Fix style defects in three queued draft posts

[krMaynard]

Editorial fixes found while auditing the 22 queued drafts for publish-readiness. All three posts remain drafts (published: false); English and every existing locale (ja/zh/ko) updated in the same commit.

Changes

apple-transparency-report — the real defect. Dropped the unsourced claim:

consistent with the well-known pattern of German police running device identifiers in theft cases

"consistent with" already carries the hedge; "well-known" asserted something about the world with no source behind it.

korea-network-act-report — the "A, not B" construction appeared 3× in one post. Reworded the H2 so the remaining two land harder:

The CSAM cliff is a definition, not a trend → The CSAM cliff is a definition change

korea-naver-kakao-requests — label-colon became a plain sentence:

The biggest caveat: these are… → The biggest caveat is that these are…

Only en and zh changed here: the ja and ko translations already used the natural "the biggest caveat is that…" construction.

Verification

• 10 files, 10 insertions / 10 deletions — each replacement asserted to match exactly once.
• npm test → 86/86 pass.
• All three posts still published: false in every locale, so nothing goes live on merge.

With this merged, all 22 queued drafts are clean and publish-ready.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UaTzkTu85JrQGeuJWGwQtk

[krMaynard]

Self-review

Re-read the full diff line by line. The three requested style fixes are correct in all four languages, and each replacement was asserted to match exactly once before writing (no silent partial edits).

One issue found during review, and fixed in 3b9e776.

Reading the Japanese Apple paragraph I noticed a defect I introduced myself in PR #212: converting parenthetical em-dashes to 、 joined two independent clauses with a comma where Japanese requires 。:

…米国（129,774件）とブラジル（54,425件）を合わせたよりも多い**、**これは盗難事件で…

Scanned for the whole class (plain-form predicate + 、 + demonstrative) and found 6 instances across 5 files, now all corrected to 。:

File	Was
ja/blog/apple-transparency-report.html	…よりも多い、これは
ja/blog/dsa-own-initiative-scope.html	…皆無である、これは
ja/blog/snap-transparency-report.html	…報告している、これは
ja/blog/snap-transparency-report.html	…71%）だ、これは
ja/blog/turkey-5651-reports.html	…だった、これは、
ja/blog/eu-ai-act-training-transparency.html	…から生まれる、それは

Verified zero remain. This does widen the PR slightly beyond the three requested fixes, and it touches two already-published posts (dsa-own-initiative-scope, eu-ai-act-training-transparency) — flagging that explicitly since it's a correction to live content. It's a 、→。 punctuation fix only, no wording changed, so no cross-locale update is required.

Checked and found clean:

• Chinese and Korean don't have the equivalent problem (both tolerate comma-joined clauses; zh uses ， and ko uses , idiomatically here).
• All three target posts remain published: false in every locale, so merging publishes nothing.
• npm test → 86/86 pass after both commits.

[krMaynard]

Critique: eu-ai-act-training-transparency (English)

Posting an editorial review of the EU AI Act post, since this PR touches it (the ja comma-splice fix). This is about the English content, so none of it is a blocker for merging #221 — it's a separate follow-up if you want it. Strongest points first.

Overall

Good post with a genuinely non-obvious thesis: who filed tells you more than what they filed, and it's mostly a timing artifact. That's the sharp, defensible core. Two weaknesses: the piece applies its own thesis too cleanly (its data contains two exceptions it never reconciles), and it states the thesis ~5 times.

Substantive: the taxonomy has two counterexamples in its own data

The clean binary at the "who's missing" section — filers are either brand-new post-cutoff flagships or open/European labs that filed proactively — breaks in two places the post names itself:

1. Microsoft (Phi-4) fits neither bucket. Large US lab, small-model family, not an openness-pitch European lab. Why did Microsoft file when Anthropic/Mistral/xAI didn't? Never addressed. A reader spots it because it's in the filer list. Either it undercuts the timing story or there's a reason (Phi-4 shipped post-cutoff?) that would strengthen it — but it has to be stated.
2. "Fable 5" as a "current flagship" collides with the timing defense. Absentees are excused via the 2027 transitional deadline (pre-Aug-2025 models). But Fable 5 is called Anthropic's current flagship in a July 2026 post. If it's genuinely current (post-cutoff), Anthropic is obligated now and already non-compliant, and 2027 doesn't cover it. The honest, stronger version: some absences are innocent timing, at least one high-profile one may already be late — which makes the "who's missing" signal more interesting, not less.

Thesis repeated too often

Appears in the description, intro, the H2, the full "timing seam" paragraph, and again in the closer. The final paragraph is essentially the timing-seam paragraph restated with a "not X, it's Y" construction + recap — the exact ending pattern the no-slop pass targets elsewhere. Trim it to the one new clause (keeping the dataset updated).

A sharper point left on the table

The top-band saturation observation is good, but the middle text band is 1 billion to 10 trillion tokens — four orders of magnitude in one bucket. The claim that size fields "usefully separate a small/older model from a modern large one" rests on a band nearly as uninformative as the top one. Naming the middle band's width lands harder.

Unacknowledged tension

The post credits the coarseness as a deliberate trade-secret compromise ("set low by design"), then predicts pressure to "add finer buckets or report an actual number" — without noting that the coarseness is why providers agreed to file at all, so the AI Office can't easily tighten it. One sentence closes the loop.

Accuracy & verifiability

• 11 models, 7 named. "Eleven models from seven providers," then seven parenthetical names. Reader can't reconcile 7 → 11 (Phi-4 family presumably covers several). Name all 11 or say "the Phi-4 family accounts for N."
• Link the primary sources. The whole piece is about a non-registry assembled by hand, yet it links none of the individual filings (the Meta/OpenAI PDFs, Microsoft card, SmolLM3 Gradio widget). Per-filing links would prove the scavenger-hunt point, enable verification, and protect the load-bearing model IDs (Gemini 3 Pro, Muse Spark, GPT-5.5, Fable 5) that carry the dataset's authority.
• Penalty figures (3% turnover / €15M) and the 2025/2026/2027 dates are internally consistent and match the AI Act GPAI regime. Good.

Prose

• Intro is a three-clause run-on: "...into a small, comparable dataset, it's live at [link], and the exercise..." Two independent clauses spliced around a link. Break into two sentences.
• List ambiguity (same spot the JA translator restructured): "Anthropic, including its current flagship, Fable 5, Mistral, and xAI have not filed" — four commas let a reader parse Fable 5/Mistral/xAI as all Anthropic flagships. Parenthesize: "Anthropic (including its current flagship, Fable 5), Mistral, and xAI."
• Bolded paragraph lead-ins are a good scannable device — keep.

If only three things

1. Reconcile Microsoft and Fable 5 with the timing thesis (one sentence each) — biggest credibility gain.
2. Add per-filing source links — assertion → evidence, and de-risks the model IDs.
3. Cut the recap ending and fix the two prose snags.

Any content edits here would need mirroring into all 7 translations.