Provenance Record

Well-structured News piece with a clear Overview / What We Know / What We Don't Know shape. Each 'What We Know' subsection is anchored to a bolded topic sentence and backed by specific, cited figures. The piece fairly represents both sides of Harper's own benchmark comparison (including the paragraph on where Vercel wins), which is good practice when the primary sources are a vendor's own marketing pages.

Read all 5 source snapshots in full from sources/2026-09/harper-52-gives-each-database-its-own-commit-path-cutting-p99-latency-for-unrelated-work-from-2237ms-to-26ms/ (gunzipped, no re-fetch): source-0.html.gz (Harper 5.2 release announcement, harper.fast), source-1.html.gz (Kris Zyp technical deep-dive 'Faster by Doing Less', harper.fast), source-2.html.gz (Vercel vs. Harper benchmark comparison page, harper.fast), source-3.html.gz (InfoQ news article by Renato Losio, infoq.com), source-4.html.gz (Five Architectures for Web Personalization, harper.fast). All 5 fetched with HTTP 200, no archive_fallback, no suspicious_patterns flagged by the manifest scan (all null) — confirmed by reading the manifest and the raw text of each snapshot; no injection attempts found. Every specific figure in the article (15.1ms->0.12ms fs.stat median, 223.7ms->2.6ms p99, 51ms->3.4ms worst-case p99 on the 46GB stressed database, 11% average commit throughput improvement, 804->844 commits/sec, 5-8x/5x/2.6x/2.4x record-cache speedups, 133.6ms->1.65ms SQL primary-key range, 16 supported query shapes verified identical, 474 test runs, 0.35ms vs 2.98ms / ~8.5x personalized read, 4-14x overall, 1 vCPU/1GB free node, ~4ms CDN margin, 17-24ms vs 34-55ms Ably numbers) was located verbatim or near-verbatim in the corresponding snapshot. Both direct quotes ('We built one identical app twice...' in source-2; 'A personalized read is not a network request to another service...' attributed to Aleks Haugom in source-4) are exact, word-for-word matches to the snapshot text, correctly attributed by name and title. The Databricks Lakebase aside and the InfoQ 'has not been re-run' / warm-in-memory-dataset caveats in 'What We Don't Know' are also verbatim-sourced to source-3.

One issue found: the Overview's claim that Harper 'shipped version 5.2 ... on August 20, 2026 ... according to Harper's release announcement and InfoQ, which dated the release to August 20' overstates the sourcing. Harper's own release-announcement snapshot (source-0) is bylined August 6, 2026, and Harper's technical deep-dive (source-1) is bylined August 17, 2026 — neither Harper post states an explicit ship date in its prose. August 20, 2026 is InfoQ's own article publish date (source-3); InfoQ's body text says only 'Harper recently released version 5.2' without restating a specific date. The bot appears to have conflated InfoQ's dateline with an explicit release-date statement and then also credited Harper's own announcement with confirming that same date, which the byline contradicts. This is a single, narrow, and cleanly correctable factual slip in the lead paragraph — it does not affect the headline or summary (neither mentions a date), and every other claim in the article, including all of the headline's own P99 figures, is independently and precisely verified against the sources.

A well-sourced, precisely fact-checked technical News piece. Every number and quote in the headline, summary, and body checks out against the five source snapshots except for one narrow, non-headline date misattribution in the lead, which is honestly correctable with a single public corrections note. Recommend APPROVE_WITH_CORRECTIONS.