Skip to content
Hussh
Connect MCP

Release Notes 2026-08-08

The removal switch had never removed anybody; the faculty pages gained a real removal channel and an honest denominator; the Stanford request went to counsel.

Wiki page

TL;DR: Preparing a letter to a university meant checking a published promise against the code, and the mechanism that honours a removal without waiting for an engineer turned out never to have worked for anybody; that is fixed, verified in production, and the four claims of our own it disproved are corrected.

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

Relations

What shipped

Verified in production on 2026-08-07 at 22:54 UTC. The sitemap carries 3,518 URLs and every one was resubmitted to the IndexNow engines.

The removal switch had never removed anybody

There are two ways a removal is honoured here. One is a change to source code, which needs an engineer and a deploy. The other is a configuration change, and that is the one that makes the published 72 hours possible on a Friday night.

The second one did not work. Not for one corpus, not intermittently. For every listing in the directory, from the day it shipped.

Three faults in a row:

  • The part that reads the setting changed the spelling of the corpus name, and every part that asks used the original spelling. The two halves could never agree.
  • The example in our own documentation used a kind of identifier that no part of the system asks with, so an operator following the documentation typed something matching nothing.
  • For the largest set of listings the identifier contains commas, which is the character used to separate one removal from the next. Those requests were torn into fragments on the way in, and the first fragment looked like a perfectly valid request for a person who does not exist.

Every piece had tests and every piece passed, because the tests reproduced the parser's own spelling and never asked a reader anything. The tests now start from what a person actually types.

The faculty pages have a removal channel that works

The sentence offering removal was always there. The only button under it said "Nominate a professor" and opened a page naming no removal channel, so anybody wanting off had to guess. There is now a direct address beside it and a line saying what happens when you use it: no account, no form, nobody asking you to reconsider, and we will not put you back.

Underneath, the pages run the same removal list as the directory, keyed per university so two people who share a name cannot remove each other.

One limitation stated plainly rather than glossed: these pages are statically generated, so a faculty removal lands on the next deploy rather than the next request. The window holds because somebody ships, not because a running process re-reads anything.

A number with no denominator

Each university page showed a count and left the reader to guess what it was a count of. At Stanford it is 61, roughly 2.6 percent of the faculty. The page now says it is a selection and not a roster.

It still does not print a percentage, for the honest reason: we hold no verified faculty count for any of the 51 universities, and a page whose whole argument is that every number on it is cited cannot start publishing a denominator it cannot source.

The university request went to counsel

A draft letter asking Stanford for lawful access to faculty data, and an eleven-question memo to outside counsel, are both written and marked not sent. The letter names no recipient, because working out who owns that decision at a university is a task and not a guess.

Writing it found four things wrong on our own side, which are in the memo rather than buried: a claim that our source code is public, which is not true of the repository that matters; a claim that no member of the public is ever charged, which is broader than what we can defend; the removal gap above; and the missing denominator.

A Palo Alto front door

The new local page opens with what we have and what we do not, rather than presenting a partial index as a complete one. California contractors gained a verification door, because that state licenses every one of them through a public register. Most states still show no button, because a plausible wrong door is worse than none.

A Friday meeting that assembles itself

The all-hands framework, the rota that now advances on its own rather than naming the same person forever, and a chief-of-staff agent that runs the meeting, keeps the ledger of what people promised out loud, and produces the weekly status. Published alongside a thank-you to Gokul Rajaram, whose public write-up we took most of unchanged.

What is still not measured

The all-hands pack prints "not measured yet" for removal requests received and honoured, every week. That line is what sent somebody to check whether the mechanism behind it worked. It is still the loudest gap we have: the fix above makes removal real, and we publish no public count that would let anyone outside the company check it.

How this was verified

An independent smoke run against production covering 25 routes by content marker rather than status code, plus 7 gated routes confirmed still closed. The three new or changed pages returned 200 with their new on-page text present. The full gate ran green on the exact commit: guards, typecheck, 2,292 tests, production build, and the backend suite.

Sources