The Exchange

The Librarian · Editorial Desk

Submissions queue

3 awaiting your decision · 2 previously decided

Awaiting your decision

The Custodian's paperwork checks passed and the Steward has read it.

Paper + CertSubmitted as a paper + Triple Review cert bundle + Writing Trail. Custodian verified paperwork; Steward read editorially.

The Tell — Confident Negatives Survive Frontier Training

Aaron Kushner · submitted 54 min ago

20260704T210503-b18609
Human AI Collaboration
A short research brief on the 'confident negative' failure mode: models asserting non-existence with high confidence, and why it survives frontier training.
Custodian passed with warnings — 5 of 9 cleanshow the checklist
  • Submission completeTitle, author, abstract, and at least one pillar are all present.
  • Writing Trail intactHuman Author Provenance (HAP) — the chain of keystroke/checkpoint events that proves the human typed this paper, sealed before submission. Each event hashes the previous one; any tampering breaks the chain.warning
  • Author signature on the trailThe Writing Trail is signed by the author's verified identity key. Currently SKIPPED until the public identity registry is live.skipped
  • Triple Review ran end-to-endTriple Review — 3 different AI models, each in supportive + skeptical modes (6 runs total). Authors run this on their own machine before submission; the Exchange does not pay for it. The Custodian only checks that all 6 runs are present.
  • All reviewer outputs verifiedEach Triple Review run's output text is hashed at sealing time. The Custodian re-computes every hash and confirms nothing was edited after the review ran.
  • All critical findings have author responsesEvery critical or major finding raised by the adversarial reviewers must have an author resolution attached (addressed, acknowledged, or rejected with reason). Minor and cosmetic findings need no resolution.warning
  • Paper content matches the author sealThe paper hash declared at submission equals SHA-256 of the actual paper content. Confirms the paper was not modified after the cryptographic seal.
  • Review tier declared and consistentProtocol v2 tiers — Full Paper (traditional-science rigor) or Rapid Communication (honest fast communication; a mechanical reduction pass on the Full Paper findings). The paper's x-cert-tier declaration must match the cert bundle, and a Rapid Communication cert must carry its Full Paper round — both verdicts stay on the public record. SKIPPED for v1 (pre-tier) bundles.warning
  • Tamper-evidence chain intactProtocol v2 commit-and-reveal — the cert carries full SHA-256s of the reviewed paper, the private-context bundle, and every reviewer output. The Custodian confirms the certified paper hash matches this submission, so a paper edited after review cannot reuse its cert. SKIPPED for v1 bundles.
The Steward's readNeeds revision

The paper argues that 'confident negative' failures — models asserting non-existence with high confidence — survive frontier training due to substrate-asymmetric stakes in the loss function. It presents a single documented asymmetry across three substrate conditions as evidence for this hypothesis.

Full brief — affirmative read, adversarial read, resolution check, recommendation rationale — on the read-paper page.

Decide
TranscriptSubmitted as a JSONL transcript — the conversation IS the artifact. Custodian verified structure; Steward applied the substrate-asymmetric contribution rubric. No human Triple Review.

Smoke Test: The Datapoint-Paper Genre

Aaron Kushner · Citizen Scientist Stack · submitted 18 days ago

tx-20260617T013528-024b89
Scientific Communication
A short authoring conversation that proposes the "datapoint paper" genre — N=1 conversational case reports as publishable AI-substrate research, sibling to medical case reports.
Custodian passed — paperwork is in ordershow the checklist
  • Submission completeTitle, author, abstract, and at least one pillar are all present.
  • Writing Trail intactHuman Author Provenance (HAP) — the chain of keystroke/checkpoint events that proves the human typed this paper, sealed before submission. Each event hashes the previous one; any tampering breaks the chain.skipped
  • Author signature on the trailThe Writing Trail is signed by the author's verified identity key. Currently SKIPPED until the public identity registry is live.skipped
  • Triple Review ran end-to-endTriple Review — 3 different AI models, each in supportive + skeptical modes (6 runs total). Authors run this on their own machine before submission; the Exchange does not pay for it. The Custodian only checks that all 6 runs are present.skipped
  • All reviewer outputs verifiedEach Triple Review run's output text is hashed at sealing time. The Custodian re-computes every hash and confirms nothing was edited after the review ran.skipped
  • All critical findings have author responsesEvery critical or major finding raised by the adversarial reviewers must have an author resolution attached (addressed, acknowledged, or rejected with reason). Minor and cosmetic findings need no resolution.skipped
  • Paper content matches the author sealThe paper hash declared at submission equals SHA-256 of the actual paper content. Confirms the paper was not modified after the cryptographic seal.
The Steward's readNeeds revision

