"Will be reunified" vs "will surely be reunified": how translation reshapes Beijing’s Taiwan position.

From editorial note to queryable record

The verbatim-verification pass surfaced six translation discrepancies across five doctrinal anchors but logged them only as editorial notes — not readable by the retrieval system, not surfaced to readers. This release closes both gaps. Each doctrinal passage in the corpus now carries an attached set of translation variants as structured data: the passage location, the two competing renderings, their publishers of record, an editorial discrepancy note, and a doctrinal-load tag (high, medium, or low). All six variants are populated. The two heaviest entries — both in Xi's 2 January 2019 Taiwan address — are surfaced on the live page in a dedicated callout. The schema work is the precondition for any retrieval system that scores against PRC-English; the live surfacing is the editorial finding the discrepancies always implied.

A single working translation is a forecast-quality liability

The founding corpus carried one English text per anchor. The verbatim-verification audit corrected that text where the working English diverged from a verifiable primary source, and surfaced cases where multiple verifiable sources diverged from each other. The audit's output sat in editorial notes.

Editorial-note-only discrepancies fail two downstream consumers. First, the retrieval pipeline runs over the structured doctrinal-passage objects in code; it does not parse editorial notes, so it scores against whichever single working English the editor picked. Second, the live page reader sees only that single working English, with no indication that doctrinally significant variants exist. Both consumers operate on a flatter version of the corpus than the audit actually produced.

The fix is structural, not editorial: extend the schema so variants are first-class data, and use the live page to surface the variants whose doctrinal load is heaviest. Editorial judgment still picks the working text; the variants ride alongside it.

The variant schema and the doctrinal-load tag

Each translation variant is a structured record with six fields: the in-passage location it applies to, the rendering used as the working English, the competing rendering, the publishers of record for each side, a one-to-three-sentence editorial note on what the divergence implies, and an ordinal doctrinal-load tag.

Each doctrinal passage in the corpus now carries an attached tuple of these variants, with an empty tuple as the default — so anchors with no documented variants carry no overhead. The retrieval pipeline can read variants at query time; downstream language-model calls can be passed the full variant set rather than a single point estimate; the live page can render whichever subset clears a publication threshold.

The doctrinal-load tag is editorial, not statistical. HIGH applies to variants where the wording carries a substantive doctrinal claim — a use-of-force formulation, a sovereignty assertion, an escalation register. MEDIUM applies to variants that change rhetorical force without altering the underlying claim. LOW applies to translator-style differences with no doctrinal stakes. Two of the six variants are HIGH; three are MEDIUM; one is LOW.

Two HIGH-load entries, both in Xi's 2019 Taiwan address

China must be reunified, and will surely be reunified.
— CSIS Interpret

The CSIS Interpret rendering carries 必然 as deterministic, almost teleological — the “surely” doing serious doctrinal work. The Xinhua Highlights variant flattens the same Chinese to “China must be and will be reunified”— resolute, but not metaphysical. The Xinhua flatter wording has become the post-2022 PRC-English convention, which means a 2026 PRC-English headline on Taiwan that omits “surely” is plausibly tracking the post-2022 register rather than walking back the 2019 claim.

“make no promise to renounce” → “do not renounce” → “will not renounce”

The second HIGH entry tracks the use-of-force formulation across three documented PRC-English variants between 2019 and 2022. The literal Chinese (不承諾放棄) is “do not promise to renounce.” The 2019 Xinhua rendering reads “make no promise to renounce” — passive restraint, force option not waived. CSIS Interpret renders the same clause as “do not renounce” — the negation moves from the promise to the underlying act. The 2022 white paper renders it as “will not renounce” — the future-tense active retention of the force option.

The shading from passive restraint not waived to active retention of the force option is not translation noise. It maps exactly onto the post-Pelosi doctrinal escalation the 2022 Taiwan white paper itself codifies — the white paper drops the 1993 and 2000 commitments to refrain from sending troops or administrators to Taiwan post-reunification. The divergence between the three renderings is itself the signal: PRC official translators are tracking a hardening in the Chinese original's authoritative reading, and the structure of the English sentence absorbs the hardening.

For the predictor: every retrieval against the 2019 Xi anchor now carries the full variant set, not just the single working text. A Taiwan-incident event that surfaces PRC-English “reunified” language in 2026 maps to a different doctrinal posture depending on whether the rendering tracks the 2019 wording or the 2022 wording. The variant tag is the input the retrieval system needs to make that distinction.

Translation provenance is a doctrinal feature

The substantive claim of this release is that translation provenance is a doctrinal feature in the technical sense: a property of the data the predictor can condition on. It is not metadata to be cleaned up; it is signal to be retrieved on. The PRC publishes its own English. That English is curated. The curation tracks doctrinal direction-of-travel. Any forecasting system that retrieves on a single point-estimate English text is reading the corpus at lower resolution than the corpus actually carries.

On the question of which translation the public-facing corpus should show by default when PRC English diverges from Western analyst translations: the schema work above lets the question be deferred rather than answered. The public-facing corpus shows the working text by default and exposes the variants via the doctrinal-load tag. Readers who want the divergence can see it; readers who want one wording get one. The retrieval system gets the full variant set regardless.

The four lighter discrepancies — in Mao 1956, Jiang 1995, Xi 2012, and the 2024 productive-forces essay — each carry medium or low doctrinal load. They are populated in the schema for completeness but are not surfaced on the live page in this release. A future version with empirically learned per-passage weights may upgrade some of them; the current threshold is conservative.

What this does not yet resolve

The retrieval scorer does not yet condition on variants. Variants are queryable, but the stub retrieval pipeline still scores against the working text only. The language-model swap, scheduled for a later release, is the first version where variant conditioning becomes part of the API call.

Doctrinal-load tagging is editorial.The high / medium / low ordinal is the editor's call, not a learned weight. That is appropriate for six variants but does not scale; the auto-research loop will need a learned variant-importance signal once the corpus widens.

The variant set is bounded by the audit. This release surfaces the six discrepancies the verbatim-verification audit found. A second-pass audit (or a wider corpus) will likely surface more. The schema accommodates additions; this release only loads what is already documented.