QueryPort

Changelog

What we changed on this site and in our own gateway: dates, links to the thing that changed, and the numbers we measured. Cosmetic work does not make it here.

Only what is visible from the outside or backed by a measurement. Repainting a button is not an update.

Entries: 7
  1. new section

    Case studies: five tools, measured, one of them a rollback

    This domain had no evidence on it at all. It now has the same five cases the Russian side carries, with the same numbers — visits across three thirty-day windows — the same counting method, and the same fifth card about the release we rolled back after reading the complaints. The numbers come from one shared source rather than being retyped, so the two languages cannot quietly disagree about what was measured.

    Read the cases
  2. new section

    Field notes: three of our own failures, each with a tool

    This domain had no reading section at all. It now has one, and it is not a translation of the Russian side: same three incidents, written for someone integrating with a service rather than hiring one. A 200 response carrying a refusal, a suite that was green at 298 checks and covered one class of path out of four, and our own error endpoint that accepted text from anyone and forwarded it toward a person. Every note ships a tool that runs entirely in your browser: nothing you paste is uploaded, and none of them fetch a URL.

    Read the notes
  3. measurement

    Compared two vision profiles and published the inconvenient half

    We took an invoice with exactly one arithmetic error planted in it and ran it four times through each profile. On the English invoice both profiles found the error four times out of four — the expensive one bought nothing. On the Russian twin of the same task the cheap profile found it once out of four. Both numbers are printed on the demo page: the profile is chosen by measuring your own data, not by reading somebody's table.

    Run the comparison yourself
  4. new section

    Opened a capability showcase with live demos

    Five task profiles — text, text extraction from an image, image understanding, heavy understanding and speech synthesis — can now be tried by hand, with your own file or on our sample. Each capability says where it fails: a showcase without limits reads as advertising.

    Open the showcase
  5. gateway demo

    The answer stopped looking like a draft

    Answers come back with formatting — lists, tables, headings — and until now that formatting reached the screen raw: asterisks, hashes and the pipe characters of a table were left for the visitor to decode. Lists now look like lists, tables scroll on a phone, and the formula notation that sometimes comes back is turned into ordinary signs. Prices in dollars are left alone.

    See it on a live answer
  6. home page

    The status line stopped promising a freshness it did not have

    The line "checked less than a minute ago" was baked into the page as it was served, so on a tab left open for an hour it was simply untrue — the browser polls nothing. It now shows the time of the check and says outright that the page does not repeat it. While we were there, every button and link in the header, the footer and the cookie bar was raised to a size a thumb can hit.

    See it on the home page
  7. home page

    The gateway can now be tried on the home page

    Until this day the gateway could only be read about on this domain. The widget existed and its English wording was written, but it was never placed on the page — so nobody saw it. Now your prompt goes through the same entry point our own products use, and the answer comes back onto the page, with a plain note about where the text goes and how long the answer stays in the request log.

    Try it