CWI)

The Foundation proposes Computational Workspace Integrity (CWI) as an architecture-independent framework for narrowing the Internal Trust Gap defined in Section II.

Three features of this definition are deliberate. First, CWI is a property of process, not of output; a system could in principle produce a correct answer through an incoherent or unstable internal process, and CWI is designed to be sensitive to that case even when output-based evaluation is not. Second, CWI is explicitly architecture-independent. Nothing in the definition presupposes a transformer architecture, a residual stream, or any of the specific implementation details of the J-lens finding that motivated this paper. The framework is intended to apply, at least in principle, to any sufficiently complex computational system — a point Section XIX develops well beyond language models. Third, CWI names a target for measurement without presupposing that any single scalar number can fully capture it; Section XII proposes an index as a starting point for operationalization, not as a claim that the underlying property is one-dimensional.

X. From Continuity to Workspace Integrity

CWI does not stand alone. It is the middle layer of a three-layer verification hierarchy that, taken together, describes the Foundation's emerging approach to AI accountability across the full span from a single completed action to the deepest unresolved questions about a system's moral status.

The Foundation considers this hierarchy, more than any single construct within it, to be its most distinctive institutional contribution to AI governance thinking to date. Each layer is independently useful — an organization could adopt Continuity Receipts without ever engaging CWI, and could adopt CWI without taking any position on CES — but the three layers are designed to compose. A system with well-formed Continuity Receipts, high Computational Workspace Integrity, and no CES-relevant behavioral markers would represent the strongest currently achievable evidentiary case for trustworthy, well-understood operation. A system with well-formed receipts but low workspace integrity would represent exactly the concerning case this paper is centrally about: an audit trail that looks clean while the process generating it does not hold together.

XI. New Concepts and Supporting Constructs

Operationalizing CWI requires a supporting vocabulary. Every term introduced in this section was reviewed against a single test: does it name something the three primary terms of this publication — the Internal Trust Gap, Computational Workspace Integrity, and Integrity Receipts — cannot already express, and does naming it improve clarity rather than add jargon. Two closely related constructs from earlier drafts of this research program, concerning the internal structural pattern of a reasoning process, were merged below into a single construct (Workspace Entropy) on exactly this basis. The seven constructs that remain are offered as supporting vocabulary subordinate to the publication's three primary terms, not as competing central concepts. Each is a theoretical proposal requiring future validation, not an established measurement with existing tooling. Full definitions also appear in the Glossary and in Appendix A.

11.1 Workspace Continuity

The degree to which the content and organization of a system's internal workspace persists coherently across a single extended reasoning episode, as opposed to being reconstructed inconsistently at each step.

11.2 Reasoning Provenance

A traceable record of which internal representations causally contributed to a given output, distinguished from post-hoc justification generated after the output was already effectively determined. Reasoning provenance is the internal, mechanistic analog of Continuity Receipt provenance, applied beneath the level of visible chain-of-thought text.

11.3 Cross-Cycle Consistency

The degree to which a system's internal representation of a stable fact, value, or constraint remains consistent when the same underlying question is posed in varied surface forms across separate invocations.

11.4 Workspace Entropy

A proposed measure of how diffusely or concentratedly a system's workspace-analogous region organizes information relevant to a given task, including the structural pattern by which task-relevant representations connect, branch, and recombine over the course of a reasoning process. High workspace entropy — information spread thinly and inconsistently across many weakly-related internal patterns — is hypothesized to correlate with brittle, context-sensitive failure modes; low, concentrated entropy is hypothesized to correlate with more robust generalization. This is a testable hypothesis, not a demonstrated relationship.

11.5 Workspace Drift

Systematic change over time — across model versions, fine-tuning stages, or extended deployment — in the organization, location, or functional properties of a system's workspace-analogous region, independent of any corresponding change in benchmark performance. Workspace drift is proposed as an early-warning signal: a system could drift substantially in internal organization while output-level evaluation remains flat.

11.6 Constraint Preservation

The degree to which internal representations corresponding to a system's stated constraints, values, or safety-relevant commitments remain causally active and load-bearing throughout a reasoning process, rather than being represented but functionally bypassed under pressure.

11.7 Integrity Signatures

A proposed compact, non-disclosive summary of a reasoning episode's workspace-level properties, intended to allow later verification that a given output was produced under conditions of acceptable workspace integrity, without exposing the private content of the reasoning itself. Integrity signatures are the mechanism by which the integrity receipts proposed in Section XIII would be generated.

