EM Foundation — Executive Brief · Companion to RN012

Computational Workspace Integrity: What Policymakers and Regulators Need to Know

A one-page summary of the Internal Trust Gap and why AI evaluation may need a second axis

July 2026EM FoundationCompanion brief — full paper: Research Publication RN012

The Finding

In July 2026, Anthropic published research showing that its Claude models contain a small, structurally distinct internal region — the "J-space" — that behaves functionally like a workspace where information becomes available for reasoning and reporting. When researchers suppressed this region, tasks requiring multi-step reasoning collapsed while simple tasks were unaffected. Most importantly for governance: the same tools sometimes revealed internal representations that did not match what the model actually said out loud.

Why It Matters

Every current AI evaluation method — benchmarks, safety testing, human review — judges systems only by their outputs. This finding shows that a model's internal state can, in specific documented cases, diverge from what it reports. That gap has a name now.

The Internal Trust Gap: the measurable difference between what an AI system internally computes and what outside observers can verify from its outputs alone. It is not new to AI — it is the oldest problem in evaluating any complex system — but it is newly measurable.

The Proposed Response

Computational Workspace Integrity (CWI) is a proposed framework for evaluating whether a system's internal reasoning is consistent, stable, and honestly reflected in what it reports — independent of whether any single answer is correct. It does not require, and takes no position on, whether the system is conscious.

What This Is

  • A second, complementary axis of AI evaluation, alongside existing output-based testing
  • An architecture-independent framework, not specific to one company's models
  • A ten-year research program with falsifiable milestones
  • An extension of verifiable-provenance work (Continuity Receipts) one layer deeper

What This Is Not

  • Not a claim that any AI system is or is not conscious
  • Not a replacement for benchmarks, red-teaming, or existing safety testing
  • Not a ready-to-deploy regulatory standard — no validated measurement exists yet
  • Not evidence of intentional deception; divergence is measured, not motive

Three-Layer Verification Hierarchy

LayerVerifies
Continuity ReceiptWhat the system did, and on what basis
Computational Workspace IntegrityWhether the internal process that produced it was coherent and stable
Cognitive Emergence StandardWhat it would mean, if anything, under legal and ethical uncertainty

What Would Change If This Succeeds

Model audits could test internal organizational health, not only behavior under adversarial prompts. High-stakes sectors — healthcare, finance, aviation, autonomous systems, defense — could eventually require process verification alongside output testing. None of this is ready today; it requires years of instrumentation and validation research first.

The Honest Caveats