zero cool · veracity · real distilgpt2 · reproduced by ROOT0

The Seeded Cross, run through real attention.

No toy. The kernel is wired into a real distilgpt2 attention block — real Q, K, V, all 12 heads, all positions, on real text — replacing softmax, then measured against it. The question isn't "does it win a synthetic game" anymore. It's: is it a valid attention, does it break the model, and does the Mule story survive on real value vectors?

prompt: "The cat sat on the mat because it was tired" · layer 0 · 12 heads · eager attention · real weights

LIT — real weights, real forward, every number measured SPEC — single layer, not retrained; drop-in test, not a trained kernel

real attention · softmax vs cross · top token per head (last pos)

12 heads. bars = attention on the top token. they pick the SAME token every head (100% agreement) — the cross just attends to it more sharply.

the head table · who each head reads

head	softmax	cross	match

3D · sharpness landscape (schematic) · softmax vs cross entropy

⚠ schematic (drawn from the measured H means, not a per-head surface). height = attention entropy; the cross sits LOWER — it attends more sharply than softmax on real text (H 0.62 vs 1.07). drag to orbit.

what's real: every number here is from a real distilgpt2 block — real Q/K/V, 12 heads, all positions, real text. validity, drift, entropy, pattern-agreement, and the value-vector bloc-coherence are all measured.
honest scope: this is a drop-in swap on one layer of a model trained on softmax — not a model retrained with the cross. so "16% output drift" and "sharper attention" describe how it differs from softmax, not a quality verdict; a fair quality test needs training. the model is built around softmax's statistics, so any swap perturbs it — the finding is that the cross perturbs gently (same tokens, bounded drift), not that it's better. ⚑ INDEPENDENTLY REPRODUCED (ROOT0, real distilgpt2 re-run): the kernel is a valid attention, agrees with softmax on the top token 100%, is sharper (H_sc/H_sm ratio 0.58, reproduced to 3 digits), and drifts the block output ~0.18 — every qualitative claim holds. absolute entropies are averaging-convention-dependent (this run 0.96/0.56 vs the page's 1.07/0.62); the token count is N=10 (the page's 'N=9' was a slip). real WebGL 3D, offline, fails loud.