[bong-water-water-bong]

#805 (plan P6) lists one protocol item that was never built: "spare verify rows filled with the drafts' 2nd/3rd candidates". This PR settles it — it should not be built as specified, and the evidence is a test rather than an argument.

Why it cannot work here

The loop's acceptance rule is argmax(logits[i-1]) == draft[i] (npu/mtp/draft.h:317). A draft is accepted iff it is the base's own argmax at that position, and the argmax that grades depth i is produced by row i-1 of the same batch. Two consequences:

1. a candidate at the same position that is not the argmax is refused with certainty, whatever carries it — so 2nd/3rd candidates at a depth can only add rows that are certain to be rejected;
2. a correct candidate deeper in the chain is never scored while its parent is refused, and an alternative at a depth cannot even be placed in the same batch (it would be graded by its parent's row, which does not exist yet) — it needs another batch, i.e. exactly the pass the loop already pays to re-draft.

Measured — tests/mtp_draft_test.cpp [16], toy base given a real two-level ranking

rank-2 candidate at every depth ................ accepted 0/92
the same head, argmax moved onto the runner-up
  (the control that makes that zero mean something) accepted 20/20
right at depths 2..k, wrong at depth 1 ......... accepted 0, reached_at[1] == 0
the same head at k=8 (M=16 rows) ............... accepted 0 of 184 rows

The control is the point: the identical head against the identical base is accepted every time once the argmax moves, so 0/92 is a measurement of the acceptance rule, not a broken harness. And widening the batch to the full M=16 changes nothing, because the batch's rows are the chain, not candidate tests.

A real tree verifier (per-row positions + attention mask) would be a different design, and with a recurrent GDN state carried row by row it would also need a state per branch — the "no per-token state banks" constraint this same plan imposes. The two levers that did move this loop are already recorded: the hidden variant (post-final-norm) and the drafter's stream cursor.

Contents

• tests/mtp_draft_test.cpp — section [16] (8 checks) plus the two toy heads and the rank_two / swap_top_two knobs it needs.
• docs/npu.md — the determination next to the break-even paragraph, with the counts and the code line.
• tools/mtp_acceptance.cpp — the stale "batch it harder, npu plan P6: the shared drafter protocol - tau-gated chain, 4-row verify with leaf fill, deferred GDN commit; acceptance per target in the leaderboard #805's leaf fill" comment now says what is true.

Verified

./mtp_draft_test                                   -> all checks passed (existing checks + [16] + the control)
ctest -R 'mtp_draft|mtp_depth_pool|npu_leaderboard' -> 3/3 passed
tools/copyright.py --check (tree staged)            -> rc 0
g++-16 -std=c++26 -fsyntax-only tools/mtp_acceptance.cpp (CI's flags) -> rc 0

refs #805.

[bong-water-water-bong]

#1305 ready for review at 919dff8

[github-actions]

Preparing review...

[bong-water-water-bong]

VERDICT: READY at 919dff8475e65c0ac3a3f91f293e3b883e1d0a2e. The determination is correct for this verifier, and the test measures it with a control that makes the zero mean something.

The rule, read in the code: npu/mtp/draft.h:313-319: draft 0 is checked against the last step's logits, draft i>0 against row i-1 of the batch, if (argmax(lg, V) != d[i]) break;. It is a linear chain that stops at the first mismatch. The loop is greedy by construction (draft.h:44: emitted tokens identical to plain greedy decode), so a non-argmax candidate is rejected with certainty, and nothing below a refused parent is ever graded. Both consequences in the body follow directly.

Measured here: g++-16 -std=c++26 -O2 -fopenmp tests/mtp_draft_test.cpp npu/mtp/mtp_head.cpp → all checks passed, rc 0. Section [16] prints exactly the body's numbers:

PASS  leaf fill (1): the base's rank-2 candidate at the SAME position is refused every time
PASS  leaf fill (1) control: with the argmax on the runner-up, the same head is accepted every time
INFO leaf fill (1) control: same head, argmax moved -- accepted 20/20
INFO leaf fill (1): rank-2 candidate at every depth, accepted 0/92
PASS  leaf fill (2): a correct depth-2 candidate is never scored while depth 1 is refused
PASS  leaf fill (2): widening the batch (k=8, M=16 rows) buys exactly nothing when the parent is refused
INFO leaf fill (2): correct deeper candidates, accepted 0 at k=4, 0 at k=8 (of 92 / 184 rows)

copyright.py --check rc 0. 3 files, +139/−3; the tools/mtp_acceptance.cpp change is comment-only. (My local -fsyntax-only of it stops at nlohmann/json.hpp, a submodule this worktree doesn't have checked out. main's copy stops at the same line, so that's the environment, not the PR.)

Nits, not blocking

• tools/mtp_acceptance.cpp:489-492: the new sentence ends …cost is the base or the head. (P6's "leaf fill" … [16].), but the next, untouched line still begins // the head. Measured here…, so the comment reads "…the base or the head. (…) the head. Measured here". Drop the leftover the head. on that line.
• The doc says "this verifier", which is right. One word, "greedy", in the bold lead would stop a reader from carrying the conclusion to a sampling verifier (speculative sampling can accept non-argmax drafts).

refs #805 is right: this closes the leaf-fill item as "not this design" and leaves the tree-verifier question named.