# How the 28 September search check was counted

This is a dated check of Vedokrok's own site search, using public release `MHC-RPUB-20260920-75ad787a`. The [full result file](https://vedokrok.com/research/search-check-2026-09-28.json) contains every query, chosen destination, rank, result count and search mode. It also records the snapshot and SHA-256 hashes of the benchmark and implementation files used for the run.

## The questions

Editors wrote 37 ordinary task phrases and chose one Vedokrok page each phrase ought to retrieve. Twelve deliberately out-of-scope phrases checked whether search would return an unrelated page. These are tests written by the project, not queries sampled from a representative user population.

For each task phrase, the evaluator searched the selected site catalog and recorded the position of its chosen page. Position 1 means the first result; `null` means the page was not returned. “In the first five” counts a page only when its position is 1–5. A control counts as a correct silence only when search returned zero results. The file includes both default search and the explicitly broader search mode. The figures shown on `/lab` use default mode.

## What happened

In default mode, 31 of 37 chosen pages were first. All 37 were in the first five. The other six appeared second or third. All 12 out-of-scope phrases returned nothing. Fifteen of the 37 task searches used the search fallback mode; the result file identifies them individually.

Earlier misses led us to change the site and add one newly discovered phrase to this same set. The result is therefore a regression check, not an independent holdout or a blind evaluation. A selected destination being retrieved says nothing about whether it was the best page for a person, whether that person could use the advice, or how Google or an AI assistant would rank or cite the site.

## What you can verify

The result file lets you recalculate the counts above and inspect any query with a weaker rank. You can try a phrase in [site search](https://vedokrok.com/search), although today's search or public release may differ from this frozen run. The website source repository is private, so this note and result file do **not** make the software run independently reproducible. They make the reported inputs, counting rule and outcomes public. A future public code release or independently collected query set would support a stronger claim.

If you find a mismatch between the file, this explanation and the site, [report it](https://vedokrok.com/contact?topic=Correction%20or%20source%20concern).