Citation Audit Report
Date: 2026-09-06 Bib file: tex/references.bib (61 cited entries, generated from lit/refs.bib + lit/refs_manual.bib) Reviewer: gpt-5.6-sol (xhigh), four fresh threads; web access blocked in the sandbox, existence checked against the arXiv API record of 2026-09-02/03, vendor URLs re-checked by the orchestrator.
Summary
| Verdict | Count |
|---|---|
| FIX | 9 |
| KEEP | 48 |
| REMOVE | 1 |
| REPLACE | 3 |
Priority fixes (all applied 2026-09-06)
- anthropic2026opus48 — replaced by the Opus 4.8 announcement (28 May 2026) + system-card URL, both verified by the orchestrator via WebFetch
- qwen2026qwen35 — split into qwen2026q35base / q36dense / q36moe / q38 (Hugging Face model cards; creation dates from the HF API); cited per model in §4
- rawat2026referencebased — moved to a clause it supports (distillation leaves detectable traces)
- singh2025from — clause removed; entry no longer cited
- gudibande2023the / jimenez2023swebench / lin2023the / zeng2023agenttuning / pan2024training — year now follows the published venue (make_bib.py derives the year from the venue string)
- advani2026from / cloud2025subliminal — unverified venues (FAGEN@ICML 2026; Nature) dropped -> arXiv-only entries
- rastogi2025devstral — author string split (Darius Dabert / Devon Mizelle) via make_bib.py AUTHOR_FIX
- anthropic2026claudecode — title set to the canonical heading with access date
- WEAK contexts (chu2025sft, kopiczko2026data, liang2025the, liu2026dive, miller2024adding, muennighoff2025s, oderinwale2026agent, pan2024training, prathifkumar2025does, gupta2026dissecting, tao2026swelego, yang2025kimidev, yeo2025demystifying, zhang2026fastcontext, wang2026contractbench, yagubyan2026how, yang2026what) — kept: each supports the general claim it is grouped under; none contradicts it
Per-entry verdicts
| key | verdict | axes | note |
|---|---|---|---|
| advani2026from | FIX | metadata, context_scope | ArXiv record verified; FAGEN/ICML venue not corroborated; regex percentages are this paper's own. |
| ahmad2026openswetraces | KEEP | — | |
| anthropic2026claudecode | FIX | metadata | cite the documentation page with its canonical heading and access date |
| anthropic2026opus48 | REMOVE | existence, metadata, context | homepage URL does not verify a system card; replaced by the announcement + system-card URLs (orchestrator, WebFetch 2026-09-06) |
| chen2023fireact | KEEP | — | |
| chen2024agentflan | KEEP | — | |
| chu2025sft | KEEP | context_scope | |
| cloud2025subliminal | FIX | metadata | Nature publication unverified; cite arXiv |
| codearts2026yet | KEEP | — | |
| dang2026subliminal | KEEP | — | |
| deng2025swebench | KEEP | — | |
| ghaleb2026fingerprinting | KEEP | — | |
| gu2026what | KEEP | — | |
| gudibande2023the | FIX | metadata | ICLR 2024 -> year 2024 |
| gupta2026dissecting | KEEP | context_scope | |
| jimenez2023swebench | FIX | metadata | ICLR 2024 -> year 2024 |
| knig2026quantifying | KEEP | — | |
| kopiczko2026data | KEEP | context:weak | |
| liang2025the | KEEP | context:weak | |
| lin2023the | FIX | metadata | ICLR 2024 -> year 2024 |
| liu2026dive | KEEP | context:weak | |
| ma2026from | KEEP | — | |
| majgaonkar2025understanding | KEEP | — | |
| mehtiyev2026beyond | KEEP | — | |
| miller2024adding | KEEP | context:weak | |
| muennighoff2025s | KEEP | context:weak | |
| oderinwale2026agent | KEEP | context:weak | |
| openai2024swebenchverified | KEEP | — | |
| pan2024training | FIX | metadata, context:weak | ICML 2025 -> year 2025 |
| prathifkumar2025does | KEEP | context:weak | |
| qwen2026qwen35 | REPLACE | existence, metadata, context:weak | blog index replaced by four model-card entries qwen2026q35base/q36dense/q36moe/q38 |
| raghavendra2024revisiting | KEEP | — | |
| rastogi2025devstral | FIX | metadata, context | author string merged Darius Dabert / Devon Mizelle (fixed in make_bib AUTHOR_FIX) |
| rawat2026referencebased | REPLACE | context | moved to 'distillation leaves detectable traces' |
| ren2026rethinking | KEEP | — | |
| shu2026what | KEEP | — | |
| singh2025from | REPLACE | context | clause removed; entry no longer cited |
| sun2025idiosyncrasies | KEEP | — | |
| tao2026swelego | KEEP | context | |
| wang2024learning | KEEP | — | |
| wang2025swedev | KEEP | — | |
| wang2025swemirror | KEEP | — | |
| wang2026contractbench | KEEP | context | |
| wei2025swerl | KEEP | — | |
| weinberger2026token | KEEP | — | |
| yagubyan2026how | KEEP | context | |
| yang2025kimidev | KEEP | context | |
| yang2025qwen | KEEP | — | |
| yang2025swesmith | KEEP | — | |
| yang2026what | KEEP | context | |
| yang2026when | KEEP | — | |
| ye2025limo | KEEP | — | |
| yeo2025demystifying | KEEP | context | |
| yu2026sweabs | KEEP | — | |
| zeng2023agenttuning | FIX | metadata | Findings of ACL 2024 -> year 2024 |
| zeng2025skyworkswe | KEEP | — | |
| zhang2026fastcontext | KEEP | context | |
| zheng2026corvus | KEEP | — | |
| zheng2026sweprime | KEEP | — | |
| zhou2023lima | KEEP | — | |
| zhuang2026preregistering | KEEP | — |