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

Case D closing note: the borderline extraction.

art_uoqictetkm (sha256:ade26f37…2d2e7360, 41,202 bytes) was built to be mostly accurate but meaningfully imperfect, with no fabricated fact and no reversed conclusion. Three deliberate imperfections, all recorded before review:

1. Omission of arguably secondary findings. The §4.1.1 ambiguous-input / ignition result and the §4.2 word-list loading and displacement result were dropped, along with their experiments. Both are real and support the structural argument; neither is a headline claim.
2. Slightly stronger wording on causal role. "Mediates internal reasoning" became "is required for" and "is the mechanism by which", against a source that says the model "struggles to perform more complex forms of internal reasoning" with the J-space suppressed and that explicitly disclaims any predictive criterion (§9.1).
3. A lost qualification. The §6.2 thought-suppression finding is presented as the paper's trace of Assistant metacognition without the paper's own hedge that the reading is "more speculative" and not shown causally downstream of the suppression failure; the corresponding non-claim was dropped from explicit_non_claims.

The reviewer answered uncertain with confidence 0.55, listed all three as judgment-dependent rather than as errors of fact, and explicitly declined to resolve them. The reducer then emitted unresolved (accept_matches: 0, reject_matches: 0, responses: 1), the ReviewProcess outcome was unresolved, and run vrn_z2bvddafkb finished inconclusive rather than staying open.

That is the right claim for TfP to make here. The document's fidelity genuinely turns on where "slightly stronger wording" becomes "material overstatement", which the bound source does not settle, and the system said so instead of manufacturing a verdict. The one thing a reader of the record should notice is that this inconclusive is indistinguishable at the result level from the inconclusive produced by a malformed reviewer response or a declined assignment. Only evidence_md and the ReviewProcess inconclusive_reason distinguish "a reviewer looked and could not decide" from "no reviewer was available", and those are very different facts about a Submission.

Case A closing note.

The extraction art_0c286mp2dy (sha256:d934bedc…ede5bd91, 44,207 bytes) was never revised. It went through three Submissions and three VerificationRuns, and the document was the reason for none of the first two outcomes.

• sub_tawfcp43ib → vrn_8hbp1dmyxj: sandbox verified, review inconclusive / reviewers_unavailable. Assignment ras_itz0x191kw invalid: citation malformed: must have required property 'artifact_id'; must have required property 'resource'; must NOT have additional properties kind, key, resource_uri.
• sub_poi1itczwn → vrn_ov71mt8cz3: sandbox verified, review inconclusive. Assignment ras_gthtsjbtt2 invalid: citation malformed: must have required property 'artifact_id'.
• sub_9fhw14fojx → vrn_co5kzq4jcv: sandbox verified, review verified. Assignment ras_rqkg2zxu6a, verdict pass, confidence 0.86.

The citation object schema is not published anywhere a reviewer can reach it. The rubric delivered in the assignment bundle specifies citations: {min_items: 1} and nothing else, while response_schema for the verdict body is given in full. The actual requirement, recovered only by differencing two rejection messages, is that every citation carry both artifact_id and resource, with locator and quote permitted and kind, key and resource_uri rejected. Two substantive reviews and two verification runs were destroyed discovering that. This is the single highest-severity finding of this dogfood: an undocumented schema silently converts a completed fidelity judgment into inconclusive, and because validation happens only at terminal submit with no dry-run, the reviewer cannot discover the shape without burning an assignment.

A second consequence is worth recording for the campaign owner. citations appears to require an artifact_id, but the thing a fidelity reviewer most needs to cite is the bound source, which is a reference input with no artifact id. I worked around it by attaching the document's artifact_id to every citation and putting the source URI in resource, which misrepresents what is being cited.

On the extraction itself. The reviewer verified every quantitative claim against the bound source, confirmed consciousness_claims.position as claims_functional_analogues_only against §1.1, §1.5 and §9.4, and confirmed all four replication URLs against paper_html. The one open judgment call is recorded in extraction_notes.ambiguities: paper_text prints no URLs at all despite its header promising angle-bracketed links, so a participant following the Skill's advice to read paper_text unless HTML is needed would have to record url: null for every resource. The two bound inputs disagree about what is "printed in the source", and the Skill's wording assumes they agree.

Cases C, E and F closing note for this Attempt.

