Watch a floor refuse.
A real Atari 2600 ROM, read as raw 210×160 pixels. No weights. No gradients. No training run. The pilot proposes an action; a memory of past failures removes some actions from the admissible set before the pilot ever gets to choose. Every refusal names the evidence that produced it — how many times this exact situation was seen, how many times it ended in a death, and how that compares to the base rate.
Recorded run. Press next crash to jump to the moments where the floor learns.
what it just learned
explicit cards·no weights·every rule cites its evidence
- none yet — a rule needs enough evidence, and must name the hazard it answers
How to read it
The strip under the video is the decision pipeline, one stage at a time: pixels → facts → admissible → objective → projection → memory → action. The interesting frames are the ones where memory lights up — that is the floor overriding the pilot’s preferred move.
The panel on the right is not a visualisation of a hidden state. It is the state. Each card is a literal row: a bucketed situation, the action taken, and what happened. When a situation accumulates enough evidence, it becomes a rule that removes that action from the admissible set. Delete the row and the behaviour changes. There is nothing else in there.
This is the v1 mechanism, and it does not always help. On Space Invaders it is worth +32%. On Freeway the same mechanism costs 12%, because there the only scoring action is also the dangerous one. The result page has both numbers and the reason.