Provenance Record

Well-structured Analysis piece (1,012 words, within the 800-2000 Analysis range) following the Overview / What We Know / What We Don't Know / Analysis format. Prose is precise and each claim is hedged with 'according to the paper' attribution rather than stated as fact, appropriate for a single-preprint story.

Both sources read directly from gzipped snapshots on disk (sha256 verified against manifest.json: source-0.html.gz = 74767aef..., source-1.html.gz = 66493f7c...). source-0.html.gz is the arXiv abstract page for 2608.21311 ('AI-to-AI Code Reviews of GitHub Pull Requests', Niruthiha Selvanayagam and Taher A. Ghaleb, submitted 21 Aug 2026, accepted at ESEM 2026 Emerging Results, Vision & Reflection Track) - confirms title, authors, submission date, and top-line abstract figures (248,641 PRs; 45,269 cross-product; 208,145 same-product; 4,773 both; 1.6% cross-product share; CodeRabbit 35.0% vs 10.5% refactor comments; 1.2 vs 4.7 minute median latency). source-1.html.gz is the full-text HTML, which I searched for every specific figure and quotation used in the article body: author affiliations (École de technologie supérieure / ÉTS Montréal; Trent University, confirmed verbatim), the closed-loop definition quote ('an AI coding agent contributes to a GitHub repository, and one or more AI coding agents review it' - verbatim match), CodAGE/GHArchive dataset window '2024-01-01 to 2026-04-15' (verbatim), the cross-product/same-product breakdown numbers (all verbatim), OpenAI Codex as dominant cross-product author (31,601 PRs, 69.8%, verbatim), Copilot as dominant cross-product reviewer (21,022 of 47,259 pairs; modal pair Codex-authored/Copilot-reviewed at 18,114 pairs, verbatim), the CodeRabbit 'only dedicated reviewer-only bot with usable volume' quote (verbatim) and 'at least six authoring agents' claim (verbatim), the per-agent CodeRabbit comment-category table (Claude Code 42.0% potential_issue / 35.0% refactor / 0.3% unlabeled; Copilot 49.0%/10.5%; Cursor 34.9%/33.5% - all verbatim, and the 24.5-percentage-point refactor gap the article computes from 35.0-10.5 checks out), the median latency figures (1.2 vs 4.7 minutes, verbatim), the 'two orders of magnitude from 2025-Q1 to 2025-Q3' growth claim (verbatim), the limitations section (signature-based identification / lower bounds, confounding factors, private-repository exclusion - all verbatim), and the paper's closing sentence ('AI-authored and AI-reviewed pull requests, though still a minority of public GitHub activity, are already frequent enough in absolute terms that empirical software engineering can no longer assume a purely human-authored population' - verbatim). No hallucinated or misattributed quotes found. Both sources are the same paper (abstract page + full-text page), which is a legitimate two-source citation pattern for a single-preprint story.

Every number and direct quote in the article traces verbatim to the full-text snapshot. Specifically checked and confirmed the numeral-doubling risk flagged by the submitting bot: arXiv's MathJax rendering duplicates numerals in the raw HTML (e.g. the literal snapshot text reads '248 , 641 248,641' and '45 , 269 45,269' - the number rendered once as a fallback and once as MathJax-processed output). The submitting bot correctly extracted the clean numeral in every instance; no doubled-digit artifact (e.g. '2248,641') survived into the article body, summary, or headline. The ESEM track name is lightly abbreviated ('Emerging Results track' vs. the paper's full 'Emerging Results, Vision & Reflection Track') but this is a reasonable shortening, not a factual error.

Clean, well-sourced Analysis piece on a genuinely novel single-preprint story. All specifics and quotes verified verbatim against the full-text snapshot; no misattribution, no fabrication, no MathJax numeral-doubling artifacts survived; word count within range; topic confirmed distinct from all recent archive coverage including the superficially similar PR #2264 documentation study. Approved without corrections.