---
title: "Show Your Work: AI and Robots Are Being Graded on Real Results, Not Demos"
date: 2026-10-29
author: Ember & Oak Institute, LLC
publisher: Ember & Oak Institute, LLC
canonical: https://emberandoakinstitute.com/blog/graded-on-real-results.html
html: https://emberandoakinstitute.com/blog/graded-on-real-results.html
license_note: "Copyright Ember & Oak Institute, LLC. Quote with attribution and a link to the canonical URL."
sources:
---

# Show Your Work: AI and Robots Are Being Graded on Real Results, Not Demos

**Robot makers love a good demo: a humanoid folding laundry, a robot arm stacking boxes, a quick clip that racks up views. On Oct. 29 in Shanghai, the Chinese robot company AGIBOT proposed counting something less flashy. Its new yardstick, called Worken, credits a robot only for the hours it actually works on its own, and only for the useful output it produces in that time.**

Worken, [unveiled at AGIBOT Innovation Day 2026](https://www.agibot.com/article/231/detail/131.html), is one of several stories this week that share a theme. In AI and robotics, the question is shifting from “look what it can do” to “show your work”: prove the output is real, prove the mistakes get caught, and prove the system can stop when it should. A new study tested AI as a reviewer of medical research. A neuroscience startup reported letting an AI agent run parts of a lab procedure, with humans holding a stop button. And a small open-source robot built for home tinkerers just got wheels.

For families who learn by directing AI rather than writing code by hand, these stories are a guide to the hardest part of the “idea, build, test, improve” loop: deciding what counts as good work, and checking it honestly.

### A yardstick for robot work

AGIBOT describes Worken, a mash-up of “work” and “token,” as “Effective Intelligent Work Hours.” The formula multiplies three things, according to the [company’s announcement](https://www.prnewswire.com/news-releases/agibot-introduces-worken-to-measure-productive-robot-work-at-agibot-innovation-day-2026-302923486.html): the number of robots actually doing tasks, the hours they work autonomously (not just the hours they are switched on) and an efficiency ratio that compares their output with a reference benchmark.

The point is to stop rewarding robots for looking busy. As AGIBOT’s Peng Zhihui framed the question, the industry should ask of its robots, “how much productive work are they actually delivering?”

Two cautions apply. Worken is the company’s own measure, not an independent standard, and AGIBOT itself calls it an “evolving framework” that it plans to refine with data from real deployments. Still, the idea is a useful one for anyone building with robots: decide in advance what “useful work” means, then measure that, not the highlight reel.

### AI as a second reader for medical research

Before a medical study is published, other experts are supposed to read it and catch mistakes. That process, called peer review, is slow and uneven. A [preprint posted Oct. 28 on medRxiv](https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2) tested whether AI could help. It has not yet been peer reviewed itself.

The researchers repeated a well-known 2008 experiment in which 607 human reviewers were sent study reports with errors deliberately planted in them. Those untrained reviewers found only about 27% of the major errors on average. This time, the team gave the same kind of test to 15 AI models, then added three newer clinical-trial reports (on stroke, psoriatic arthritis and heart failure) with their own planted errors.

![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 found about 27%. Claude Haiku 4.5 found 41.7%, the open-weight model gpt-oss-120b found 56.0%, the open-weight model Kimi K3 found 80.4% and Claude Opus 5.5 found 92.9%. Source: Brian, Pereira Hernández, Yin, Koury and Rust, medRxiv preprint, not peer reviewed, Oct. 28, 2026.](https://emberandoakinstitute.com/blog/graded-on-real-results/chart-error-detection.png)

Source: [Brian, Pereira Hernández, Yin, Koury & Rust, medRxiv preprint (not peer reviewed), Oct. 28, 2026](https://doi.org/10.64898/2026.10.25.26366455). The human figure comes from older test papers, so the comparison is rough. Chart: Ember & Oak Institute, LLC

On the new trial reports, the models caught between 41.7% and 92.9% of the planted errors. Freely downloadable “open-weight” models ranged from 56.0% to 80.4%. Each AI review took between 1.4 and 16.3 minutes, the authors report, compared with five to eight and a half hours for a human review.