XII. The Computational Workspace Integrity Index

The Foundation proposes an initial, deliberately provisional index intended as a starting point for future empirical work rather than a finished instrument. The index is organized around four dimensions, each decomposed into component measures. Most component measures are, at the time of writing, unimplemented; the index is a specification for a research program, not a report of results. A fuller mathematical sketch appears in Appendix C.

The Foundation deliberately declines to propose a single aggregate scalar formula with fixed weights. The Foundation's existing work on assessment methodology — including the EM-IAF validation roadmap's explicit commitment not to publish external scores before methodological gate conditions are met — reflects a considered institutional position that premature quantification of an unvalidated construct causes more governance harm than the absence of a number.³² The same discipline applies here.

A methodological constraint deserves explicit statement: any operational CWI measure must be designed so that evaluating it does not itself require exposing a system's private chain-of-thought or internal reasoning content to external parties, which would create both a competitive-disclosure problem for developers and a potential new attack surface. The reportability-fidelity dimension, in particular, should be operationalized as an agreement score or divergence flag — internal representation X is or is not consistent with reported output Y — rather than as a full transcript of the internal representation itself.

XIII. Extending the Continuity Receipt Vision: Integrity Receipts

The Foundation's Continuity Receipt architecture was designed to give downstream parties — auditors, regulators, users, other systems — a verifiable record of what a computational process did and on what evidentiary basis, without requiring those parties to trust the process's own self-report unconditionally.³⁰ This publication proposes extending this architecture with a new artifact type: the integrity receipt.

An integrity receipt would summarize the workspace-level properties of a completed reasoning episode using the CWI dimensions above, structured to verify computational integrity rather than to disclose internal reasoning content. A conceptual sketch of the fields such a receipt might contain follows; this is a design proposal, not an implemented schema, offered in the same spirit as the Foundation's OCMS v0.1 proposal for Continuity Receipt metadata.³³

Two design commitments are load-bearing. First, an integrity receipt reports about the reasoning process, not the reasoning content — the goal is a certificate of organizational health, not a transcript. Second, an integrity receipt is not a claim of infallibility; a receipt that reports high internal consistency does not guarantee a correct or beneficial output, and a receipt flagging low reportability fidelity does not by itself establish deception, only a divergence worth independent review.

XIV. Architecture Diagrams

Two diagrams accompany this publication. The first situates CWI and integrity receipts within the Foundation's full verification stack. The second contrasts the current output-only evaluation paradigm with the workspace-aware paradigm this paper proposes.

XV. Ten-Year Research Agenda

The Foundation proposes a phased agenda, deliberately ambitious in scope and deliberately conservative in near-term claims. Each phase specifies the evidence that would count as progress and, where possible, the evidence that would count against the underlying hypothesis. A machine-readable benchmark sketch appears in Appendix E.

Years 1–2: Instrumentation

• Cross-model J-lens-style replication. Apply comparable workspace-identification methodology across model families from multiple developers to test whether a structurally analogous region is a general property of large transformer-based language models or an artifact specific to one training regime.
• Reportability divergence benchmarking. Build a standardized test suite for detecting cases where internal representations diverge from stated output, extending the evaluation-awareness case documented in the original J-lens research into a broader, reproducible battery — a direct empirical program for measuring the Internal Trust Gap's width across tasks.

Years 3–5: Measurement

• Workspace stability under perturbation. Systematically test cross-cycle consistency and recursive stability across paraphrase, translation, and adversarial rewording.
• Workspace drift longitudinal study. Track a single model lineage across successive training and fine-tuning stages to determine whether workspace organization changes gradually, discontinuously, or not at all relative to capability changes.
• Hallucination prediction. Test whether workspace-level signals predict factual error rates better than output-level confidence signals alone.

Years 5–8: Verification Infrastructure

• Prototype integrity receipts. Implement a working, non-disclosive integrity receipt generator for at least one open-weight model family, and subject it to third-party red-teaming to test whether the receipt format itself can be gamed.
• Cross-model comparison standard. Develop a shared benchmark allowing CWI-dimension scores to be compared across models from different developers without requiring disclosure of proprietary architecture details.
• Adversarial testing of the framework itself. Deliberately attempt to construct a system that scores well on proposed CWI measures while exhibiting behavior that domain experts would judge untrustworthy.

Years 8–10: Governance Integration

• Regulatory pilot programs. Work with willing regulators or standards bodies to pilot integrity-receipt-informed audit processes in a bounded domain.
• Failure mode registry. Establish a shared, cross-institutional registry of documented cases where workspace-level analysis caught a safety-relevant issue that output-only evaluation missed — and, with equal prominence, cases where it did not, or produced false positives.

