{ "slug": "graded-on-real-results", "title": "Show Your Work: AI and Robots Are Being Graded on Real Results, Not Demos", "url": "https://emberandoakinstitute.com/blog/graded-on-real-results.html", "date": "2026-10-29", "category": "AI news", "tags": [ "artificial intelligence", "robotics", "AI peer review", "AI safety", "open-source robots", "measuring AI", "AI news" ], "summary": "AGIBOT’s Worken metric, a medRxiv preprint on AI peer reviewers, an AI lab agent that refused to retry a faulty step and Reachy Mini’s new wheeled base all point to one lesson: define good work, test it honestly and keep a stop button.", "images": { "hero": { "url": "https://emberandoakinstitute.com/blog/graded-on-real-results/hero.jpg", "alt": "Painted storybook scene in a cozy Mat-Su Valley log home on an autumn morning: a girl in a mustard sweater kneels on the floor holding a clipboard while a small friendly robot on two wheels rolls along a masking-tape path between little orange cones. Her mother kneels beside her, smiling, with one hand resting on a big red stop button on a wooden box. Birch trees, light snow and a mountain show through the window, and a wood stove glows behind them." }, "charts": [ { "url": "https://emberandoakinstitute.com/blog/graded-on-real-results/chart-error-detection.png", "alt": "Horizontal bar chart of the share of deliberately planted errors found on three 2026 clinical-trial reports: human reviewers in a 2008 study on older papers about 27%, Claude Haiku 4.5 41.7%, gpt-oss-120b 56.0%, Kimi K3 80.4%, Claude Opus 5.5 92.9%.", "source": "https://doi.org/10.64898/2026.10.25.26366455" }, { "url": "https://emberandoakinstitute.com/blog/graded-on-real-results/chart-false-alarms.png", "alt": "Bar chart of error types correctly left unflagged on clean versions of the 2026 trial reports: Claude Opus 5.5 71.4%, all 15 models combined 83.0%, Qwen3.8-27B 97.6%.", "source": "https://doi.org/10.64898/2026.10.25.26366455" } ] }, "key_facts": [ { "fact": "AGIBOT’s Worken multiplies active robots × autonomous working hours × an output-efficiency ratio; AGIBOT calls it an evolving framework.", "sources": [ "https://www.agibot.com/article/231/detail/131.html", "https://www.prnewswire.com/news-releases/agibot-introduces-worken-to-measure-productive-robot-work-at-agibot-innovation-day-2026-302923486.html" ] }, { "fact": "In a 2008 study, 607 untrained human reviewers found about 27% of planted major errors on average.", "sources": [ "https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2" ] }, { "fact": "On three 2026 trial reports, 15 AI models caught 41.7% (Claude Haiku 4.5) to 92.9% (Claude Opus 5.5) of planted errors; open-weight models 56.0% to 80.4%.", "sources": [ "https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2" ] }, { "fact": "On clean papers, specificity ranged from 71.4% (Claude Opus 5.5) to 97.6% (Qwen3.8-27B), 83.0% overall; detection and false alarms traded off.", "sources": [ "https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2" ] }, { "fact": "AI reviews took 1.4 to 16.3 minutes versus 5 to 8.5 hours for a human review; all errors were visible in text; authors are affiliated with ResearchHub Foundation; preprint not peer reviewed.", "sources": [ "https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2", "https://github.com/rustlab1/llm-peer-review-benchmark" ] }, { "fact": "Exin Therapeutics’ founder reported an AI agent planned and guided a gene-therapy injection in an epilepsy mouse model with a surgeon and a veterinarian holding an e-stop; it declined to retry after a calibration fault (“I’m not retrying”); self-reported, not reviewed.", "sources": [ "https://x.com/OcanaSantero/status/2115442964574908902" ] }, { "fact": "Reachy Mini is getting a phone-driven two-wheeled base for the Wireless model, with first orders shipping end of January 2027, plus a no-computer mobile app and a ‘vibe code it’ feature.", "sources": [ "https://pollen-robotics.com/reachy-mini/", "https://pollen-robotics.com/reachy-mini/getting-started/", "https://store.pollen-robotics.com/products/reachy-mini-hoverboard", "https://www.roboactu.fr/pollen-robotics-reachy-mini-roues-application-mobile/" ] } ], "citations": [ { "name": "AGIBOT, Innovation Day 2026 announcement of Worken (Oct. 29, 2026)", "url": "https://www.agibot.com/article/231/detail/131.html" }, { "name": "AGIBOT via PR Newswire, “AGIBOT Introduces Worken to Measure Productive Robot Work at AGIBOT Innovation Day 2026” (Oct. 29, 2026)", "url": "https://www.prnewswire.com/news-releases/agibot-introduces-worken-to-measure-productive-robot-work-at-agibot-innovation-day-2026-302923486.html" }, { "name": "Brian, Pereira Hernández, Yin, Koury & Rust, LLM peer-review preprint, medRxiv (Oct. 28, 2026; not peer reviewed)", "url": "https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2" }, { "name": "DOI for the medRxiv preprint", "url": "https://doi.org/10.64898/2026.10.25.26366455" }, { "name": "LLM peer-review benchmark code (GitHub)", "url": "https://github.com/rustlab1/llm-peer-review-benchmark" }, { "name": "Gabriel Ocana-Santero (Exin Therapeutics), thread on X (Oct. 28, 2026; self-reported, not peer reviewed)", "url": "https://x.com/OcanaSantero/status/2115442964574908902" }, { "name": "Pollen Robotics, Reachy Mini", "url": "https://pollen-robotics.com/reachy-mini/" }, { "name": "Pollen Robotics, Reachy Mini getting started", "url": "https://pollen-robotics.com/reachy-mini/getting-started/" }, { "name": "Pollen Robotics store, Reachy Mini wheeled base", "url": "https://store.pollen-robotics.com/products/reachy-mini-hoverboard" }, { "name": "RoboActu, Pollen Robotics Reachy Mini wheels and mobile app (Oct. 29, 2026)", "url": "https://www.roboactu.fr/pollen-robotics-reachy-mini-roues-application-mobile/" } ] }