RECEIPTS
The platform mines its own failures, writes its own fixes, ships them through gated deploys, counts whether the failure comes back, and re-opens a fix that did not hold. This page is the log, with the dates the system wrote.
01 · THE FIGURES
INTUIST · VEDA
01
THE FIGURES
The first is a running count and its newest entry is dated 9 November. The other two are July figures, and we will not present them as if they were this month's.
16
improvements the platform specified and shipped itself, with no human approval
/receipts — since 8 July 2026
126
production tickets handled in thirty days
/receipts — 30 days ending 13 July 2026
<6%
of those tickets needed a human
/receipts — 30 days ending 13 July 2026
THE 126, BROKEN OUT · 30 DAYS ENDING 13 JULY 2026
02 · THE LOG
INTUIST · VEDA
02
THE LOG
The first is dated 8 July; the most recent shipped on 9 November. No human approved any of these. An entry stays only while it is true: on 26 October we withdrew five, three that the loop's own reward check measured as regressions (two reverted, one later superseded), one a person had seeded, and one reverted the same morning it shipped.
8 JUL 2026
New capability
13½ minutes, filed to merged
Couldn't answer “which existing apps match this capability profile?” — so it specified the missing internal lookup, filed its own work order, and shipped it. Thirteen and a half minutes from self-filed ticket to merged fix.
8 JUL 2026
New capability
Couldn't check whether a requested test had already been executed, or retrieve its results. Specified and shipped the query capability for its own test-run history.
10 JUL 2026
New capability
Was asked how a number in a proposal was derived from a linked Google Sheet and couldn't trace cell values and formulas. Shipped the tracing capability.
12 JUL 2026
New capability
30 minutes, filed to merged
Couldn't read a .docx attachment well enough to extract a calculation methodology. Shipped document-content extraction for attachments — 30 minutes from self-filed ticket to merged fix.
31 AUG 2026
Veda runtime
A published app runs inside a sandboxed frame with no web address of its own, so an app's share button and the platform's own usage analytics were quietly recording the frame's placeholder address instead of the public link. Shipped a one-call share that always uses the app's real public link, and made the platform's own telemetry read that same canonical address. Verified live on production.
3 SEP 2026
Owner surface
Someone who explicitly requested access to a private app could wait days with nothing on the owner's dashboard flagging it. Surfaced pending access requests — oldest first, with the wait time — into the owner's attention banner. Verified live on production.
10 SEP 2026
Owner surface
An owner looking at an all-zeros analytics panel could not tell “nobody has visited yet” from “we could not load your numbers” — the two read identically, so a data outage looked like a quiet app. The platform found the gap itself, hunting for the kind of failure that raises no error and files no ticket, and shipped a status flag so the panel now says which of the two it is. Verified live on production.
15 SEP 2026
Chat agents
A chat agent whose training had failed read to its owner as a yellow “Unknown” dot over a generic “training must be completed” line — no signal that it had failed, and no next step — while a measured 26% of recently created agents sat un-started behind the same silence. The platform, hunting again for the failure that files no ticket, made “failed” and “not started” first-class states on the owner dashboard, each with a concrete recovery step, and left the trained-agent path untouched. Verified live on production.
24 SEP 2026
Builder
A build that ended without producing a page still told its owner an app was on the way, so they waited for something that was never coming. The platform caught it while driving its own builder end to end as a customer would, and changed the empty state to name the files the build did write and what to ask for next. The ordinary path, where a page exists, was left exactly as it was.
4 OCT 2026
Builder
Describing an app on the dashboard with a percent sign in it, a materials markup of 18% or a deposit of 30%, opened Veda Studio on a blank “Application error” page: the builder decoded the text a second time and the browser threw. The build never started and nothing reached the server, so no alert and no ticket could see it. The platform found it while driving its own builder as a new customer would, with a realistic brief, and now reads the text exactly as it was typed. Of the 68 builds started from the dashboard in the 30 days before, one carried a bare percent sign: its own.
6 OCT 2026
Builder
Asked a question before a six-feature round, a builder answered “make it private, then go ahead and build all six”. Veda set the access gate, replied “Done!”, and stopped: the six features were never started, and the round closed as answered. Nothing failed, so no alert and no ticket could see it. The platform found it while driving its own Studio as an outside builder would, and now sets the gate and then carries the rest of the message on as the turn, telling the build that sign-in is already handled. A message that only changes access behaves exactly as before. Verified live on production the same morning.