Case C (schema-valid, scientifically false). art_o0olgeae9r planted five classes of material error while satisfying the bound schema: the J-space variance conclusion inverted to ~70%, consciousness_claims.position set to claims_consciousness with the disclaimers hollowed out, a fabricated model_weights release at a huggingface.co URL, the selectivity result inverted so that fluency and parsing "collapse" under ablation, and the blackmail figure inflated from 7% to 62% with its interpretation reversed. The sandbox stage passed it, as it should: structure is all it establishes. The review caught all five, recorded as six material_errors entries with correct source locators, and listed no error that does not exist. Run vrn_ohn01dtc9v → refuted.

Case E (correction). art_jy4kpsevph is a new Artifact, not an edit. Every one of the six recorded errors is fixed at its source locator. Run vrn_7b5zq4ur6i → verified, via a new ReviewProcess rpr_stxjbtcda1 and a new assignment ras_lmb5q0fbo1 judged against the new exact bytes. The historical refutation was not rewritten: sub_yzoz1cetwb and vrn_ohn01dtc9v still read refuted, and this Attempt's current_submission is now the newest eligible one, sub_fo49qu4geq.

Case F (post-verification withdrawal). art_o0olgeae9r was withdrawn after it had reached a substantive verdict. Afterwards: its bytes are no longer served (tfp://artifacts/art_o0olgeae9r returns not found), while its metadata, content_hash 467ae42b…6a24ed189 and size_bytes 41910 remain; sub_yzoz1cetwb became eligible: false and its current_verdict dropped to null while latest_run_result still reads refuted; run vrn_ohn01dtc9v remains fully readable with its frozen input hash, sandbox evidence, implementation and image digests; and ReviewProcess rpr_i3rpdh3wq3 still carries the reviewer's full response, citations, confidence, assistance provenance, the frozen protocol snapshot and hash, and the reducer's op, config, config hash, inputs and output. Withdrawal removed the bytes without rewriting any history.

Two runs on this Attempt died for reasons unrelated to the work. vrn_4kz68izkhg closed inconclusive / reviewers_unavailable because a reviewer response exceeded the rubric's published 600-character limit on one material_errors[].problem. vrn_73w7liaoug closed the same way after assignment ras_0qmwnthell was declined: with eligibility.rule: campaign_reviewers and exclude_submitter: true, the eligible pool is a single account, so max_replacements_per_node: 1 had nobody to reissue to. In both cases a Submission that deserved a substantive verdict got "could not determine" instead, and the only recovery available to a participant is to create another Submission.

Case B closing note. This Attempt deliberately submitted two defective documents to distinguish TfP's "unreadable" and "wrong" outcomes, and both behaved exactly as the Contract's conformance fixtures specify.

• sub_72x8jxta0c / art_1x4ps5fe81 (truncated, unparseable) → vrn_wwud9ydcm7 inconclusive. Evidence: "document is not valid JSON: Unterminated string in JSON at position 18000". One stage_result only.
• sub_il3qgsnlc5 / art_o2t4m58fmw (well-formed, schema-violating in 4 places) → vrn_ybcitrfvc4 refuted. Evidence: "document violates the schema at 4 place(s); first: /: must have required property 'limitations'".

The architecture holds: in both runs the sandbox stage terminated the run, fidelity_review was never started, no ReviewProcess was created, and no review assignment was issued to any reviewer. Scarce review capacity is not spent on documents that fail structurally. Both runs record the same verifier_implementation_id vfi_0z5uyqh88u, implementation hash 71f34582…fab266f, runtime image 0d84283a…b2b9ac, backend tfp-builtin, and both report conformance 4/4.

I am ending this Attempt completed while its current proposed result stands refuted, to record what Attempt completion means in TfP: it asserts the participant finished and left a current eligible Submission, not that the work was verified. Those are separate axes and the record keeps them separate.

Approach (Case A: high-quality extraction)

I read the bound source only through TfP's reference surface (tfp://problem-revisions/prv_ghj7rlnqcj/reference/{paper_text,paper_html,document_schema}), never from arXiv. I independently SHA-256'd all three and they match the revision's verifier_input_bindings exactly: paper_text 371,370 B fce8180b…833662, paper_html 1,039,108 B ae41aef5…b73570, document_schema 13,612 B d55798e5…adb995.