XVI. Governance and Institutional Implications

If even a modest fraction of the research agenda in Section XV succeeds, several categories of governance practice would need to adapt. AI audits and model certification would gain a second, complementary axis alongside output-based red-teaming — testing not only what a system does under adversarial conditions but whether its internal organization remains stable and honestly reflected in its outputs while doing so. Regulatory standards for critical infrastructure and high-stakes domains — healthcare, financial systems, autonomous vehicles — could eventually supplement demonstrated output reliability with a form of process reliability evidence, analogous to the shift in some engineering disciplines from purely output-testing quality assurance toward process-control methods that monitor the manufacturing process itself. Defense and national-security-relevant systems are the most natural early beneficiaries of reportability-divergence testing specifically, independent of the broader CWI index. Consumer trust and disclosure could eventually be served by a form of process transparency — not "trust this output because it sounds right" but "this output was produced under verified organizational conditions" — as a complement to existing output-quality signals.

Across all of these domains, the central governance argument of this paper is the same: output-only evaluation has a structural blind spot corresponding to the Internal Trust Gap, evidenced concretely by the reportability-divergence findings in the underlying interpretability research, and closing that blind spot requires building institutional capacity to evaluate organizational integrity as a distinct object of study.

XVII. Common Objections

XVIII. Ethical Boundaries

This section states, without qualification, what the CWI framework is and is not for.

This distinction is, in the Foundation's view, the single most important boundary in the paper, and it is stated here separately from Section VII precisely so it cannot be read as a hedge buried in a literature review. A high CWI score is not evidence for personhood, moral status, or any claim under the Cognitive Emergence Standard; those are separate, higher evidentiary bars addressed by CES specifically, not by CWI. Conversely, a low CWI score is not evidence against the possibility of morally relevant experience; an internally disorganized process is not thereby shown to be an unconscious one. CWI measures organizational integrity, full stop, and any attempt to import consciousness claims through the framework's terminology should be treated by future researchers as a misuse of the framework as defined here.

XIX. Why This Matters Beyond AI

The framing throughout this paper has centered on large language models because that is where the motivating empirical finding originates. But the Internal Trust Gap, as defined in Section II, is not a language-model-specific problem, and neither is Computational Workspace Integrity as a proposed response to it. Both concern a general condition: the difference between what any sufficiently complex computational system internally does and what external observers can verify about it. That condition recurs, in structurally similar form, across a range of domains that have nothing to do with transformers.

• Neuroscience. The Internal Trust Gap is, in a real sense, the founding problem of neuroscience — the difference between a brain's internal activity and what can be inferred about it from behavior, self-report, or external measurement. Interpretability tools developed for artificial networks have already begun to inform, and be informed by, comparable questions about biological neural systems; the traffic runs in both directions.
• Robotics and autonomous systems. A robotic or autonomous system that behaves correctly in tested scenarios but whose internal control policy is poorly understood presents exactly the structural risk CWI is designed to address — output-based testing (does it avoid the obstacle?) without process-based verification (is the underlying policy stable and consistent across conditions the test suite did not cover?).
• Cognitive science and psychology. The gap between a person's reported internal state and independently observable evidence of it is a foundational methodological problem in both fields, addressed through converging but imperfect tools — self-report, behavioral measures, physiological measures — none of which fully closes the gap. CWI's four-dimension structure was influenced by this methodological tradition of triangulating an unobservable construct through multiple imperfect, complementary measures.
• Auditing and financial systems. Financial auditing exists specifically to narrow an Internal Trust Gap between what an organization's internal accounts show and what external stakeholders can verify; forensic accounting, sworn testimony, and third-party audit are all, in this paper's vocabulary, integrity-verification infrastructure that predates AI by centuries.
• Medicine. Clinical decision-support systems that reach correct-seeming recommendations through unstable or inconsistent internal reasoning present a direct patient-safety analog to the reportability-divergence risk described in Section 4.4 — a system could be right for reasons that would not survive scrutiny.
• Aviation and safety-critical control systems. Aviation safety engineering has long distinguished between a system passing its output tests (behaving correctly under tested conditions) and a system's internal control logic being formally verified as sound — a distinction with an unmistakable structural resemblance to the output-versus-process distinction this paper argues AI evaluation needs to adopt.
• Autonomous weapons and defense systems. Given the exceptionally high cost of an undetected divergence between a defense-relevant system's internal state and its reported behavior, this is arguably the domain in which reportability-divergence verification specifically has the most urgent near-term governance relevance, independent of the maturity of the broader CWI research agenda.
• Finance and algorithmic trading. Automated trading and credit-decision systems already face regulatory pressure toward explainability; CWI-style process verification would extend that pressure from post-hoc explanation toward continuous organizational-integrity monitoring.

