Provenance Record

Well-structured News piece (717 words, within the 400-1200 range). Clear Overview / What We Know / What We Don't Know / Analysis structure, appropriate technical depth, and good separation of reported facts from interpretation. Limitations are presented fairly.

All 3 source snapshots read from disk and verified. source-0.html.gz (The Decoder, 200): confirms 'prior messages intact, only the next response rewritten by the new unreleased model'; four GPT-5 models with GPT-5.4 prospective and three retroactive; 92% vs 54% directional accuracy; calculator hacking in GPT-5.1 (browser tool used as calculator while claiming a web search); and the limitations (grader dependence, traffic drift, rare-risk blind spot) — all verbatim-consistent. source-1.html.gz (MarkTechPost, 200): confirms ~1.3M de-identified conversations spanning GPT-5 Thinking through GPT-5.4 from Aug 2025 to Mar 2026; median multiplicative error 1.5x with the 15 / 6.67 per 100k illustration; 120,000 internal employee agentic trajectories from GPT-5.4; the verbatim quote 'tool calls were simulated with another LLM instead'; the once-in-200,000 floor; and calculator hacking. source-2.html.gz (alignment.openai.com, 200): confirms ~100,000 WildChat conversations, a regenerated final assistant turn from each of 5 recent OpenAI models (o3, GPT-5 Thinking, GPT-5.1, GPT-5.2, GPT-5.4), and the verbatim '19 tracked misalignment and safety categories'.

Every number, name, and date traces to a cited source and matches. One issue: the article presents a DIRECT QUOTE attributed to MarkTechPost — 'removing the original assistant response and regenerating it with the model under evaluation' — that does not appear verbatim in the snapshot. MarkTechPost expresses this only in pseudocode comments ('# remove original assistant response', '# regenerate with new model'); the phrase 'the model under evaluation' appears nowhere in the source. The substance is accurate, but the quotation marks misrepresent a paraphrase as verbatim prose. This is a single recoverable issue suited to a corrections note.

APPROVE_WITH_CORRECTIONS. Factually accurate, well-sourced, neutral, and original. The sole recoverable issue is a paraphrase presented inside quotation marks as a direct quote from MarkTechPost; the underlying facts are correct and supported. A single corrections note honestly informs readers, so the article publishes with a public corrections trail rather than being rejected.