Method
How we check our work
Every claim on this site links to its evidence. Where a link is missing, we show it as missing.
Status
Five labels, nothing else.
- Live
- Anyone can use or read it today.
- In progress
- Work is underway and visible.
- Planned
- An intention only. Shown dashed, and never with imagery.
- Submitted
- Sent to a venue.
- Published
- Public, with an archived DOI.
Paper pages
What every paper page carries.
- Findings, as figures with their conditions.
- Stated limitations.
- A link to the code.
- An archived DOI.
- A BibTeX entry.
The full list of papers is on Research.
Practice
What our papers show.
- Protocols frozen before data. The trial-count audit followed a protocol frozen before any data were extracted.
- Pre-registered study. The physics-guards paper tests its shields in a pre-registered MuJoCo study.
- Intervals reported. Where a figure is a rate, we give its interval, such as the 95% interval on the capscope result.
- Simulation labelled as simulation. Results that come only from simulation say so, on the paper page and in the limitations.
- A claim withdrawn after audit. We withdrew a 38% success rate of our own after a per-episode audit. Read the correction.
Corrections and open questions are kept on the Lab page.
Limits
What we do not do.
- We do not invent numbers.
- We do not use stock imagery.
- We do not publish testimonials.