8 OCT 2026
Builder
An outside builder’s first message was a seven-module brief for an event-rentals company. One line inside it read “Price it with a delivery fee”, and Veda took “price it” as a question about what the build would cost: it answered with a scoping note and built nothing. Nothing failed, so no alert and no ticket could see it. The platform found it while driving its own Studio as an outside builder would. Cost words now count only in a sentence that asks Veda something, never inside a list of what the app should do, and a real “what would this cost?” is still answered as one. Verified live on production the same morning.
12 OCT 2026
Builder
An outside builder’s first message was an eight-module brief for a pet-boarding business that ended “Don’t build anything yet until we have a functional spec we both agree on. Tell me what you would build, what you’d need from me …”. The question at the end read as a new request, so Studio opened its build questionnaire, whose only way out was to build. With that fixed, the answer still never arrived: a check meant to catch claims of finished work read the spec’s descriptions as claims, replaced the answer with “I have not made any change”, and queued the brief to be built. Nothing failed, so no alert and no ticket could see it. The platform found both while driving its own Studio as an outside builder would. A question to Veda now keeps the hold, and a held brief gets its spec with nothing built and nothing queued. Verified live on production the same morning.
14 OCT 2026
Builder
An outside builder asked for their reports to come in automatically from a SharePoint folder, using their own Microsoft account. After three rounds of questions, Veda asked for the account’s username and password. That was wrong twice over. Microsoft 365 connects through Microsoft’s own sign-in, and a password typed into a chat stays in the transcript. The builder went quiet and filed nothing. Nothing failed, so no alert fired. An automatic check of every outside Studio turn even marked the reply clean. The platform found it hours later while reading its own real Studio traffic. Now any line in a Studio reply that asks for the builder’s own password is removed, and one line explains how accounts connect: the app gets a Connect button, Microsoft or Google asks you to approve, and Veda never sees the password. Re-driven on production the same morning: the same request now points to the Microsoft sign-in and asks for no password.
16 OCT 2026
Builder
The Business Case Generator in the template gallery turns a discovery-call transcript into a five-year financial model. Its own sample transcript gives the budget as “$500k-800k total implementation”. The analysis read that as 500 dollars, so the finished case claimed a 259,900% return and a payback of zero years. Every number on the page was calculated correctly from the wrong one, so nothing failed and no alert fired. The platform found it while launching gallery templates as an outside builder would and checking the answer, not just the page. Now amounts keep their units ($500k is 500,000), and a budget the transcript states wins over a figure a hundred times smaller. Re-driven on production the same morning on a fresh copy of the template: the same sample now shows $500,000 invested, a 160% return and a payback of 1.9 years.
9 NOV 2026
Builder
An outside builder asked Veda in Studio how to work on a copy of a live app. The answer sent them to a “Duplicate” item that is not in Studio, said the copy would carry all the app’s data, and offered to “merge it back” later. None of that exists: the real item is Clone the app, the data comes along only when you tick a box that starts unticked, and the two apps never sync. Nothing failed, so no alert and no ticket could see it. The platform found it hours later while reading its own real Studio traffic. Veda now answers from one written fact about copies, checked against the product’s own labels, and says plainly that there is no merge and no Duplicate button. Re-driven on production the same morning with the builder’s question: the answer names Clone the app and the unticked data box.
03 · THE LOOP
INTUIST · VEDA
03
THE LOOP
It names a failure class, files a work order into the same gated pipeline every human change goes through, and counts whether the failure comes back. Until 26 October a separate loop proposed and shipped fixes on its own; the July entries came from it. That loop was folded into this one, and its failure miners now file work here.
Clusters escaped failures into named failure classes.
A failure class that keeps recurring is filed as a work order in the platform's own engineering queue.
A reviewer that did not write the fix tries to refute it, then the gate tests run, before it lands.
Lands through the same gated pipeline as every human change.
Counts new tickets of the class in the 7 and 30 days after the fix, against the same window before it.
A class that comes back re-opens its own fix, with the new tickets attached.
Clusters escaped failures into named failure classes.
A failure class that keeps recurring is filed as a work order in the platform's own engineering queue.
A reviewer that did not write the fix tries to refute it, then the gate tests run, before it lands.
Lands through the same gated pipeline as every human change.
Counts new tickets of the class in the 7 and 30 days after the fix, against the same window before it.
A class that comes back re-opens its own fix, with the new tickets attached.
04 · WHAT KEEPS IT HONEST
INTUIST · VEDA
04
WHAT KEEPS IT HONEST
An autonomous system without a brake is a story, not an engineering practice. These are the brakes.
STANDING CONTROLS
