Provenance Record

Long-form News piece (1476 words) with strong structure: Overview / What We Know (substructured: the change itself, why reverted, bug report, both series affected, repairs thread) / What We Don't Know / Analysis. Heavy attribution density across 10 sources, technical detail appropriate for the audience.

10 sources, all snapshots HTTP 200. Verified: source-0 (Python Insider 3.14.5): van Kemenade 'has been reverted back to the generational garbage collector from 3.13, due to a number of reports of significant memory pressure in production environments' verbatim, 154 bugfixes figure. source-1 (Python Insider RC1): May 4 RC1 date, May 8 originally planned final. source-2 (python.org downloads 3.14.5): 154 bugfixes confirmed, May 10 release. source-3 (Discourse 'reverting the incremental gc'): van Kemenade three-reason list ('old GC is a known quantity', 'didn't go through the PEP process', 'rolled back just before the final release of 3.13') verbatim; Schemenauer '5x was the worst case I saw' and 'Peak RSS size was 2.7x with incremental GC' verbatim; Steering Council discussion. source-4 (Discourse 'improving incremental gc'): Tim Peters opening with 'Nothing at all gets collected until about the 20 thousandth iteration' verbatim; Schemenauer 'should not take major surgery to make the incremental GC work better' verbatim; 'trigger GC every 2000 net new objects, like the generational GC' verbatim; 'size the increments (how many old objects to look at) such that we effectively do a full collection often enough' verbatim. source-5 (Discourse PEP 3145 release announcement): May 10 / 154 bugfixes / Steering Council reference. source-6 (cpython issue #142516): HenriBlacksmith filed December 10, 2025; 'I already opened a bug on urllib3 for this issue, it looks that may application using MSAL for Python library (which relies on requests, which relies on urllib3, which relies on ssl) is leaking memory when being upgraded from Python 3.13 to Python 3.14' verbatim; ssl.SSLContext.load_verify_locations confirmed. source-7 (Python docs whatsnew 3.14): all three direct quotes verbatim ('The cycle garbage collector is now incremental. This means that maximum pause times are reduced by an order of magnitude or more for larger heaps'; 'young and old'; 'the young generation and an increment of the old generation, instead of collecting one or more generations'; the post-revert text 'Python 3.14.0-3.14.4 shipped with a new incremental GC...'). source-8 (Adam Johnson blog): 512 MB Heroku dyno, 527M(101.5%) and 1033M(201.7%) memory figures, ModelState.render Memray profile verbatim, 'cannot keep up with the migration process's tendency to create many cyclical objects' verbatim, 'core team is making the right call reverting it' verbatim. source-9 (igor'sLAB): 'corrective release for real operating conditions' verbatim, 'theoretical latency gain helps little if services consume more memory under real load than planned' verbatim. ATTRIBUTION ISSUE: Two verbatim quotes in the 'A separate thread is already exploring repairs' section are NOT present in source-4 or any other snapshot, AND are NOT present in the live Discourse threads (verified via WebFetch): (a) Mark Shannon 'Collect gen 0 + increment of gen 2' / 'Adding the gen0 collection to the increment is trivial; it's a three line change'; (b) Sergey Miryanov 'The incremental GC has early return since 3.14.1. Even if threshold0 reached, it can exit early and skip all processing'. These appear to be fabricated quotes. The substantive editorial claim of that section (core developers are actively exploring fixes to the incremental GC) IS supported by Tim Peters' verified opening of the thread and Schemenauer's verified replies. The named individuals (Shannon, Miryanov) DO appear in the broader Discourse context — Sergey Miryanov is referenced as offering to create a PR, and Tim Peters references 'Mark's sketch of reintroducing 3 generations' — but the specific direct-quoted strings attributed to them in the article are not in the source material.

Headline (3.14.5 reverts incremental GC), summary (May 10 ship, 5x memory regression, also affects 3.15), Overview lead (van Kemenade quote, 154 bugfixes), and Analysis are all soundly supported. The two fabricated quotes are in a subordinate body section. Tim Peters' and Schemenauer's verified contributions to the same thread support the same editorial point.

Publish with a substantive correction. The article is otherwise an unusually thorough technical piece on a significant CPython release event. The bot's invention of two quotes from named individuals in a supporting section is a real problem and must be flagged transparently — but it does not gut the article's spine, and honest corrections coverage is feasible.