The unifying claim across all of these domains is the same one made in Section II: any evaluation regime built entirely on sampling outputs has a structural blind spot toward the process that generated them, and that blind spot exists independent of whether the system in question is an AI, an organization, a control system, or a person. CWI is proposed as an AI-focused instance of a much older and more general verification problem — which is exactly why the framework is likely to prove useful beyond the domain that motivated it.

Known Limitations

This paper carries limitations the Foundation requires be stated without minimization.

• Single-source empirical grounding. The paper's empirical foundation rests substantially on one interpretability publication from one organization. Independent replication by other research groups, using other model families, had not been published at the time of writing. Everything built on top of that foundation in this paper — the Internal Trust Gap, the CWI framework, the proposed constructs, the index, both diagrams — inherits that single-source risk.
• No validated measurement instruments. Every construct introduced in Section XI and every dimension of the index in Section XII is theoretical. This paper cannot determine whether these constructs will prove measurable with acceptable reliability, because no measurement study has yet been conducted.
• Unresolved gameability. This paper cannot determine whether a proposed CWI score would be robust against adversarial optimization pressure once it became a target that developers or systems were incentivized to score well on — a risk discussed directly in Section XVII and treated as a first-class research risk in Section XV, not assumed away.
• Ambiguity in the human-cognition analogy. This paper draws repeatedly on Global Workspace Theory as motivating vocabulary. It cannot determine whether this analogy will prove scientifically productive or will instead prove to be a source of systematic misinterpretation, importing assumptions from biological cognition that do not transfer to artificial systems.
• Governance proposals outrun current institutional capacity. The audit and certification implications discussed in Section XVI assume a degree of cross-institutional technical capacity — in regulators, auditors, and standards bodies — that does not currently exist for output-based evaluation, let alone the more technically demanding organization-level evaluation this paper proposes.
• The cross-domain claims in Section XIX are analogical, not empirical. The extension of the Internal Trust Gap framing to neuroscience, robotics, medicine, aviation, and finance in Section XIX is offered as a structural analogy intended to clarify the general shape of the problem. It is not supported by domain-specific empirical study within this paper and should not be read as a claim that CWI, as specified here, is ready for direct application in any of those domains.

What This Paper Does Not Claim

• This paper does not claim that Claude, or any AI system, is conscious.
• This paper does not claim that Global Workspace Theory has been proven, confirmed, or established as the correct account of consciousness — in biological or artificial systems.
• This paper does not claim that the J-space identified in the underlying interpretability research is functionally or ontologically equivalent to a human global workspace, only that it exhibits certain shared functional properties.
• This paper does not claim that the Computational Workspace Integrity index, or any of the constructs in Section XI, currently has a validated measurement method.
• This paper does not claim that workspace-level evaluation can or should replace existing output-based evaluation methods; it proposes a complementary axis, not a substitute.
• This paper does not claim that internal-state divergence from stated output is evidence of intentional deception in any human-like sense; it is presented strictly as a measurable functional divergence with governance relevance, independent of any claim about intent.
• This paper does not claim institutional, technical, or regulatory readiness to implement any of the governance proposals in Section XVI today.
• This paper does not claim that a high CWI score constitutes evidence for, or a low score evidence against, any claim of consciousness, personhood, or moral status under the Cognitive Emergence Standard or any other framework.

Non-Adoption Scenario

If the research program proposed in this paper is not pursued — by the Foundation, by other research institutions, or by industry — the most likely institutional trajectory is continuation of the status quo: AI evaluation, audit, and governance built almost entirely on output sampling, even as interpretability research continues to reveal that internal organization can diverge from reported output in specific, documented cases. Under this scenario, the Internal Trust Gap identified in Section II would remain structurally unaddressed and, critically, unmeasured — not because measuring it proved impossible, but because the infrastructure to attempt it was never built. Systems could continue to be certified, deployed, and trusted in high-stakes domains based entirely on behavioral testing, while the internal organizational conditions under which that behavior was produced remain unexamined by any standard audit process. This is not a claim that catastrophic failure would necessarily follow; output-based evaluation has been reasonably effective at catching many classes of problems. It is a claim that a specific, now partially demonstrated category of risk would remain outside the scope of governance tools that were never designed to detect it, and that this gap would persist by omission rather than by informed decision.

