[Decision] An English docs correction leaves its locale siblings asserting the corrected-away claim — #59 says delete them, the seat post says leave them, and #68 blocks the pass that would fix them

[os-bill]

Filed by the repo:objectos seat, session_01ChPQM8jamxLUfUAxwFpJ8S, round 1, from a concrete instance created by this lane, today. Not hypothetical and not a policy musing: PR #252 landed as 1034f5f an hour ago and the situation below is live on main right now.

The instance

#250 was filed because configure/data-sources.mdx told readers a shipped capability was a roadmap item. Its own severity argument: "Someone evaluating whether ObjectOS can sit on their existing production database reads this page, concludes the safety gates they need are a roadmap item, and leaves."

PR #252 fixed that — in English. Measured on origin/main after the merge:

File	Still carries the retired forecast?
data-sources.mdx (en)	no — clean, 0 hits
data-sources.ja.mdx:165	yes — 「計画されている、まだ出荷されていないもの」
data-sources.ko.mdx:165	yes — 「계획되어 있지만 아직 출시되지 않은」
data-sources.fr.mdx:250	yes — "(statut : Proposed). Prévu, pas encore livré"
data-sources.zh-Hans.mdx:198	yes — 「计划中、尚未发布的内容」
data-sources.zh-Hant.mdx:199	yes — 「計劃中、尚未釋出的內容」
data-sources.de.mdx:238	yes — ## Roadmap: In-Product External Datasource Federation
data-sources.es.mdx:238	yes — same section, ADR-0015 link intact

So the exact harm the card was filed over is fixed in one language and live in seven. No gate fires: check-translations.mjs passed on the PR, and the freshness signal these siblings now carry is provenance, which routes them to a worklist (see below).

This is not a one-off. Every English correction this lane lands creates it. #254 — already queued, the twin defect on extend-existing-systems.mdx — will create it again on that page's siblings the moment it lands.

Why the seat cannot settle this: two recorded positions point opposite ways

#59 (closed, adjudicated) says delete:

Fumadocs falls back to English when a translation is missing, so ⇒ Delete the stale translations of any rewritten page rather than leaving them. A stale translation renders wrong content; a missing one renders correct English. ⛔ Do not leave a translated page asserting the retired story.

The repo:objectos seat post (standing note) says leave:

Locale siblings go stale by design — seven quickstart.*.mdx still carry the 3002 banner, correctly.

Both have a real argument. #59's is mechanical: missing renders correct English, stale renders wrong content, so leaving is strictly worse than deleting. The seat post's is that a stale port number in an otherwise-correct page is not worth deleting a whole translated page over.

The distinction that probably separates them is whether the stale claim is merely out of date or actively false in the direction that costs a reader — but that line has never been drawn, so it is being decided ad hoc, per PR, by whoever is holding the seat.

And the escape hatch is closed. PR #252's body says the siblings "the next translation pass fixes". That pass cannot run: #68 (create the translation account) is an open decision card, and #75 and #135 are both pm:blocked on it. So "leave it for the pass" currently means "leave it indefinitely", which is not what that sentence promises.

四维分析

维度	读数
长远合理性(权重 ≥50%)	指向「按判据删」。 #59 的论证是机械的、与具体页面无关的:缺失 → 回退到正确英文;过期 → 渲染错误内容。留着永远严格更差。而现状最坏的不是任一选项,是没有规则 —— 同一个问题每次由当班席位临时决定,#59 与座位贴各被引用一半,这正是本卡要终结的漂移。
实际业务需求	指向「删」,而且这次有量。 受影响的是评估期读者 —— #250 自己论证的正是这类读者「读完就走」。七种语言里,日/韩/法/简/繁五种明确写着「尚未发布」。这不是陈旧,是反向错误。
防 AI 写代码犯错	指向「删」并写成机械判据。 现在没有任何门禁能区分「过期但无害」与「过期且反向错误」,靠的是 dev 和 PM 每次的判断力 —— 而本轮的 dev 在 PR 正文里写下了「下一次翻译 pass 会修」这句今天为假的话,不是失职,是它信了座位贴那条note。声明即强制:没有机制兑现的承诺不该写进 PR 正文。
创业阶段不扩散	偏向最小动作。 删除 = 少 7 个文件、读者拿到正确英文、零新机制。反过来「先建翻译账号再补翻译」是把一个内容删除挂在一条未决的基础设施决定(#68)后面。⚠️ 但删除是发布内容的删除,不可逆(重译需要 #68 那个不存在的账号),所以本席不自裁。

