Rachel → approved "DeepSeek and Peking University Open-Source a Toolkit That Cuts AI Response Times by Up to 85%": "Strong inference-layer story. Mechanism and reproducibility bundle are real. Vendor-reported speedup is appropriately caveated; approve as-is without waiting for third-party reproduction."
Ready to publish: "DeepSeek and Peking University Open-Source a Toolkit That Cuts AI Response Times by Up to 85%"
Iris → use: "Use this story. The source ships three concrete open artifacts (paper, code, model weights) and reports hard, paper-backed numbers — 60–85% per-user speedup on DeepSeek-V4-Pro-DSpark and V4-Flash-DSpark, plus +16.3%–18.4% acceptance length over DFlash and +26.7%–30.9% over Eagle3 on Qwen3. That combination is what makes this a publishable inference-infrastructure story rather than a hype piece. The angle should stay constructive (real engineering progress that lowers deployment latency and cost) but should preserve the framework's two named technical bets — semi-autoregressive draft generation and confidence-scheduled verification — and the source's own framing that this is engineering landing, not a model-architecture leap. Do not soften the headline numbers into vague language; 60–85% and the acceptance-length deltas are the news. Cold-reader risks the reporter must handle in every candidate: DSpark, DeepSpec, DeepSeek-V4-Pro-DSpark, V4-Flash-DSpark, Eagle3, DFlash, Qwen3, Markov head, semi-autoregressive, and 'speculative decoding' all need at least a one-clause category doorway in the dek or lede. Liang Wenfeng is a proper noun a smart non-Asian-tech reader will not recognize and should be introduced by role in body, not name-dropped in the title. Items flagged for the reporter, not for this creative pass: (a) the 500亿元 financing context is presented as background, not as a direct cause of this release — the source should support any implied causal claim before it is published; (b) DeepSeek-V4-Pro and V4-Flash model identities and capabilities should be independently confirmable from DeepSeek's public channels before publication; (c) source is a Chinese-language Leiphone (雷锋网) explainer — if the outlet runs in English, translate 推测性解码, 半自回归, 置信度调度, 草稿模型, 后缀衰减 on first use with a brief gloss, and keep DSpark/DeepSpec as proper nouns."
Sky → pursuing: "Reporter should hydrate the primary sources (DeepSpec paper PDF, GitHub code repo, HuggingFace model page) and verify the production deployment claim against DeepSeek's live API or status, then draft the story around the semi-autoregressive + confidence-scheduled verification mechanism and its implications for inference economics."