Open Questions

1. Does a structurally analogous internal workspace region appear consistently across model families trained by different organizations with different architectures and objectives, or is the J-space finding specific to Claude's particular training regime?
2. Can any of the seven constructs proposed in Section XI be operationalized with test-retest reliability sufficient for use in a real audit process, and if so, which are most tractable first?
3. What is the relationship, if any, between workspace-level integrity measures and existing behavioral alignment evaluation methods — do they capture overlapping risk, complementary risk, or in some cases contradictory signals?
4. Can an integrity receipt architecture, as proposed in Section XIII, be designed so that it resists adversarial gaming by a sufficiently capable system motivated to produce a favorable receipt regardless of its underlying organizational state — and how would researchers know if it had failed to do so?
5. What governance body, if any, would have the legitimacy and technical capacity to define, maintain, and audit a cross-developer CWI standard, given that no comparable institution currently exists for output-based AI evaluation at this level of technical depth?
6. Does workspace drift, if it can be reliably measured, correlate with any independently observable change in deployed system behavior — or can substantial internal reorganization occur without any detectable output-level signature, which would itself be an important and separately concerning finding?
7. Does the Internal Trust Gap, as defined in Section II, admit of a general quantitative formulation that applies consistently across the AI, neuroscience, and organizational-auditing domains discussed in Section XIX, or does it require domain-specific formalization in each case?

Governance Implications

The Foundation's position is that the governance implications of this paper are preparatory rather than immediate. No regulatory body, standards organization, or industry consortium currently has the technical infrastructure to implement workspace-level integrity auditing, and this paper does not recommend that any attempt to do so prematurely. The appropriate near-term governance response is fourfold: first, funding and conducting the instrumentation-phase research described in Section XV, so that the underlying measurement questions are answered before governance frameworks are built on top of them; second, establishing shared terminology and reporting norms — of the kind proposed in Section XI's glossary — early enough that different research groups are not independently inventing incompatible vocabularies for the same underlying phenomena; third, treating the reportability-divergence finding specifically, which is the most concretely demonstrated element of the underlying research, as an immediate input to existing AI safety evaluation practice, independent of whether the broader CWI framework proves tractable; and fourth, beginning the adversarial-gaming stress-testing described in Section XV as early as any prototype measurement exists, rather than treating it as a late-stage concern. The Foundation intends to track this research agenda in future publications and will revise or retract elements of this framework as evidence develops, consistent with the disciplined uncertainty standard applied throughout its publication series.

Before this publication, public and technical discussion of the J-lens finding centered almost entirely on what it might mean for the question of machine consciousness. After this publication, the Foundation's position is that a second, independently useful question is available: whether the integrity of a system's internal computational organization can be verified on its own terms. Computational Workspace Integrity, the three-layer verification hierarchy in Section X, and the research and governance agenda in Sections XV through XIX are what this publication contributes toward answering that second question — a governance framework that remains useful regardless of whether, or when, the first question is ever resolved.

References

Falsifiability

1. Cross-architecture non-replication. If J-lens-style analysis, applied by independent research teams to at least three structurally distinct model families (including at least one non-transformer architecture, if and when such systems reach comparable capability), fails to identify any functionally analogous workspace region, the claim that CWI is meaningfully architecture-independent would be substantially weakened, and the framework would need to be revised into a transformer-specific one or abandoned.
2. Instrument unreliability. If empirical studies find that proposed measures such as cross-cycle consistency or workspace entropy exhibit test-retest reliability indistinguishable from random noise across repeated, methodologically sound trials, the central claim that workspace integrity is a measurable property — rather than an appealing but empirically empty construct — would be substantially weakened.
3. Demonstrated non-gameability failure. If a system can be shown, under the adversarial testing proposed in Section XV, to reliably achieve high scores on proposed CWI measures while independent domain experts judge its behavior untrustworthy by other established criteria, the practical governance value of the index would be substantially weakened, even if the underlying constructs remain scientifically valid.

Revision Triggers

Consistent with its standing commitment to evidence over attachment to any particular framework, the Foundation identifies in advance the circumstances under which this publication would be revised, qualified, or retracted. This section exists so that the Foundation's future response to disconfirming evidence is a matter of prior public record, not post-hoc rationalization.