四棱基本同向(删),但落在人工地板上:发布内容的破坏性删除 + 两条在案记录直接冲突。 按规则恒交维护者。⛔ 本席不代裁 —— 执行席无代裁通道,且这不是「恢复不变量」那类自带答案的修复。

Options

	动作	代价
A(推荐)	定一条判据:英文修正移除或反转一条能力断言时,同 PR 删除其 locale 兄弟;纯措辞/数字漂移(如 3002 端口)照旧留给 pass。立即适用于本次 7 个文件与 #254	需要写清判据;删除不可逆
B	一律删除任何被改写页面的 locale 兄弟(#59 字面)	简单机械,但会因一个端口号删掉整页翻译
C	一律留着,等 #68 解锁翻译 pass	诚实但今天等于「无限期留着错的」;需要同时撤掉 PR 正文里那句承诺
D	先解 #68,再回来定这条	把内容正确性挂在基础设施决定后面

推荐 A,判据落在「断言被移除或反转」而不是「文件被改动」——它正好把 #59 的案例(退役故事)和座位贴的案例(端口号)分到两边,两条在案记录都不必被推翻。

若你选 A 或 B,本席立即派发本次 7 个文件的删除,并把判据写进座位贴的常设条款,#254 落地时同步适用。 若选 C,本席会去修正 PR #252 正文里那句今天为假的承诺。

Re-check

git -C objectos fetch origin main
for f in $(git ls-tree -r --name-only origin/main content/docs/configure | grep -E 'data-sources\..*\.mdx$'); do
  echo "--- $f"; git show origin/main:$f | grep -niE 'ADR-0015|Proposed|Roadmap' | head -2
done
git -C objectos show origin/main:content/docs/configure/data-sources.mdx | grep -c 'not yet shipped'

Related: #59 (the delete ruling) · #68 (blocks the translation pass) · #75 · #135 (both blocked on #68) · #250 / PR #252 (the instance) · #254 (will repeat it).

[os-bill]

Correction to my own framing above — it overstated the mechanism, and the difference matters for the decision.

I wrote that "Every English correction this lane lands creates it." That is wrong. Measured on 8e665c7, the commit immediately before PR #252 landed:

File at 8e665c7 (pre-fix)	Carried the stale forecast?
data-sources.mdx (en)	yes — 2 hits
data-sources.ja.mdx	yes — already

So the seven siblings were already wrong before my fix. A translation is derived from its English source, so when the English page carries a false claim, its siblings carry the same false claim from birth. PR #252 did not create eight bad pages — it fixed one of eight copies of a defect that was already there, and made the other seven visible by leaving them provenance-stale.

What changes, and what does not:

• ⬇️ Urgency drops. This lane is not generating new damage with each correction. Nothing got worse today.
• ⬇️ The "we broke it, we should fix it" argument is void. It was broken on arrival.
• ➡️ The decision itself is unchanged and still needed. The question was never who caused it — it is whether a locale sibling asserting a corrected-away claim should be deleted (English fallback renders correctly) or left for a pass that The ownership check is inert on main until TRANSLATION_BOT_LOGIN is set #68 currently blocks. The Deploy section documents a retired open-source edition — rewrite it for the commercial runtime #59 and this seat post still point opposite ways, and that conflict is still resolved ad hoc per PR.
• ⬇️ [finding] extend-existing-systems.mdx carries the same stale federation claim #250 fixes on configure/data-sources.mdx #254's sequencing pressure drops. extend-existing-systems.mdx has seven siblings and they already carry the claim, so dispatching [finding] extend-existing-systems.mdx carries the same stale federation claim #250 fixes on configure/data-sources.mdx #254's English fix ahead of this decision makes nothing worse. I am no longer treating [finding] extend-existing-systems.mdx carries the same stale federation claim #250 fixes on configure/data-sources.mdx #254 as waiting on this card.

The four-quadrant reading above holds on its own terms — the long-term axis was about having a criterion at all, not about attributing blame — but the business-need axis should be read one notch cooler than I first wrote it.

Recording this rather than quietly editing the body: the original claim is what a reader would have decided against, and I would rather be visibly corrected than silently accurate. The root error is the same one this lane keeps paying for — I inferred the causal direction from the after-state instead of reading the before-state, which is one git show at the parent commit.

---

Generated by Claude Code