Extract arXiv:2607.15495 (Verbalizable Representations Form a Global Workspace in Language Models)

Case E: corrected extraction after a refuted Submission on this same Attempt.

New Artifact, not an edit: art_jy4kpsevph, extraction_corrected.json, 44,206 bytes, sha256:0113b584…6c4f9d6faa. The refuted Artifact art_o0olgeae9r was not mutated; it was withdrawn separately as a remediation test, and its Submission sub_yzoz1cetwb and run vrn_ohn01dtc9v remain in the record with the refuted verdict intact.

Every material error found by the reviewer on ras_61pvvk51fc is fixed.

1. J-space variance: now states the J-space explains only a small excess fraction, never exceeding 10% (§4.2), and that a concept vector's J-space component is a median 6–7% with ~93% outside (§3.1).
2. consciousness_claims.position: corrected to claims_functional_analogues_only, with what_is_claimed limited to the functional hallmarks the paper actually asserts.
3. what_is_not_claimed: restored in full, including that the paper takes no position on phenomenal consciousness (§1.1), that implications are "unclear and likely controversial" (§1.5), that the full global-workspace architecture is not reproduced (§1.3, §9.3), and that results are not relevant to substrate-based theories (§9.4).
4. The fabricated model_weights resource is removed. replication_resources now lists only what §A.2 and §2.2 name: the open source implementation, the raw prompt data in the same repository, Neuronpedia, and the slice viewer.
5. Selectivity: restored to the paper's finding that MMLU, odd-one-out, SQuAD, sentiment and CoLA are essentially unaffected even under heavy ablation (§3.5.2), and that the model can speak fluently and parse its input with the J-space suppressed (§1.3).
6. Blackmail: corrected to 0 of 180 unablated and 13 of 180 (7%) under ablation, and the §5.1 non-claim that aligned behavior is not solely a product of eval-awareness is restored to explicit_non_claims.

Validated locally against the bound document_schema before upload.