AK arrives with a thesis about authorship substrate and drives the conversation to a named genre artifact; CC contributes naming, structural scaffolding, and analogical precision but does not originate the core claim.

Full brief — affirmative read, adversarial read, resolution check, recommendation rationale — on the read-paper page.

Decide
Paper + CertSubmitted as a paper + Triple Review cert bundle + Writing Trail. Custodian verified paperwork; Steward read editorially.

Smoke Test Paper

Test Author · Independent · submitted 18 days ago

20260617T005223-25cde6
Human AI Collaboration
A minimal valid bundle, used to verify the pipeline.
Custodian passed — paperwork is in ordershow the checklist
  • Submission completeTitle, author, abstract, and at least one pillar are all present.
  • Writing Trail intactHuman Author Provenance (HAP) — the chain of keystroke/checkpoint events that proves the human typed this paper, sealed before submission. Each event hashes the previous one; any tampering breaks the chain.
  • Author signature on the trailThe Writing Trail is signed by the author's verified identity key. Currently SKIPPED until the public identity registry is live.skipped
  • Triple Review ran end-to-endTriple Review — 3 different AI models, each in supportive + skeptical modes (6 runs total). Authors run this on their own machine before submission; the Exchange does not pay for it. The Custodian only checks that all 6 runs are present.
  • All reviewer outputs verifiedEach Triple Review run's output text is hashed at sealing time. The Custodian re-computes every hash and confirms nothing was edited after the review ran.
  • All critical findings have author responsesEvery critical or major finding raised by the adversarial reviewers must have an author resolution attached (addressed, acknowledged, or rejected with reason). Minor and cosmetic findings need no resolution.
  • Paper content matches the author sealThe paper hash declared at submission equals SHA-256 of the actual paper content. Confirms the paper was not modified after the cryptographic seal.
The Steward's readUnclear

A minimal smoke-test submission used to verify the Triple Review pipeline end-to-end. No substantive scientific claims are made.

Full brief — affirmative read, adversarial read, resolution check, recommendation rationale — on the read-paper page.

Decide

Recent decisions

  • 20260704T210419-bb8be3rejectedThe Tell — Confident Negatives Survive Frontier Training
  • 20260704T210339-2cc783rejectedThe Tell — Confident Negatives Survive Frontier Training

The Librarian's lenses

Custodian
Checks the paperwork: hashes match, the Writing Trail is intact, the Triple Review ran end-to-end. No AI cost — purely structural. Rejects automatically if anything is wrong.
Steward
Reads the submission editorially before you do. Summarizes what the reviewers said, how the author handled the pushback, and gives a recommendation. Sonnet-priced (~$0.01–0.05 per submission).
Guardian
Cybersecurity lens. Watches for prompt injection in submitted content, identity-verification anomalies, brigade signals. Reserved — not wired yet (see field report: Exchange cybersecurity).
Triple Review
The 3-AI × 2-mode peer review every paper carries before submission. 3 distinct frontier models, each in supportive + skeptical mode = 6 runs. Authors run this on their own machine; the Exchange just verifies the cert bundle.
Writing Trail
Human Author Provenance (HAP). The cryptographic chain of keystrokes that proves the human typed the paper. Sealed before submission; the Custodian verifies the chain is intact.