A model with no attention admits

Yes — and it is witnessed, not promised. mamba2-780m, 48 SSM layers and zero attention anywhere, was first refused with nineteen itemised findings, then admitted the day graph schema 6 landed: a registered Mamba2 operator judgment consumes the SSM semantics, the container encodes with no attention surface and no FFN surface — the mixer is the whole block — and ordinary LQL opens it with the source checkpoint deleted. And it generates: the generic executor paritied against the fp32 reference — argmax exact at all 430 teacher-forced positions, trajectories token-for-token — and ordinary INFER … GENERATE reproduces the reference continuation from the artifact alone.

PASSED · 19 blocking → 0 · closure at encode · generates through ordinary LQL under the deletion invariant, paritied to the fp32 oracle · 2026-08-30