Corrections
When we get something substantively wrong, we fix it and log the fix here; small copyedits live in each article's revision history instead. Found an error? Use the "Report issue" button on the article (no account needed). So far we have received 2 reports and reviewed 1, taking a median of 4.8 days. 1 is waiting, the oldest since September 26, 2026. That is our record rather than a promise. See the verifiability policy for how checking works.
321 articles were corrected in the last 30 days. Every correction note and resolved report is listed below, newest first. A fix applied to several articles with the same note is shown once, listing each article; "See the change" opens that version in the article's revision history.
- editorial
Kilo Code's lineage and facts. It forked Roo Code (10 Mar 2025), not Cline, so it is a grandchild of Cline, not one of 'two significant forks' of it; the $8M seed was announced 10 Dec 2025, not 'late 2025'; Orchestrator mode and Memory Bank are now deprecated per Kilo's docs; and the comparison-table row now reflects VS Code/JetBrains/CLI/cloud, MIT and 500+ models. Added kilo_code link and seven references.
- editorial
Attributed the $100,000 store seeding and the error list to METR, which is the source for them, and added the Union Street location and the candles oversupply METR also reports; linked Andon Labs and Claude Sonnet 4.6.
- editorial
Vending-Bench was described as running "roughly a year of in-simulation days" (runs are capped at 2,000 messages) and Grok 4's top placing was dated to August 2025 with no source. Grok 4 took the top spot on July 10, 2025; the leaderboard ranks by minimum net worth, and as of 1 Oct 2026 Grok 4 is second behind Gemini 3 Pro. Sourced ref 38 to the archived August 13, 2025 update and removed two unsupported claims (Grok 4 run traces that abandoned the simulation, and xAI not leading with the eval in its marketing: Andon Labs was in fact invited to present the update).
- editorial
The Vending-Bench 2 figure was called "simulated profit" on a "multi-day" benchmark and carried no citation. It is the mean closing bank balance after one simulated year from a $500 start; sourced to Andon Labs and dated, and the unverifiable "widely cited as a concrete signal" framing removed.
- editorial
Vending-Bench 2 figure was labelled "simulated profit" on a "multi-day" benchmark. $5,478.16 is the mean closing bank balance after one simulated year from a $500 start; added the Andon Labs sources, the error bar and the leaderboard date.
- editorial
Fixed reference 1's title, the METR scaffolding chronology (July/December 2023, not early 2024), the de Macedo paper's boundary list (it never names LangChain and names five neighbours), the SWE-agent Lite figure's framing, the METR 26pp/8pp figures and the elicitation attribution, the lm-evaluation-harness quote source and Zenodo DOI, the MMLU table label and the HELM read-out description, the Open LLM Leaderboard quote, the tau-bench authorship, the test-harness definitions' sources, the TechTarget URL and the harness.io founding wording; added OpenRouter's own dated app rankings with its stated caveats, identified omp as Oh My Pi, Command Code and Freebuff, and added Harness-Bench's measured 23.8-point harness gap, HELM's maintenance mode and simple-evals' deprecation.
- editorial
There was no ROSCon 2021 (the in-person event was cancelled; ROS World 2021 ran online), and TurtleBot 4 was announced by Clearpath Robotics and Open Robotics, with iRobot supplying the Create 3 base
- editorial
Three calendar days began in 2022 not 2024; the PyCon framing is not on every edition; 2012 and 2013 venues are on the archive; the sponsor-prospectus figures are a 2026 projection; MuJoCo attribution; bronze sponsor conflict; organizer naming
- editorial
Remove an argument not made by the cited outlet; AnewFold demo status; schema.org addresses not a footer; preprint corresponding-institution strings; spin-out citation
- editorial
Canonical Vending-Bench slug
- editorial
The Arena misbehavior findings were Anthropic's report of Andon Labs' testing of Claude Opus 4.6 specifically (not "the highest-performing model"), run length is 2,000 messages with o3-mini reaching 222 simulated days (not "dozens"), runs consume about 25M tokens over 5-10 real hours, tool steps are 5/25/75 minutes or 5 hours, search is Perplexity, and the Inc. reference did not support the meltdown passage (replaced with the Andon Labs eval page). Update: Vending-Bench was deprecated on 18 November 2025; adds Vending-Bench 2 (methodology, system prompt, 1 Oct 2026 leaderboard with standard errors and run counts, trend fits, $63k headroom estimate), Vending-Bench Arena rounds 1-13, the per-model behavior analyses, and the Gemini 4 Argon result.
- editorial2 articles
Huawei pulled the Ascend 960DT forward to Q1 2027 in September 2026
- editorial
FlashMLA shipped Feb 2025 for V3/R1 and is not the V2/V3 architecture; DeepSeek-V3 is 671B parameters
- editorial
660 TFLOPS is the April 2025 update, not a later one (day one was 580); TFLOPS unit
- editorial
DeepSeek-V2 5.76x throughput and 42.5% training-cost figures are whole-model versus DeepSeek 67B, not MLA alone; remove the back-computed 8,700 tps figure
- editorial
The Sep 2025 row should not state H800 SXM5
- editorial
State the evidence for the #EP column reading, fix the V1 requirements and NVLink citations, and source the MoonEP comparison
- editorial
The README Performance section was removed on 14 April 2025, not in the July 2025 refactor; NVRTC citation; TileLang affiliations; install-time build wording
- editorial
TileLang is external work (tile-ai), not DeepSeek's own DSL
- editorial
TileLang is the tile-ai project, not DeepSeek's; the paper reports ~70 lines reaching up to 98 percent of FlashMLA on H100 (95 percent was versus AITER on MI300X)
- editorial
The Ascend build defaults to dav-3510 (Ascend 950), so the validation hardware is stated; inclusive 2^22 bound; LiteTopK attribution
- editorial
Moe_topk_gate_forward parameters are required, not optional
- editorial
Give all four per-device Triton speedups (the 1.08x-1.25x range dropped the A100 1.03x figure); fork count; average-speedup wording; re-attribute a paraphrase to the Zhidongxi post
- editorial
Reference authors and dates, attribution of coverage claims, and the Gemini 4 naming citation
- editorial
CWE-bench rows are not all Google-run (Artificial Analysis published v1); reference and attribution fixes