# Research with a claim ledger

Choose sources by what they can establish. A paper can describe its experiment;
a release announcement can establish what a vendor announced; neither establishes
that the same result will occur in a specific business.

1. Search the topic and an opposing interpretation. Record a candidate angle.
2. Open primary sources. Record the actual URL after redirects and access date.
3. Check the source's publication or revision date separately from the news event.
4. For each important claim, record the supporting passage in brief notes, the
   population or model, the metric, and the limitation. Do not store whole articles.
5. Look for updates, retractions, version changes and conflicting measurements.
6. Resolve conflicts or describe them. Do not average incompatible percentages.

For workplace studies, distinguish experiments, observational studies, surveys,
and vendor customer stories. Report measured outcomes rather than translating
speed into headcount savings. Compare quality and time, including review and
rework. A satisfaction survey is not a productivity experiment.

For frontier models, record exact model/version, announcement date, product/API
surface and comparison settings. Recheck pricing and availability on the day of
publication if mentioning them. Describe vendor benchmarks as vendor-reported.
An independent study on a different model can inform evaluation design but cannot
validate the newly released model. Avoid claiming a model is the newest without
checking the whole relevant release history.

For agent skills, consult the specification and the target product's docs.
A readable Markdown file, a native installed skill and an agent with tools are
three different things. Never assume cross-product execution compatibility.

Suggested evidence record fields: slug, researchedAt, readerDecision, thesis,
originalContribution, sources[{url,title,publisher,publicationDate,accessedAt,
claims,limitations}], contradictions, verification, editorialReview. Use null
rather than guessing a date. Keep private maintenance notes out of public bundles.