PP4 Round 12: what the greedy margins establish

Source/CPU evidence only. The historical admission remains NOT QUALIFIED. The proposed registration is pp4-r12-logit-grid-tie, for NEW PP4 experiments only. NBW, TV1 and W8 retain their existing rules.

Primary observations and their limits

Read-only campaign PP4-429-admission-W-20261027T133630Z stopped at cmd/0156-admission-B-greedy.log, after three divergent outputs:

Reference	First differing token index	Margin nats
0	97	0.0
1	78	0.5
2	63	1.0

admission-B-greedy.json SHA256: ace208cbb3135015577c163417f4504da9ab95ebefd8910e97d2dd7c78ef93eb. These are NBW diagnostics rounded to four decimal places. Its gap uses the BETTER of the two tokens, so zero does not prove that the experiment token was the top. No top-20 values, token IDs, absolute logits or runtime dtype were saved. Prompts 3–5 were never checked. This artifact cannot retroactively pass the new rule. Its observed margins can supply regression cases only when supplemented by explicitly synthetic top-20 evidence.

The binding Round 12 brief also reports DEP1's prompt-2 margin of 1.0. The named DEP1.report.md available during this review contains the Round 3 budget and earlier reports, without that flip's primary values. Treat the flip as coordinator-provided context, not an independently recovered top-20 receipt.

Read-only X86host PP4-server-R11-20261027T132720Z/summary.json records candidate 42928552727ee85561600d886a704d3342ad969a, DIAGNOSTIC COMPLETE, no stop and pair_restored=true. Its admission field explicitly says NOT QUALIFIED. The coordinator's baseline-reproduction/graph-reference result in the brief is separate from this summary's narrower fields. Round 12 changes no worker byte.

Dtype and API source chain

Image-source paths below are relative to /opt/dsv41/data/src-image/; they were read, never executed as servers.

vllm/models/deepseek_v41/nvidia/model.py:1601–1609,1638–1649 constructs ParallelLMHead and LogitsProcessor, and passes hidden states straight through that processor. The production-mounted K1 model is offload/kvpack/overlay/vllm/models/deepseek_v41/nvidia/model.py, SHA256 bc3aa8b06ff40c65e93f31e11c3cad38fcb4fb6425946e61829f041bac283d3f, exactly matching BOTH saved admission identities. Its two compute_logits method source spans are identical to the image's. The image file's complete hash differs because K1 has other changes; do not confuse the two files.
vllm/model_executor/layers/logits_processor.py:96–101,153–200,211–223 uses the head's quant_method.apply unless head_dtype explicitly differs from hidden dtype. Generation defaults to model dtype at vllm/config/model.py:2487–2506. The V4.1 constructor supplies neither scaling nor soft-cap overrides, so defaults scale=1/soft_cap=None apply. TP gathering and vocabulary slicing do not restore lost mantissa bits.
vllm_exl3_moe/config.py:234–241 selects Exl3HeadMethod for an EXL3 head. head.py:125–144 selects the BF16 mrow output for small BF16 inputs, or performs EXL3 FP32 accumulation and casts back to x.dtype. The saved admission env has VLLM_EXL3_HEAD_MROW=1 and EXL3_HEAD_MROW_MAX_ROWS=8 on both ranks. Thus changed verify/prefill row counts can change head kernels; the ordinary BF16 activation path returns BF16 logits on either branch. Unquantized heads similarly follow activation dtype. An explicit FP32 head override is a different configuration, and must not be presumed absent solely from a config hash. No live tensor dtype was measured in this lane.
vllm/v1/sample/sampler.py:80–103,307–309,312–360 computes raw logprobs with logits.log_softmax(dtype=torch.float32), promotes sampling logits to FP32 before processors, and gathers top-k plus the sampled token. ModelConfig defaults to raw_logprobs (config/model.py:285). For speculative rows, rejection_sampler.py:145–180,203–212,248–265 promotes model logits to FP32, keeps raw values separately from processors, reassembles accepted and bonus logits, then uses the same compute_logprobs/gather_logprobs path. PP4's probability verifier overlay does not increase head precision.
vllm/entrypoints/openai/completion/serving.py:650–735 serializes those values under token_id placeholders, clamping only below -9999. The grid gate refuses clamped/nonfinite/malformed data. Logprobs are FP32 values, not BF16 log-softmax outputs. Subtraction cancels the common log-partition up to FP32 rounding, exposing the already rounded head-logit differences.

Relevant read-only image SHA256:

model.py          66005a8edde825612bbb9e041a5a218132cf97123bba9b401890122de6314a89
logits_processor  35cf8235cb529ba2d37358fe9b9e5a29f40b5c3cbcc9e6750c8480c51bd7da83
sampler.py        a67286eb78aac98185b6922ac33d8c754ca5ff9788d893965ba52f2b5a0e18d1
rejection_sampler 9acaefa7ff3ff02649adf3a2edc4a7281d0089c23566fe5ff140c5f820d88ac4
exl3 head.py      295faadd33db3431e3ddaa54c7b2a410cd91b62e9b1957efcd68e5f8add1b744

Grid interpretation and the next window

BF16 has eight significant binary digits. Within a binade, spacing is 2^(floor(log2(abs(logit)))-7): 0.5 at magnitudes [64,128), 1 at [128,256), and 2 at [256,512). Spacing changes at binade boundaries, including on the lower side of an exact power of two. FP32 promotion cannot undo that rounding. CPU synthetic BF16 logits followed by the actual torch FP32 log_softmax show 0.5 gaps represented as 0.5000000596046448 and 1.0 gaps as 1.0; this explains why comparisons need FP32 rounding bounds rather than decimal equality.

The observed 0/.5/1 margins are CONSISTENT with this mechanism. Their logprob magnitudes do not reveal the absolute head-logit exponent: log-softmax is invariant to a common shift. Nor does a top-20 lattice prove its fundamental spacing; the finite sample could omit intermediate representable values. The registered checker therefore uses a conservative observable step: the smallest positive top-20 gap must itself be a binary power of two <=1, and EVERY pairwise gap must be an integer multiple within FP32 subtraction error. It never computes a fine GCD to rescue an off-grid set, chooses a larger tolerance from the divergent margin, or infers a grid from an all-equal set. Sparse or mixed-binade evidence may conservatively STOP.

Changed batch/reduction/EXL3 paths, changed attention/cache states, wrong token alignment, stale KV, and an actual verifier/failure-mask defect can ALL change the winning token. Prefix-cache reuse by the old diagnostic could also alter its shape. Rounded margins alone distinguish none of these.

The next coordinator window obtains a fresh salt for EVERY greedy generation and a DIFFERENT salt for its teacher-prefix probe; it saves all top-20 IDs and unrounded values, both compared IDs, exact experiment-to-top gap, derived step, all-pairs residual/tolerance and token counts. Only an experiment token actually at the top, both tokens present, <=one observable step, equal token lengths and at most three divergent prompts can pass. Every failure STOPs before cells; W8 still requires six distinct exact IDENTICAL results.

Fresh full-top20 evidence passing that rule supports only the registered narrow numerical allowance, not correctness of every verifier decision. A non-top experiment token, off-grid set, >one step or >3 flips falsifies that allowance and stops the window. To establish causation rather than compatibility, a separately authorized diagnostic must compare the SAME token prefix with fresh salts across target-only and controlled verify shapes, record pre-softmax dtype/ absolute top logits, and inspect argmax versus emitted verifier token before later tokens diverge. Shape-dependent neighboring BF16 bins support rounding; emission disagreeing with that shape's own argmax supports a verifier/state bug. That additional runtime instrumentation is not added or run by Round 12.