Unmodified Qwen3.8 passes the 12-case quantization gate

The NVIDIA NVFP4 checkpoint passes 10 deterministic reasoning tasks and two long-context needles before Rocket changes any dense precision.

Question. What answer baseline must every dense-family precision change preserve?

setup
nodes	head and worker, two GB10 nodes
kernel	6.17.13-rocket64k, 64 KiB pages, both nodes
engine	vLLM 8e685d198, TP2, EP2, FP8 KV, no speculation
checkpoint	nvidia/Qwen3.8-Flash-Next-NVFP4 at fc694b54fb0174e0913e6adf86691ef85a4ead47
Rocket commit	a94c7119cacb750e92190e1b873137de85c26902
sampling	greedy, 2,048 completion-token cap
replicates	1
variance	unavailable

scripts/numerics/qwen38-quality.py --out qwen38-baseline-quality.json

arith_chain    PASS  8.4s
trains         PASS  6.9s
logic_grid     PASS  3.6s
sqrt2          PASS  2.3s
counting       PASS 15.4s
units          PASS  7.1s
date           PASS  3.4s
modular        PASS 11.8s
ordering       PASS  6.7s
negation       PASS 16.2s
needle_8000    PASS 10.2s
needle_64000   PASS  9.6s
score          12/12

class	passed	total	longest prompt
deterministic reasoning	10	10	111 tokens
needle retrieval	2	2	104,088 tokens
total	12	12	104,088 tokens

The negation case consumes 299 completion tokens, including 279 reasoning tokens. A 256-token trial returned no final answer, so precision comparisons retain the 2,048-token cap.

Verdict. Accepted. Any isolated precision map that regresses one passing case is rejected. Answer parity, completion length, and reasoning-token changes are retained per case.

Next.

• capture per-layer activation ranges for linear-attention input projections and output projections
• compare the first NVIDIA-layout NVFP4 linear-attention overlay against this JSON with --compare
• add a larger perplexity and agent-task corpus before accepting the final precision map

Reopen if.

• the checkpoint, tokenizer, vLLM revision, or CUDA driver changes