Empryo edits your files and keeps working while you are not watching. That asks for more trust than a green badge can earn. I build it alone, and I have to keep three operating systems and three ways of running it all true at once, which is more than one person can hold in their head. So 290 hostile checks do the asking. 211 of them run on macOS, Linux and Windows alike. On Linux and Windows they run the compiled release bundle, built exactly as the one you download, never a developer checkout.
They interrupt a running answer and ask something else. They rate-limit the model and kill the provider mid tool-loop. They hand it a filename with a space and an accent, a file marked read-only, a project with no files at all. Whatever breaks becomes a check that runs on every patrol from then on. And because a check can only re-ask a question somebody already asked, other cells are built to work the opposite problem: going looking for the bugs nobody thought to test for.
Each cluster is one of the promises below: its size is how many checks guard it, its brightness how deeply. The dashed edge is the frontier, where guarded ground ends.
The dark outside is real. 96 of the 374 things this product can do have a check aimed at them today. The count comes from the product's own registries, so a new feature shows up as a gap the day it lands. The blue cells spend their lives out there.
On your machine, first try — not just on ours.
how we attack itOn Linux we install the shipped build from the web and make it prove every piece of itself came along; on macOS we run the bundled installer into a throwaway home.
Everything above measures the engine. This measures you: the eight things a person does, from installing it to giving it up, walked end to end on a machine with nothing on it. A step only counts when a check starts from a machine with no Empryo and no config on it, installs from a signed origin, and drives what happens next. Never from a machine I arranged first.
3 of 8 today. It is published on its own and averaged into nothing. Until a step has been walked end to end I would rather show the low number than let a bigger one speak for it.
Six kinds of cell run this system. Meet them once here; their silhouettes turn up everywhere else on this page.
One promise written down: what to exercise, and what must hold once it runs.
Takes one seed to one operating system and one surface, and on Linux and Windows runs it against the build you would download.
Fires on every run without being asked. Crash text, garbled screens, leaked escape codes: caught even when no seed was looking for them.
Drives the product the way a person would, hunting for what is missing rather than what is broken.
Attacks every new finding before anyone believes it. One pass came back 18 refuted of 23. What it refutes never becomes a false alarm.
Born from a confirmed defect, it recognises that bug on sight, forever.
seed killer T cell memory cell
Every seed asks a question somebody already thought to ask, against the real build, on every platform that can honestly run it. A red result has to survive a hand reproduction before anyone calls it a defect. This loop keeps the promises already made.
dendritic cell negative selection seed
A check cannot notice an absence. The desktop app shipped with no right-click menu in any editable field. Paste by mouse was impossible there, and every test stayed green, because the assertion did not exist. A user found it immediately. Dendritic cells exist for that class of bug: they drive the product like a person, note what should be there and is not, and hand every claim to negative selection to be attacked. What survives becomes a seed.
Both loops end in the same place: a check that runs on every patrol, forever.
A red check is not a bug yet. This is the path from a suspicion to a scar.
latest pass · Aug 16, 202652 defects claimed49 confirmed1 refuted
Everything above says what is watched. This says what happened the last time it ran. A full patrol: unfiltered, past the smoke tier, more than one machine.
Across 3 adversarial passes, 105 defects were claimed and 84 survived two hostile reviewers. 19 were refuted and 2 disputed. Of the confirmed ones, 1 still carry an open marker: a confirmed defect that is not fixed keeps a check that reports known rather than passing, so it cannot hide in a green patrol.
Linux and Windows run the release archive you download, never a developer checkout. A filled dot means the check drives it: a real terminal read back character by character, a real window clicked. A half dot means it only gets watched as it starts. When a surface cannot be driven honestly on a platform, the report says so instead of claiming a pass.
