fix: referent channel (each unit now a T³) + QK gate + magnetic softmax, trained under load K∈{0..3}
tier: lit · W1 grad 2.0e-08 · W4 constructive K=2: 0.823 · determinism 0.0 · TTF-1 task-identifiability correction filed · 2026-07-09
BEFORE / AFTER — DISTRACTOR & DEPTH CLIFFS
dashed red = TTF-1, solid teal = TTF-2 · left: distractor load (K* moved 1→6, at K=1: 0.34→0.94) ·
right: depth truncation (zero-margin weld → 10-layer margin) · baseline itself rose 0.79→0.95
3D — THE HEALED TUG-OF-WAR (K=1)
same specimen class that broke TTF-1: unit 0 now ignores the
marked distractor (wrong referent) and locks to the
true cue — competition + referent match, watched working
NEW AXIS — REFERENT COLLISION: THE BINDING RESOLUTION
distractor referent approaches the query; accuracy holds 0.87+ down to Δ=0.15 rad (8.6°) — unbroken across the sweep ·
floor caveat: referent noise σ=0.05 rad, so the smallest tested gap is still ~3σ; true resolution limit lies below the grid
F−1 DELTA TABLE — ALL AXES
every axis improved except cue blur — unchanged by design: that fracture is classifier geometry (45° class spacing), not retrieval; the correct negative control
🖍 TODDLER CORNER
Last time our robot face-planted when there were two secrets in the room. Today we fixed it, two ways at once:
1. We gave every secret a name tag (that’s the “referent”) and told the robot WHOSE secret we want. 2. We made the secrets compete: instead of grabbing everything marked, the robot now holds a little contest and only the best match wins. (Real robots call this contest “softmax.” We built it out of magnets.)
Now: two secrets? Easy. Four? Fine. It takes six competing secrets before it stumbles. And even when two name tags are nearly identical — tilted apart by less than a clock-minute — it still picks the right one.
Also nice: the robot got less brittle everywhere else too. Learning to handle a messy room made it calmer about missing rings and noisy tags. Practice with chaos builds sturdy robots.
One thing did NOT change: if you smudge the feeling itself past a certain point, no amount of clever finding helps — the feelings just look alike. That test staying broken is how we know our fix fixed the right thing. A good repair leaves the other cracks honest. lol