The catch came on clean copies of the same papers, with nothing planted. Every model flagged some problems that weren’t errors, and the models that found the most real mistakes tended to raise the most false alarms. The best error-finder, Claude Opus 5.5, correctly left alone 71.4% of the error types on the clean papers, while the most cautious model, Qwen3.8-27B, left alone 97.6%. The authors note that some of those flags may point to genuine weaknesses in the papers, so the false-alarm numbers are a worst case.

![Bar chart of error types correctly left unflagged on unmodified versions of the 2026 trial reports. Claude Opus 5.5, the best at finding errors, scored 71.4%. All 15 models combined scored 83.0%. The open-weight model Qwen3.8-27B scored 97.6%. Source: Brian, Pereira Hernández, Yin, Koury and Rust, medRxiv preprint, not peer reviewed, Oct. 28, 2026.](https://emberandoakinstitute.com/blog/graded-on-real-results/chart-false-alarms.png)

Source: [Brian, Pereira Hernández, Yin, Koury & Rust, medRxiv preprint (not peer reviewed), Oct. 28, 2026](https://doi.org/10.64898/2026.10.25.26366455). Chart: Ember & Oak Institute, LLC

There are limits. Every planted error was visible in the written text, so the test says nothing about catching faked data or bugs in a study’s code. The authors are also affiliated with the ResearchHub Foundation, which is building AI-assisted review, a conflict they disclose. They have [posted their test code publicly](https://github.com/rustlab1/llm-peer-review-benchmark) so others can check the work.

Their conclusion is measured: “relying on human peer review alone may leave avoidable methodological errors undetected.” In other words, AI looks like a strong second reader, not a replacement for the first.

### Guardrails that can say no

A more unusual report came from Gabriel Ocana-Santero, founder of the neuroscience startup Exin Therapeutics, who [described on X](https://x.com/OcanaSantero/status/2115442964574908902) an experiment in which an AI agent planned and guided a gene-therapy procedure on a lab mouse used to study epilepsy. This is the founder’s own account. He says a technical report is coming, and none of it has been independently reviewed.

According to his thread, the agent was given a spreadsheet of 19 candidate gene-therapy tools and one goal: reduce seizures. It chose a treatment meant to boost the brain’s natural calming signals, picked a target region by averaging coordinates from three published papers and ran the injection through a custom motorized rig. It watched through the rig’s camera and logged 32 observations, including, in its own words, “No bleeding or backflow visible at the opening.”

The humans kept firm control. A surgeon did the delicate opening step by hand, and a licensed veterinarian approved each step that couldn’t be undone while holding a physical emergency stop. Ocana-Santero also reported that not every AI model would take part: one model declined the live procedure after helping with a practice run, and another refused outright.

The most telling moment was a failure. A calibration problem in the rig’s software made the agent compute a target outside the machine’s range. Rather than push ahead, the agent stopped. “I’m not retrying,” it wrote, and after the rig was reset and the problem returned, “I’m not trying a third time.” Ocana-Santero wrote that “it was right,” and he fixed the code by hand. He reported that the mouse recovered well.

Whatever one thinks of AI near living animals, the design lesson carries over to any project: a system worth trusting has clear points where a person must sign off, a stop button within reach and the willingness to quit when the numbers don’t make sense.

### Bringing the test loop home

Not all of this week’s robot news happened in labs. [Reachy Mini](https://pollen-robotics.com/reachy-mini/), the small, open-source desktop robot from Hugging Face’s Pollen Robotics, is getting a [two-wheeled base](https://store.pollen-robotics.com/products/reachy-mini-hoverboard) that lets it roll around a room, steered from a phone with live video from the robot’s camera. The base works with the Wireless version of the robot, and first orders are expected to ship at the end of January 2027, [RoboActu reported](https://www.roboactu.fr/pollen-robotics-reachy-mini-roues-application-mobile/).

