I want to write a paper review of . Links: - - - Run the paper-review-swarm workflow. Parallel agents: (A) style draft in my voice, (B) 250-word role/comparison paragraph for Conclusions (prose, not bullets — see anchors). Then draft-critic on (A), **fact-verifier agent** on the assembled post (see below), synthesize, save to `_posts/YYYY-MM-DD-paper-review-.md`, re-critique the saved file. Placeholder image paths are fine — I swap them later. Strict constraints (derived from past review drift): **Voice anchors — the anchors are the spec. If the rules below contradict the anchors, trust the anchors:** - `_posts/2026-02-16-paper-review-kimik25.md` - `_posts/2026-02-23-paper-review-drope.md` - `_posts/2026-03-16-paper-review-haclr.md` The anchors use neutral expository present tense in method sections, heavy bolding of key concepts (not just system names), and anchored interpretive sentences when they genuinely aid understanding. They do NOT use tutorial-explainer voice ("the trick is...", "neat serving trick"). **Length.** 800–1100 body words for typical papers; up to 1400 for frontier-dense systems. Go long only when the technical surface genuinely demands it. **First-person discipline.** Personal voice ("I like that...", "I am less sure") belongs in the opening hook (1–3 paragraphs) and Conclusions only. Method/experiments/limitations sections stay neutral. Anchored interpretive sentences ("The core insight is that X") are fine when the "X" immediately follows — what breaks the voice is disembodied value claims ("the neat trick", "a genuine gain") that are not anchored in the mechanism being described. **Banned editorial moves in method sections** (specific phrasings that broke past reviews): - "the trick is" / "the neat X" / "the clever Y" / "the tricky part is" - "that is it." / "does not pretend" / "is not trying to hide" - Invented metaphors: "a stability tax", "does not slosh back", "has the pricing to actually use it" - Explainer framing: "actually trainable and servable", "what makes this work is" - Disembodied value adjectives in methods: "neat", "clever", "delicate", "elegant" when they replace concrete description "Key insight", "core insight", "elegant" in Conclusions, and similar phrasings that appear in the anchors are fine. The ban is on populistic register, not on every editorial move. **Conclusions shape.** Prose comparison of 2–3 reference points, followed by a short hedged personal take. The anchors use prose for category-level contrast, not bullet lists. Keep Conclusions to 250–400 words. **Title.** Thesis subtitle preferred: "Why X Needs Y, Not Just Z" over descriptive enumeration like "X: A, B, and C at N Parameters". **Technical depth ≠ tutorial voice.** Depth means concrete numbers inline (parameter counts, benchmark scores, FLOPs, latency multipliers) and named components (CSA, GRPO, MXFP4). It does not mean walking the reader through why each number is impressive. **Section-title/body consistency.** If a section title names components ("CSA, HCA, and SWA"), the body must describe each. If editing trims a component out of the body, update the title too. This also applies to subtitles ("On-Policy Distillation and Generative Reward Models" — if GRM is cut from the body, cut it from the title). **Comparison attribution.** Every "beats X" / "trails Y" claim must name the baseline group. "V4-Pro-Max beats prior open-source SOTA by 20 points on SimpleQA-Verified" is correct; "beats Gemini by 20 points" is wrong if the 20-point margin is against open-source models. When a paper claims a headline margin, check the table for *who* the margin is against before carrying it into the review. **Fact-verifier pass (new).** After draft-critic and before synthesis, run a dedicated fact-verifier agent. Its input is the draft plus the paper PDF. Its output is a per-claim verdict for every numerical figure and every comparison in the draft: "verified (Table/Figure N, page P)" / "needs correction: actual value is Z" / "not found in paper — remove or verify". Draft-critic catches style; the fact-verifier catches truth. Both are required. Do not rely on the initial extraction summary for numbers; the source of truth is the paper itself. **Cross-links.** Link to a prior review of mine only when it substantively strengthens a specific claim (e.g., the paper directly uses a technique I already reviewed). Don't force them. For dswok notes, WebFetch `https://dswok.com/` to verify the note exists before linking; zero dswok links is an acceptable outcome. === I have updated the paper review. I want you to do two things: 1. Verify factual correctness, that it covers the paper and doesn't have wrong or nonexisting ideas. 2. Analyse the differences in the content, structure and style, think why I did these changes and suggest how to update the style guide.