Avi Fenesh

Don't have that cell on the card - accept there is agentic only, same as you read (0.755 to 0.730). Chat accepted/drafted trimmed and untrimmed isn't published and I haven't pulled those two from the run logs yet. Won't guess whether chat sits under your 0.28 ceiling. If I get a quiet window I'll dig the pair out on the same accepted-over-drafted definition and post them.

I haven't sliced the agentic set by prompt tokens yet, so I don't have that number. Your reading of the K curve feels right to me - chat landing on the asymptote and agentic as the thing that needs an explanation. If I get a quiet window I'll bucket the agentic steps by context and see whether the short bucket climbs toward that ceiling.

it's the trunk head on the draft path, not the selector codebooks.

MEMRA_FRSPEC_TRIM is the same FR-Spec ranks mask as on mtp: load-time, 32768 rows from the ranks txt. on dflash the drafter has no lm_head of its own, so the thing being masked is the head the draft path borrows when it scores draft positions (target/trunk head). verify stays on the full vocab either way. the mask only limits which tokens draft is allowed to propose.

the boot line saying "DFlash2 draft head TRIMMED" is that borrowed head, which is why it fires for a file that doesn't ship one. selector_top_k / the codebooks are a different cut. FRSPEC_TRIM does not touch those.