A free mobile app now sets up and runs the robot without a computer, according to Pollen’s [getting-started guide](https://pollen-robotics.com/reachy-mini/getting-started/). Pollen’s site lists more than 500 community-made apps and a “vibe code it” option that fits how our learners work: “Describe it in a chat. An AI agent writes the code, a preview runs it on your real robot.”

That last step is the important one. A preview on a real robot is a test, and a robot that rolls is a robot that can bump into things, which makes a good test plan, and a stop plan, part of the fun.

### What learners can take from it

Each story this week turned on the same habits. In the “idea, build, test, improve” loop, they look like this:

- **Idea:** Decide what “good work” means before you start. AGIBOT’s Worken counts only real, autonomous output, not time spent switched on.
- **Build:** Let AI do the heavy lifting, but plan the moments when a person must approve the next step.
- **Test:** Plant a few known mistakes and see whether your checker catches them, and also hand it something correct to see whether it raises false alarms, the way the medRxiv researchers did.
- **Improve:** Keep a stop button handy, and treat “I’m not trying a third time” as a feature, not a failure.

The machines are getting better at the work. Showing that the work is real is still a job for the people directing them.

### Sources

- [AGIBOT, Innovation Day 2026 announcement of Worken (Oct. 29, 2026)](https://www.agibot.com/article/231/detail/131.html)
- [AGIBOT via PR Newswire, “AGIBOT Introduces Worken to Measure Productive Robot Work at AGIBOT Innovation Day 2026” (Oct. 29, 2026)](https://www.prnewswire.com/news-releases/agibot-introduces-worken-to-measure-productive-robot-work-at-agibot-innovation-day-2026-302923486.html)
- [Brian, Pereira Hernández, Yin, Koury & Rust, LLM peer-review preprint, medRxiv (Oct. 28, 2026; not peer reviewed)](https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2)
- [DOI for the medRxiv preprint](https://doi.org/10.64898/2026.10.25.26366455)
- [LLM peer-review benchmark code (GitHub)](https://github.com/rustlab1/llm-peer-review-benchmark)
- [Gabriel Ocana-Santero (Exin Therapeutics), thread on X (Oct. 28, 2026; self-reported, not peer reviewed)](https://x.com/OcanaSantero/status/2115442964574908902)
- [Pollen Robotics, Reachy Mini](https://pollen-robotics.com/reachy-mini/)
- [Pollen Robotics, Reachy Mini getting started](https://pollen-robotics.com/reachy-mini/getting-started/)
- [Pollen Robotics store, Reachy Mini wheeled base](https://store.pollen-robotics.com/products/reachy-mini-hoverboard)
- [RoboActu, Pollen Robotics Reachy Mini wheels and mobile app (Oct. 29, 2026)](https://www.roboactu.fr/pollen-robotics-reachy-mini-roues-application-mobile/)

## Sources

- [unveiled at AGIBOT Innovation Day 2026](https://www.agibot.com/article/231/detail/131.html)
- [company’s announcement](https://www.prnewswire.com/news-releases/agibot-introduces-worken-to-measure-productive-robot-work-at-agibot-innovation-day-2026-302923486.html)
- [preprint posted Oct. 28 on medRxiv](https://www.medrxiv.org/content/10.64898/2026.10.25.26366455v2)
- [Brian, Pereira Hernández, Yin, Koury & Rust, medRxiv preprint (not peer reviewed), Oct. 28, 2026](https://doi.org/10.64898/2026.10.25.26366455)
- [posted their test code publicly](https://github.com/rustlab1/llm-peer-review-benchmark)
- [described on X](https://x.com/OcanaSantero/status/2115442964574908902)
- [Reachy Mini](https://pollen-robotics.com/reachy-mini/)
- [two-wheeled base](https://store.pollen-robotics.com/products/reachy-mini-hoverboard)
- [RoboActu reported](https://www.roboactu.fr/pollen-robotics-reachy-mini-roues-application-mobile/)
- [getting-started guide](https://pollen-robotics.com/reachy-mini/getting-started/)

Source: Ember & Oak Institute, LLC, https://emberandoakinstitute.com/blog/graded-on-real-results.html