2 November: Research searches the web for real, grades its own report, and always ends with a document
Research now runs real searches, accepts only pages with a finding as sources, has a reviewer grade the draft, shows a progress card, and opens the report in the side column.
If you use Research, reports rest on better sources, show a grade, can be downloaded in five formats, and no longer end without a result. No action needed.
An outside review graded one Research report 4 out of 10. 20 of its 41 sources were search result pages, and the vendor documentation on the actual mechanism was missing. Research changed in three ways on 2 November.
Text version
Search: real queries, evidence-only sources. Review: ranked hypotheses, a reviewer grade, up to two follow-up rounds. Delivery: progress card, report in the side column, export in five formats, and a report even if the model fails.
Better sources
The planner's queries are searched, and only the target pages are fetched. Search pages, home pages and login pages never count as sources. A page becomes a source only when a claim was extracted from it. Reports carry ranked hypotheses with what speaks for and against them, how to verify them, similar cases and an ordered list of next steps.
A reviewer
A reviewer grades each draft from 0 to 10 against a fixed rubric. Below 8, up to two follow-up rounds research exactly the named gaps, and the best draft wins. The grade and a factual method line go into the report. The grade is our own check, not an outside audit.
What you see and what no longer fails
A card shows five steps (plan, sources, write, verify, report) and one row per facet. The finished report opens in the right-hand column and can be downloaded as PDF, Word, Markdown, HTML or plain text. The export runs in your browser.
One run had researched for 21 minutes and ended without a report, because the synthesis was cut at its length limit. A run now always ends with a document: one compact retry, then a report built from the extracted findings without a model. Search engines that block us are skipped for 30 minutes.
We have not yet re-graded Research with an outside review since these changes.