⊞
IX ∞ SELF-HEALING EYE
THE MACHINE THAT WATCHES ITS OWN WATCHERS
EXPERIMENT 03 / ACTIVE PERCEPTION

Break the image.
See what notices.

This thing has nine eyes. Other eyes watch those eyes. The higher eyes can even catch the lower ones lying.

START WITH
01 / THE THING BEING OBSERVED

A world waiting to be seen

SLEEPING
◈NOTHING TO REPAIR YET.CHOOSE A SUBJECT ABOVE
INPUT → PERCEPTIONERROR: --
NO IMAGE LOADED0 REPAIRS
↖ OR TAP THE IMAGE TO BREAK EXACTLY WHERE YOU TOUCH
Real image source will appear here.SOURCE ↗
02 / THE RECURSIVE WATCHERS

Who watches the watcher?

3 LEVELS
Every illuminated cell has an actual piece of image evidence behind its decision. Tap any cell to investigate.
LEVEL 019 PIXEL EYES0 / 9 CHECKED
↓ EVALUATE THEIR REPORTS ↓
LEVEL 023 ROW AUDITORSAWAITING DATA
↓ COMPARE WITH INDEPENDENT PIXELS ↓
LEVEL 03META-OBSERVERAWAITING DATA
◎
NO EVIDENCE YETThis observer checks whether its observers are trustworthy.
—
TAP A CELL FOR EVIDENCE

Each one measures a different part of the picture. The meta-observer uses its own pixel samples to check their claims.

03 / EVIDENCE TRAIL

What changed, and why?

00

Pick an image. Then try to break it.

04 / THE IMPORTANT CONTROLS

Try to outsmart it.

Automatic repairWatch the observers make real changes.
Independent meta-checkTurn it OFF, poison an eye, and watch what changes.
Show the nine regionsSee the geometry behind the decisions.
These are genuine pixel tests and programmed feedback loops, not conscious observers or LLM agents. More layers mean more independent verification, not more artificial minds.
THE ENGINEER EXPLAINS IT TO A RANDOM PEDESTRIAN 😭

Watch it catch itself messing up.

Get an image, press BREAK SOMETHING, and a chunk visibly corrupts. The nine eyes measure actual pixel differences. Three supervisors double-check their reports. One more observer checks those supervisors. New layers can independently check that observer too.

Try FOOL AN EYE, which makes one sensor report everything is fine. If independent observation is enabled, a higher layer catches that discrepancy, distrusts the sensor, and repairs the broken image anyway.

Turn the independent check off, and you can see where the system becomes blind. That's the experiment: not just having observers, but testing whether their observations match reality.

This doesn't understand images like a multimodal AI model; it compares pixels against a known reference. Search results come from Wikimedia Commons. Everything else runs locally in your browser.