After reading this page you can tell, for any brief on this site, what was done by a person, what was done by an AI agent, and how much weight each claim can bear.
A person owns each question, and AI agents do the research
A person sets the question and owns the result. Julie Davila chooses each question, sets its scope, reads the drafts, sends work back, and decides whether a brief is published. She is accountable for what appears here.
AI agents do most of the research and drafting. The agents are built on Anthropic's Claude models. One agent breaks the question into separate lines of inquiry and gives each to a research agent. Those agents search the public web, open primary sources, and write their notes to disk with a source link and an evidence grade for each finding. The planning agent then reconciles the notes, resolves conflicts between sources, and drafts the brief.
A literature tool screens the academic work. For the peer-reviewed side, the agents use Elicit to search and screen the research literature and pull findings from each included paper. Where a brief relies on an Elicit screen, it says how many papers were included and whether full texts were read.
Some briefs get an adversarial check. For some briefs, a separate review pass, sometimes run on a model from a different vendor, checks the load-bearing claims against their sources and hunts for errors. A brief says whether it had one.
What every brief carries
- The answer first. Each brief opens with what it concludes, then the evidence.
- A source for every external figure. Numbers and quotations link to where they came from. A figure the agents could see only in a search result, without opening the source itself, is marked as such or left out.
- Labelled synthesis. Where a brief goes beyond its sources and draws its own conclusions, the passage is marked Synthesis. Treat those passages as reasoned hypotheses, not findings.
- A "How this was made" section at its end: the question, the tools, the number of sources, what was checked and what was not.
How the evidence is graded
| Grade | What it means | How much weight it bears |
|---|---|---|
| Peer-reviewed | Published in a refereed venue | The most, within the study's own sample and method |
| Primary industry | An organization reporting its own results, on its own blog or in its own report | Real experience, self-reported, rarely audited |
| Vendor | A company that sells a product in the area, reporting data that supports buying it | Directional only; the incentive is named beside the figure |
| Secondary | A third party reporting someone else's result | Only as good as the original, which was not read |
| Practitioner | Widely argued or practised, with little public measurement | A starting point to test |
| Synthesis | An inference drawn in the brief itself | A hypothesis to check against your own data |
Where the method falls short
AI agents make mistakes a careful human researcher would also make, and a few a human would not. They can misread a table, lift a figure from a summary rather than the paper, or join two sources too eagerly. The steps above narrow those risks but do not remove them. When a finding matters to a decision, open the source it links to.
The briefs use public sources only. None of them contains any organization's private data, and none describes a specific organization's internal program.
Use, citation and corrections
The text on this site is licensed CC BY 4.0: share and adapt it with attribution. To cite a brief, give its title, "Julie Davila, Research", its version, and its address. To save one as a PDF, use the Save as PDF button at the top of the page or your browser's print menu; pages print in light colours with the navigation removed.
If you find an error, open an issue on the site's GitHub repository with the page, the claim, and the source that contradicts it. Corrections are listed in the brief they change, with the date.
Julie Davila · Research. Researched with AI agents and reviewed by Julie Davila: how this research is made. Text licensed CC BY 4.0. Source on GitHub.