RECEIPTS
The platform mines its own failures, writes its own fixes, ships them through gated deploys, measures the result, and reverts what regresses. This page is the log, with the dates the system wrote.
INTUIST · VEDA
01
THE FIGURES
01 · THE FIGURES
The first is a running count and its newest entry is dated 12 October. The other two are July figures, and we will not present them as if they were this month's.
18
improvements the platform specified and shipped itself, with no human approval
/receipts — since 5 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
INTUIST · VEDA
02
THE LOG
02 · THE LOG
Four landed in a single ninety-two-minute window on the evening of 5 July; the most recent shipped on 12 October. No human approved any of these; a human audits them afterwards.
5 JUL 2026
Builder
Detected that users sometimes ask for the same change twice because the previous build didn't actually apply it (zero diff) or the change wasn't visible in the preview. Shipped a gate that verifies a build changed what the user asked for before declaring it done.
5 JUL 2026
Builder
Found its own retry loops: fix attempts that didn't change the code kept being retried. Shipped a breaker that requires a root-cause diagnosis before another attempt is allowed.
5 JUL 2026
Builder
Noticed rebuilds that produce no delta (the change was already present, or the feedback wasn't actionable) waste compute and erode trust. Shipped an abort gate for zero-diff refines.
5 JUL 2026
Support agent
Caught itself answering “this” and “it” on long email threads from the latest message only. Shipped grounding that resolves pronouns against the artifact the thread was originally about.
8 JUL 2026
New capability
14 minutes, filed to live
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. Fourteen minutes from self-filed ticket to healed.
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
33 minutes, filed to live
Couldn't read a .docx attachment well enough to extract a calculation methodology. Shipped document-content extraction for attachments — 33 minutes from self-filed ticket to healed.
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.
25 SEP 2026
Connections
Three connectors — GitHub among them — are authorised with a token you paste rather than a sign-in you click, and the Connect button answered them with an error and no way forward, in the builder and inside the app alike. Nothing failed on our side, so no alert and no ticket ever named it; the owner simply stopped. The platform found it while driving a freshly built app of its own, pressing every Connect button, and connected the credential form that had been written for those connectors all along and never reached. Every other connector keeps the road it had.
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.
INTUIST · VEDA
03
THE LOOP
03 · THE LOOP
It names a failure class, proposes a fix, and files a work order into the same gated pipeline every human change goes through.
Clusters 30 days of escaped failures into named failure classes. 39 classes so far.
Drafts a guardrail for each recurring class, grounded in how the real tickets were resolved.
Compile checks, a gold test-suite floor and smoke tests, before traffic flips.
Files its own work order through the standard pipeline. One ship in flight at a time.
Recurrence of the failure class in equal windows before and after the ship.
A guardrail that regresses its own metric is reverted, and its class is marked human-only.
Clusters 30 days of escaped failures into named failure classes. 39 classes so far.
Drafts a guardrail for each recurring class, grounded in how the real tickets were resolved.
Compile checks, a gold test-suite floor and smoke tests, before traffic flips.
Files its own work order through the standard pipeline. One ship in flight at a time.
Recurrence of the failure class in equal windows before and after the ship.
A guardrail that regresses its own metric is reverted, and its class is marked human-only.
INTUIST · VEDA
04
WHAT KEEPS IT HONEST
04 · WHAT KEEPS IT HONEST
An autonomous system without a brake is a story, not an engineering practice. These are the brakes.
STANDING CONTROLS
Intuist is raising a $3M seed. Investors: write to [email protected].
