Release Notes 2026-08-16

TL;DR: The search interface answered plain-English questions from the wrong lane with a clean success code; fixed, verified, the scorecard caught the follow-on bug in twenty minutes, four untrue sentences are gone, and the mortgage lane is ready for the NEXA roster.

Status as of 2026-08-16: see body.

Relations

TL;DR: The search interface ignored the one parameter our own finder page sends, so a plain-English question came back as twenty financial advisers with a clean success code; that is fixed and verified in production, the identical defect one field over was caught by our own scorecard twenty minutes later, four sentences on the site that were not true are gone, and the mortgage lane now exists ahead of a partner roster of roughly 3,700 loan officers.

Status as of 2026-08-16: all five releases verified in production; the public scorecard reads 100 percent with seven more checks than it had when it last read 100.

What shipped

Five releases, each through the full gate: CI on the exact commit, UAT verified by content marker, production promote, independent smoke run, IndexNow resubmission.

The question you type, answered in the right lane

Our finder page has always sent your question as a parameter called q. The search interface behind it ignored that parameter entirely and answered from a default lane. Asked for help with a flooded basement in Kirkland, it returned Hot Creek Capital and Moss Bay Capital. A plumber in Stamford returned Abs Global Investments and Coinbase Advisors. A therapist in Manhattan returned a capital management firm. Every one was a clean success as far as the machine was concerned.

The page had been reading those same sentences correctly the whole time, with a parser sitting one file away that the interface never called. It calls it now, and the same sentence gets the same answer whichever way it arrives.

The worst case was not the plumber. The page has always recognised somebody describing domestic violence and put the National Domestic Violence Hotline in front of them before searching anything. The interface returned financial advisers. A safety response that exists for a person reading a page and not for software acting on their behalf is not a safety response. The hotline, its around-the-clock staffing, and the promise that nobody gets called without your say-so now travel together wherever the question is asked.

The same defect again, one field over, caught in twenty minutes

The fix above taught the interface to read the profession and the speciality out of the sentence, and left behind the ZIP code. So a question that names exactly where it is, "my kid needs a doctor near 98033", was refused for not saying where it is. The identical mistake, by the same hand, a day later.

The public scorecard caught it against the live site twenty minutes after the first fix shipped, because the same release had taught the scorecard to ask its questions of the interface as well as the page. It read 96.7 percent, the one red line was this, and the fix followed within hours. That is the instrument working, and it is the only reason the mistake cost an afternoon rather than a week.

Four sentences that were not true

Found by checking public claims against the code behind them, and corrected rather than softened:

The site asked itself whether Agent One works with ChatGPT, Gemini, Claude, Siri, Grok, Meta AI and Copilot, and answered yes. There is no finished connection to any of the seven. The page now says not yet, names what exists, and commits to naming the first real listing with a date.
The Apple partnership proposal said we send heavier work to Apple's Private Cloud Compute, in the present tense. No outside company can run a workload there. Now conditional, with the constraint stated on the card itself.
The machine-readable agent profile listed two payment and commerce standards beside two protocols we genuinely run, with nothing marking which was which. Each entry is now labelled, and a consumer is told to treat the intended ones as absent.
The developer page advertised three permission names that appear in no registry we publish. It now describes what an agent actually does, because the running server is the authority, not a marketing page.

The mortgage lane, built ahead of its inventory

The founder is in a live conversation with NEXA Mortgage about bringing roughly 3,700 loan officers onto the network. Before this release, the product could not hear the word mortgage: the question returned "I'm not sure what you need yet."

Now the sentences parse, and because we hold no mortgage inventory, the honest answer is the register every loan originator in America is actually in: NMLS Consumer Access. The release also closed the trap underneath, where a mortgage question would otherwise have returned the adjacent finance corpus of banks and appraisers with confidence.

The roster itself arrives through the front door or not at all. The standing rule is that a pending request plus a quiet scrape is worse than either alone, and the request is pending. The onboarding ask is written, the ingest path exists, and the estimate from file-in-hand to live is days.

The release-notes guard fired, twice, and was right both times

A guard added the week before refuses to promote a release to production while the public release record is stale beyond tolerance. It fired on two of these releases and stopped them until the record caught up. Without it, the largest changes of the period would have reached production while the page claiming to be the public record said nothing about them.

How this was verified

Every claim above about a live surface was checked by fetching that surface after the promote: the corrected pages by their new on-page text, the interface by asking it the actual sentences and reading the lane it answered in, the mortgage question by confirming empty results plus the NMLS door rather than a confident wrong list. The scorecard runs against production and reads 100 percent, now at 29 checks against the 22 it had the last time it read 100, and the difference is the point: the added checks are the ones that would have caught this batch's defects earlier.

Sources

hushh.ai releases
The finder
Works with, corrected
Agent profile, corrected
Verified in production 2026-08-16 UTC by independent smoke runs and direct fetches of every corrected surface.