Original research
Research
First-hand measurements and the apparatus behind them. Every page states whether it carries results or only a method — because the failure mode of a research section is a plausible methodology with invented numbers under it.
2 measured2 published harnesses awaiting a run
- Measured
GitHub's Copilot Docs Contradicted Its Own Changelog on Retirement Day
When GitHub retires a Copilot model, does its reference documentation reflect that on the day the retirement takes effect?
On 1 September 2026, six Copilot models were retired. GitHub's supported-models reference page listed all six as generally available, with no deprecation notice anywhere on it.
- Measured
Who Actually Crawls a Developer Documentation Site
On a small, new technical documentation site, what proportion of HTTP requests come from people, and how does AI-assistant crawling compare to search-engine crawling?
Fifteen days of server logs from a 129-page documentation site: 90.9% of requests were not people, and AI assistants fetched more than every search engine combined.
- Harness — no results
A Test Matrix for Copilot CLI Permission Boundaries
Does a Copilot CLI permission configuration actually refuse what it appears to refuse, and does the refusal hold when the request is phrased as something the agent wants to do?
Twenty-four permission configurations paired with the commands each should allow and refuse, so an agent's tool boundary can be verified rather than assumed.
- Harness — no results
A Seeded-Defect Corpus for Testing AI Code Review
Of ten deliberately seeded defects across seven categories, how many does an AI code reviewer find — and how many of its findings come with a reproduction rather than a hedge?
Ten deliberately planted defects with a machine-readable answer key and a scoring harness, so claims about what AI review catches can be checked rather than asserted.
How this works
Every page carries its question, environment, limitations and reproduction steps in structured frontmatter rather than in prose, so none of them can be quietly dropped in an edit. The build fails if a page claims a measurement without describing what was measured or the environment it was measured in.
Where an experiment cannot be run here, the apparatus is published and the page says so at the top. We would rather publish a corpus anyone can run than a number nobody can check. The testing methodology covers the general rules.