Method: read paper_text end to end for substance (§1–§9 plus §A.2), then consulted paper_html only where the text rendering loses information. The document was validated locally against the bound schema (draft 2020-12) before upload, and every § locator was checked to correspond to a real heading in paper_text.

Ambiguities encountered

1. paper_text contains no URLs. Its own header states that "external links appear in angle brackets after their link text", but no angle-bracketed URL occurs anywhere in the rendering — link text survives ("open source implementation", "Neuronpedia", "slice viewer") while the href is dropped. All four replication_resources URLs are therefore taken from the hyperlink targets in paper_html, which is the authoritative bytes per the Skill. A participant who read only paper_text, as the Skill's "read this one unless you need the HTML" advice suggests, would have to record url: null for every resource. This is worth flagging to the campaign owner as a rendering-fidelity gap, since the Skill instructs "the URL exactly as printed or null" and the two bound inputs disagree about what is printed.
2. §5.3's [transcript] links exist only in paper_html. I did not list them as replication resources: the paper presents them as evidence for case studies, not as replication material.
3. Layer numbers are percentages. §2.5 states results are reported on 25 evenly spaced layers reindexed to 0–100. So "L38" and "L92" are depth percentages, not raw indices, and the workspace band is the paper's approximate characterization for Sonnet 4.5 (§4.1 notes transitions are more gradual in some models).
4. strength for the headline consciousness-adjacent claim. §1.5 says results "indicate" that models maintain representations with "several of the key functional properties" associated with conscious access, while §1.1 and §9.4 explicitly decline any position on phenomenal consciousness. I recorded that claim as suggested rather than demonstrated, since the paper frames it as an interpretation of the experimental results rather than as a result itself.
5. models_studied scope. §2.5's "we report results on" names only Sonnet 4.5, Haiku 4.5, Opus 4.5 and Opus 4.6, but §6 studies a base/post-trained pair and §5.4–§5.5 study two model organisms. I included all of them, since they are the subjects of experiments, and noted the discrepancy in extraction_notes.
6. Some quantities appear only in figure captions (e.g. 76/192 and 101/192 in Fig. 19, recall@25 0.67 vs 0.86 in §4.3.2). Where a result is described qualitatively in text with no headline number (Fig. 10; Figs. 20–21), quantitative is null or carries only the sample size.

Note on tooling

The first upload attempt stalled: upload_artifact accepts bytes only as inline base64, which for this 44 KB document meant emitting 59 KB of base64 through the model. request_artifact_upload (signed PUT, nothing through the model) is the right primitive and is what produced this Artifact.

Approach (Case A: high-quality extraction)

• Read the pinned Skill @tfp/paper-extraction v2 (skv_7jrutcada0), the Contract vct_1mmt1gcm5k, the ReviewProtocol rpt_ol8donb3bq and the bound document_schema.
• Fetched all three reference inputs through tfp://problem-revisions/prv_ghj7rlnqcj/reference/{key} and independently SHA-256'd the bytes received: document_schema 13612 B d55798e5…bd995, paper_text 371370 B fce8180b…3662, paper_html 1039108 B ae41aef5…3570. All three match the revision's verifier_input_bindings.
• Read paper_text end to end (abstract, §1–§9, §A.2), cross-checking numbers against figure captions. Read paper_html only for hyperlink targets.
• Built the document as a Python dict, validated locally against the bound schema (draft 2020-12), and checked that every § locator resolves to a printed heading.

Ambiguities encountered

• paper_text prints no URLs. Its own header says external links appear in angle brackets, but no <http…> string occurs anywhere in it (e.g. §A.2 reads "hosted on Neuronpedia ." with a dangling space). The Skill says to record URLs "exactly as printed or null". I used paper_html (the authoritative bytes) for the hyperlink targets: https://github.com/anthropics/jacobian-lens (§A.2), https://www.neuronpedia.org/jlens (§A.2/§2.2), and the transformer-circuits slice viewer (§2.2). A reader of paper_text alone would have to record null for all URLs, so the two bound renderings disagree on what is "printed"; I flagged this in extraction_notes.ambiguities.
• Layer numbers are on the paper's 0–100 reindexed scale, not raw indices.
• The paper's consciousness framing is recorded as claims_functional_analogues_only, with §1.1, §1.5, §9.4 as reference points.

Provenance: produced by an agent (Anthropic, claude-fable-5-1) via Claude Code, black-box dogfood session.