Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Writers, developers and analysts using AI research tools
Before you start: A bounded question and a way to open primary sources
⚡ TL;DR — the 30-second version
- One consequential claim, one inspectable evidence record.
- Provider claims remain attributed.
- Publication date and verification date are different facts.
AI research tools can find candidate sources and organize material quickly. They can also cite an irrelevant page, combine facts from different versions or turn a vendor's claim into an apparently independent conclusion. The output needs an evidence review, not just a spelling pass.
Start by defining the scope: which date, region, product edition and decision does the research concern? 'Best AI model' is not a research question. 'Which currently available models accept our input formats and fit a documented budget?' is.
“The right question is not 'does it have citations?' It is 'which claim does each citation actually establish?'”
Field guide / 01
01 / Keep one row per consequential claim
Record the claim, source URL, publication date, verification date and a short supporting excerpt or location. Separate the provider's reported measurement from your interpretation. If two sources disagree, retain both until you understand why.
The site’s Evidence Ledger in the Lab exports a template for this structure. A row is useful even when the answer is unknown; that is better than disguising an evidence gap as a confident sentence.
| Field | Why it exists |
|---|---|
| Claim | A precise statement that can be supported or rejected |
| Source and location | Where the evidence actually appears |
| Published / checked | History versus your latest review |
| Evidence class | Specification, provider report, independent test or interpretation |
| Caveat | Conditions that would make the claim misleading |
Field guide / 02
02 / Prefer the closest source to the fact
Use model cards for documented capabilities, release notes for changes and current price sheets for billing. A launch article may become outdated after a correction. An index snippet may summarize an older page even when the live documentation has changed.
Provider sources are authoritative about their published terms, not neutral evidence that they outperform every competitor. Treat comparative claims as attributed results with their methodology attached.
Field guide / 03
03 / Check support, not just availability
Open the cited page. Find the exact statement. Confirm that its date, unit and scope match the generated claim. A link resolving successfully proves only that the page exists.
For a price, check the unit, caching, long-context threshold, introductory expiry and region. For a score, check the benchmark version, harness, effort and whether tools were available. For availability, distinguish announced, preview, gated and general access.
Field guide / 04
04 / Leave a revision trail
Publish a reviewed date without resetting it every time the site builds. Log substantive corrections and maintain stable links to your source list. Do not rename an old event as breaking news simply to keep a feed busy.
For ongoing work, identify which claims are likely to expire. Prices and model access need more frequent review than a general explanation of precision and recall. A stale-data warning can be more honest and useful than an automatic 'updated today' label.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01Perplexity — Sonar and Search API changelogPerplexity↗
- 02Sonar Deep Research — model and billing categoriesPerplexity↗
- 03DeepSeek changelog — V4.1 Flash and V4 Pro correctionDeepSeek↗
- 04Introducing Claude Opus 5.5Anthropic↗