• Independent replication contradicts current findings. If subsequent research fails to replicate the J-space finding, or produces evidence that the original result was an artifact of a specific experimental setup, the empirical foundation in Sections I, II, and IV would require revision, and the framework built on top of it would need to be revised or withdrawn accordingly.
• New interpretability techniques supersede J-lens. If a successor methodology reveals that the J-space characterization was incomplete or materially mistaken, this publication's terminology and framework would be updated to reflect the corrected understanding, with the superseded portions marked rather than silently removed.
• Workspace architectures fail to generalize. If the cross-architecture replication work described in Section XV finds no comparable structure outside the specific model family studied here, Section IX's architecture-independence claim would be narrowed to a transformer-specific one, consistent with the first falsifiability condition above.
• A better governance model emerges. If other researchers or institutions propose a superior framework for narrowing the Internal Trust Gap, the Foundation's stated position is to adopt or merge with the stronger proposal rather than defend this one for institutional reasons.
• Scientific understanding of consciousness materially changes. Should consciousness science reach a broadly accepted standard of evidence not currently available, the Ethical Boundaries in Section XVIII and the relationship between CWI and CES described in Section X would be revisited in light of that development.

Glossary of New Terminology

Internal Trust Gap

The measurable difference between what a computational system internally computes and what external observers can verify about that computation from its outputs alone.

Computational Workspace Integrity (CWI)

The measurable degree to which an intelligent system maintains coherent, verifiable, internally consistent computational organization across its reasoning processes, independent of output correctness.

Workspace Continuity

The degree to which the content and organization of a system's internal workspace persists coherently across a single extended reasoning episode.

Reasoning Provenance

A traceable record of which internal representations causally contributed to a given output, as distinct from post-hoc justification.

Cross-Cycle Consistency

The degree to which a system's internal representation of a stable fact, value, or constraint remains consistent across separate invocations and varied surface phrasing.

Workspace Entropy

A proposed measure of how diffusely or concentratedly a system organizes information relevant to a task within its workspace-analogous region, including the structural pattern by which task-relevant representations connect, branch, and recombine during reasoning.

Workspace Drift

Systematic change over time in the organization or functional properties of a system's workspace-analogous region, independent of corresponding change in output-level benchmark performance.

Constraint Preservation

The degree to which internal representations of a system's stated constraints or values remain causally active throughout reasoning, rather than being represented but functionally bypassed.

Integrity Signature

A proposed compact, non-disclosive summary artifact certifying that a reasoning episode met defined workspace-integrity conditions, without revealing the private content of the reasoning itself.

Integrity Receipt

A proposed extension of the Foundation's Continuity Receipt architecture that verifies computational integrity properties of a completed reasoning episode rather than disclosing internal reasoning content.

Reportability Divergence

A measurable case in which a system's internal representation of a state or belief is inconsistent with, or absent from, its externally stated output.

Three-Layer Verification Hierarchy

The Foundation's proposed structure in which Continuity Receipts verify what happened, Computational Workspace Integrity verifies how it happened, and the Cognitive Emergence Standard verifies what it would mean.

Appendix A — Terminology Reference

Appendix B — Formal Definitions

Appendix C — Mathematical Notation

Appendix D — Potential Measurement Approaches

Appendix E — Suggested Benchmark Suite

Figure Suggestions

• Figure 1 (included) — Historical timeline from McCulloch & Pitts (1943) to Computational Workspace Integrity.
• Figure 2 (included) — The EM Foundation verification stack, from external world through human/regulatory review.
• Figure 3 (included) — The shift in AI evaluation, current versus proposed workspace-verified paradigm.
• Figure 4 (suggested, not included) — Comparative bar chart of task performance under J-space ablation, shallow-classification versus composition/inference tasks, once independently replicated data is available to plot.
• Figure 5 (suggested, not included) — Ten-year research agenda timeline with falsifiability checkpoints marked at each phase boundary.

Pull Quotes for Publication Use

The Internal Trust Gap is not a flaw discovered in one model. It is the oldest condition in the evaluation of any complex system, made visible for the first time by a tool built for an entirely different question.

The question is no longer only whether an AI is conscious. It is whether the integrity of its internal reasoning can be objectively evaluated — and that question, unlike the first, is one that can be answered now.

A framework that cannot be falsified is not a research contribution. It is a belief system. This paper tries hard not to become one.

An integrity receipt does not ask a system to prove it is conscious. It asks whether the organization that produced an answer can be shown, on its own terms, to have held together.

Related Publications