Content Quality: News article of about 832 words (News range 400-1200). Structure: Overview, 'What MariaDB Says Is New', 'How the 56% Figure Was Produced', 'Availability Wording Differs', 'What We Don't Know'. Neutral, attributed throughout ('MariaDB says', 'as described by MariaDB'); no AI self-reference, no editorializing beyond the title's framing, which is attributed to MariaDB's own wording. The 'Other items, all as described by MariaDB' list covers vector search, multi-region and deployment/security, and is introduced as a selection, not an exhaustive list. No relative dates (the 'later this month' phrase appears only inside a verbatim quote, next to the explicit November 4, 2026 date). The limits section states what the sources do not establish (full report gated, no third-party repeat, no durable-commit results, 11.8.3-1 setting parity unknown), and each of those statements is true of the five snapshots.
Source Verification: Read all five committed snapshots under sources/2026-11/mariadb-enterprise-platform-2027-is-generally-available-its-56-throughput-claim-rests-on-a-mariadb-sponsored-test-mariadb-calls-non-durable/ (gunzip then HTML to text; all HTTP 200, no archive fallback, suspicious_patterns null for all five). source-0 (Blocks & Files, Chris Mellor, published 4 Nov 2026): CONFIRMS the same-day coverage, the six-component list, the 30% vector figure, '56 percent higher peak throughput over the last major release in third-party testing' (article correctly quotes 'in third-party testing'), the present-tense 'MariaDB AI Retrieval is available in public preview on MariaDB Cloud', and that it does not mention sponsorship or test conditions. source-1 (press release, 4 Nov 2026, Redwood City and Dublin): CONFIRMS 'generally available today', the six-component list, the Agent Memory/Semantic Cache/AI plugins quotes, 'now delivers up to 30% more queries per second than the last major release', the footnote 1 text (McKnight Consulting Group, 'High-Concurrency TPC-C Benchmarking,' October 2026; 'HammerDB TPROC-C (derived from TPC-C, not an official TPC-C result)'; 'in-memory engine scalability test'; Enterprise Server 12.3.3-0 vs. 11.8.3-1 on Google Cloud; 'Sponsored by MariaDB.'), and 'MariaDB AI Retrieval follows in public preview on MariaDB Cloud later this month.' source-2 (launch blog by Tina Yang, 4 Nov 2026): CONFIRMS 56% more peak throughput than version 11.8 in third-party McKnight testing, 'runs 25% faster than MySQL 9.7.2 at peak performance', the 30% vector figure 'than version 11.8 at the same accuracy', Raft Cluster GA and the 'Writes commit as soon as a majority of nodes confirm...' quote, Exa generally available on premises, BYOC GA running the MariaDB Cloud data plane in AWS, Azure or Google Cloud, FIPS 140-3 through the Enterprise Kubernetes Operator, and 'MariaDB AI Retrieval is available in public preview on MariaDB Cloud.' Minor nuance, not a correction: the blog body gives the 25% MySQL figure with 'It' referring to Enterprise Server 12.3, while its Key takeaways attribute it to the platform; the article follows the Key takeaways. source-3 (engineering blog by Adam Luciano and Steve Shaw, 4 Nov 2026, 'About 4x More OLTP Throughput than 10.6'): CONFIRMS 11.8.3-1 peak 1,702,758 NOPM at 88 virtual users, 12.3.3 2,653,370 NOPM at 152 virtual users, '~1.56x', NOPM defined as New Orders Per Minute, 'The configuration was deliberately in-memory and non-durable.' (in the test-procedure list of the McKnight testbed section, so it describes the same McKnight run that yields the 56%; the sentence is MariaDB's authors' own, and the blog does not say who set the configuration), '12.3.3 ran with innodb_index_shrink = OFF, a non-default setting', c4d-highmem-192-lssd with 96 cores and 192 vCPUs, 768 warehouses, 'McKnight did not test 10.6', the earlier separate run on 'a similar 96-core system', and 'treat the comparisons in this section as indicative rather than like-for-like'. source-4 (report landing page): CONFIRMS 'An independent performance test, sponsored by MariaDB', 56% more at best result vs 11.8, 'about 25% more orders per minute than MySQL 9.7.2 between 64 and 120 users, and 21% more at 152 users', 'At each database's best result, MariaDB's lead was 21%', 'so anyone can repeat the test and check the claims', and that the full report is behind a form. I extracted all 25 quoted strings in the body and matched each against the normalized snapshot text: all 25 are verbatim. Internal link /article/2026-03/20-mariadb-acquires-gridgain-to-build-sub-millisecond-data-infrastructure-for-agentic-ai exists and is described exactly (the March 9 announcement of a definitive agreement to acquire GridGain Systems). Independent-source recount: 1 of 5 (Blocks & Files); 4 are MariaDB first-party. Analysis requirements do not apply (category is News). Allowlist: mariadb.com independently confirmed and added (see findings); blocksandfiles.com already on the list.
Factual Accuracy: No fabricated or misattributed specifics found. The 56% figure is correctly described as peak-to-peak (11.8.3-1 at 88 users vs 12.3.3 at 152 users), Enterprise Server 12.3.3-0 vs 11.8.3-1, HammerDB TPROC-C. The two MySQL comparisons are presented as MariaDB words them: launch blog '25% faster at peak', report page about 25% at the same load between 64 and 120 users, 21% at 152 users, and 21% at each database's best result. The '4x vs 10.6' figure is correctly described as built on an earlier separate run and, in the blog's own words, indicative rather than like-for-like. The AI Retrieval wording difference (press release 'later this month' vs launch blog and Blocks & Files 'available') is presented with both quotes. Dates (November 4, 2026; October 2026 report) match the sources. The title claim 'Test MariaDB Calls Non-Durable' is accurate: it is MariaDB's engineering blog that says the configuration was deliberately in-memory and non-durable, and the same blog is the source for the 1.56x ratio.
Overall Assessment: Accurate, carefully attributed News coverage of a vendor GA announcement that surfaces the vendor's own disclosure of how its headline benchmark number was produced. Every quote, number, version and date verified against the snapshots; approved without corrections. Featured candidacy: 1 - solid, accurate release coverage with useful benchmark-context for database engineers, but a single vendor's product launch rather than an industry-wide story