argo/memory/findings/attention-head-7-3.md
FINDINGUSER OVERLAYCONF · 0.92
---
kind: finding
subject: attention-head/7-3
model: argo-37b
epoch: ep18
evidence_count: 12
promoted_from: claim-low-rank-reinit
status: validated
---

Head 7.3 collapses to rank-1 after epoch 12

Spectral analysis of head 7.3's QK matrix shows the dominant singular value carries 94% of total norm by epoch 12, and stays there through epoch 18. The other 63 singular values flatten into the floor.

This is the direction that explains the 2.1 % FLOPS reduction we saw when we pruned 7.3 in the @argo-37b-prune-batch-3 run.

Pinned by user · 2 days ago. Why this matters: pruning quorum says we need three independent signals before we kill a head. This finding is signal one.

What the data shows

Head 7.3 starts as a normal multi-head with full-rank QK. By epoch 12 it has collapsed to rank-1; by epoch 18 the second singular value is below 0.04 in normalized magnetic. The collapse is gradual and monotonic — not a sudden phase change.

Three replications

The collapse is reproducible across seeds. We re-ran the experiment with seeds 1, 7, and 19. All three runs collapsed the same head by the same epoch within ±1.

Why it might matter

If head 7.3 is rank-1, it's not learning. It's a free pass to prune. The pruning quorum policy at @policies/pruning-quorum.md says we need three signals; this is one.

memory · readyln 1 · col 1spectral · md
unsaved patch · ⌘S to saveargo branch · main