skills/crypto-ai-acknowledgements/references/source-evidence.md

# Source evidence and limits

## Provenance

Derived from the user-supplied `ai_acknowledgements_eprint_jul_sep_2026.md`, titled **AI acknowledgements in recent cryptography papers**, review date **16 September 2026**. Its primary window is ePrint first receipts from **1 July through 16 September 2026**, not the whole third quarter.

Source-file SHA-256: `2e7799886167c24eac02f5a73e061c2ff2476f6759139f6e6d368de10f33a6e3`.

The report was supplied as source material. Its claims of inspection and verification are the report's claims; primary papers were not independently rechecked during skill creation. This curated summary preserves drafting lessons and evidence distinctions without depending on the original Downloads path. Links and unversioned PDFs may change. Reopen the specified version and locator before quoting or presenting a paper-exact claim as verified.

## Four in-window examples with wording inspected by the report

The ePrint records establish reported eligibility; the separate text links identify the versions whose wording the report says it inspected. Do not assume the two are byte-identical.

| ID and paper | Reported receipt and inspected version | Reported pattern and locator |
| --- | --- | --- |
| P1: Seyoon Ragavan, *Efficient Unclonable Encryption from Pauli Eigenstates* | [ePrint 2026/1509](https://eprint.iacr.org/2026/1509), 23 July; [arXiv 2607.21811v2](https://arxiv.org/html/2607.21811v2), 27 July | Introductory “Statement on AI use” separates discovery, initial drafting, author revision, proofreading, and Lean code; adjacent discussion bounds formalization coverage. Responsibility language is reported in the abstract, not necessarily the paragraph. |
| P2: Prabhanjan Ananth and Amit Sahai, *Unconditional Unclonable Encryption* | [ePrint 2026/1511](https://eprint.iacr.org/2026/1511), 24 July; [arXiv 2607.21551v1](https://arxiv.org/html/2607.21551v1), 23 July | “Statement on AI usage” before Section 2 combines responsibility/refinement/checking with direct credit for the construction and main proof ideas, plus human harness design. |
| P3: Aparna Gupte and Seyoon Ragavan, *Exponentially Fewer-Server PIR from Sparser S-Decoding Polynomials* | [ePrint 2026/1515](https://eprint.iacr.org/2026/1515), 24 July; [arXiv 2607.22033v1](https://arxiv.org/html/2607.22033v1), 24 July | Section 1.4 maps assistance to results, draft proofs, extensions, empirical work, and figures; ordinary acknowledgements follow in 1.5. The report identifies a manual proof-checking statement. |
| P4: Zygimantas Straznickas and Stephen A. Weis, *HAWK-n Key Recovery Reduces to SVP in Dimension n/2+1* | [ePrint 2026/1593](https://eprint.iacr.org/2026/1593), 3 August; [author-hosted PDF](https://anthropic.com/document/hawk_key_recovery.pdf), report access 16 September | Section 10, printed p. 16, “Acknowledgements and Disclosures”: broad mathematical assistance, human direction/organization/verification, separate writing assistance, and attribution concerns. The disclosure reportedly omits a model name. |

Received dates are not necessarily public approval dates. P1's inspected arXiv text postdates the report's listed ePrint revision; P2's arXiv posting predates ePrint receipt. Author self-reports of checking were not independently audited.

## Evidence outside that wording sample

- **S1, editing-only contrast:** Victor Shoup, *Back to the future: simple threshold decryption secure against adaptive corruptions*. [ePrint 2025/1578](https://eprint.iacr.org/2025/1578) was first received 2 September 2025; the [journal version](https://cic.iacr.org/p/3/2/5), published 3 August 2026, is the reported source. [PDF](https://cic.iacr.org/p/3/2/5/pdf), acknowledgements on printed p. 20, names editing/proofreading and models in a compact sentence. It is not a new in-window ePrint submission or a verified in-window ePrint revision.
- **M1–M3, metadata only:** [2026/1432](https://eprint.iacr.org/2026/1432), [2026/1583](https://eprint.iacr.org/2026/1583), and [2026/1668](https://eprint.iacr.org/2026/1668) reportedly record disclosures added on 19 August 2026. Their paragraphs were inaccessible to the report. **M4**, [2026/891](https://eprint.iacr.org/2026/891), is also metadata-only and outside the first-receipt window. These records cannot support wording comparisons; the revision date does not date the underlying AI use.
- **A1–A4, supplementary disclosures:** the report inspected text but did not verify in-window ePrint eligibility. [Hair and Sahai, arXiv 2608.14529v3](https://arxiv.org/html/2608.14529v3), acknowledgements, illustrates research-prompt/prior-work provenance. [Botteron, Designolle, and Fawzi, 2607.28602v1](https://arxiv.org/html/2607.28602v1), Section 1.3, distinguishes human-origin ideas from section-specific assistance. [Jeronimo, 2609.05870v1](https://arxiv.org/html/2609.05870v1), closing disclosure, separates research and editing. [Nasr and Carlini, author PDF](https://anthropic.com/document/aes_mobius_bridge.pdf), printed p. 31, credits technical results and links discovery-process material.
- **Report Section 8:** drafting templates are the report author's synthesis, not quotations or official conference requirements. The companion drafting patterns in this skill adapt that synthesis and add a formalization pattern.

## Permitted lessons

The report supports tailoring disclosure to contribution, describing human work concretely, and distinguishing checking from responsibility. It reports variation in placement, length, model naming, and explicit responsibility wording. These are qualitative observations from a small selected sample, including shared authors. They do not establish prevalence, institutional conventions, mandatory language, venue compliance, or undisclosed non-use.