TEST TO FAILURE — 1 · THE TRAINED SILO UNDER LOAD

LOAD–DISPLACEMENT CURVES · FIVE AXES · ✗ = FRACTURE · ⚑ = F−1

two cliffs, three slopes: distractors and depth truncation fail catastrophically; marker strength, type noise, and cue blur degrade gracefully · x normalized to grid index per axis (loads labeled at fracture)

3D — THE TUG-OF-WAR SPECIMEN: K=1, TWO MARKED CUES

teal = unit 0 (readout) · red = true cue · violet = marked distractor · the gate passes BOTH marked units; their torques average and unit 0 settles between the two phases — binding failure, watched happening

🖍 TODDLER CORNER

We squeezed our magnet robot five different ways to find where it breaks.

1. Whisper the secret mark more quietly? It’s fine until about half-volume. Sturdy!
2. Make everyone else a little noisy? Fine for a while. Sturdy!
3. Smudge the feeling itself? Fine until the smudge is almost as big as the gap between feelings. Sensible!
4. But put TWO secret marks in the room? Instant faceplant. The robot grabs both feelings at once and ends up holding a mushy average of them. It can find marked things — it cannot choose between marked things. (Fun fact: the two big robots from before have the exact same problem. All three robots trip over the same toy.)
5. And if we remove even a few of its 60 finding-rings? Faceplant again — not because it needs all 60 to find things, but because the second half of the robot only understands the first half at exactly ring 60. They practiced together so much they can’t improvise. Best friends, zero flexibility. lol

The big lesson: practicing together made the team great AND fragile. And nobody — not the big robots, not ours — has learned to share a room with two secrets yet. That’s what the “softmax” in real attention is for: it makes the secrets compete until one wins.