Introducing the Doctrinal Event-Response Predictor.

A structured forecaster on top of sixty years of doctrine

On 7 October 2022 the U.S. Bureau of Industry and Security published its advanced-computing rule against the PRC chip industry. Inside twenty-four hours, MOFCOM and the MFA produced a struggle-and-self-reliance-framed rhetorical response. Ten months later, MOFCOM imposed gallium and germanium export controls. Twelve months later, Big Fund Phase III was finalized. The substantive cascade ran for a year. None of this was unforeseeable from the corpus: Xi's 2018 CAS/CAE self-reliance keynote, the 2020 People's Daily dual-circulation editorial, and Mao's 1956 Ten Major Relationships all pre-figured the response template.

This paper builds a structured predictor on top of that observation. Twelve doctrinal anchor passages from 1949 to 2024 carry the lineage. Nine historical event-response pairs from 2018 to 2026 are the eval set, with the realized frame, escalation tier, primary policy lever, and timing window all known. The predictor scores a held-out event against the rest, on all four dimensions, with the registry append-only. We log one live ex-ante prediction at release: the PRC response to the May 2026 talent-curb expansion is forecast as a self-reliance-framed response at tier 4, with talent restriction as the primary policy lever, within approximately 257 days. The prediction resolves on 7 February 2027.

The doctrine is the grammar of every response

State responses to foreign actions don't come from nowhere. They route through a sixty-year register: Mao on humiliation and self-reliance, Deng on hide-and-bide and selective opening, Jiang on the Eight Points, Hu on the Scientific Outlook, Xi on rejuvenation, struggle, and new-quality productive forces. The corpus is dated, sourced, and indexed. The grammar an export control or a Taiwan incident gets framed in is bounded by the passages the response's authors are willing to invoke. So the response is partially predictable from the corpus alone.

The point is not novelty of underlying data — the public doctrinal record is what it is. The contribution is a willingness to commit a structured number, on a specific event, with a fixed resolution date, in a form that can be marked right or wrong twelve months later. Mistakes are permanent in the registry. The May 2026 talent-curb expansion is the first live prediction.

The corpus, the eval set, and the live call

The three-tab board below is the working surface for the rest of the paper. Tab 1 carries the doctrinal corpus — 15 anchor passages from 1949 to 2024, filterable by era and theme, plotted on a four-era timeline. Tab 2 carries the nine historical event-response pairs from 2018 to 2026, with frame / tier / lever filters and a timeline that places each event on its realized tier lane. Tab 3 carries the open live prediction (May 2026 talent-curb expansion, resolving 7 February 2027) and the three-baseline leave-one-out backtest.

The founding corpus shipped with twelve anchors; a later update loaded the three deferred Xi-era anchors (the 2017 19th Congress political report, the 2021 common-prosperity speech, and the 2024 Third Plenum communique) into the live retrieval index, widening the corpus to fifteen.

What the nine pairs reveal

The response template is stable on two axes and unstable on a third. On frame and tier, the categorical mapping is sticky: U.S. export-control events land on a self-reliance frame at tier 4 in seven of seven cases since 2020; Taiwan-incident events land on a sovereignty frame at tier 5. That stickiness is what the most-common-by-category baseline picks up — it clears 67% framing accuracy. The unstable axis is lever: the modal response inside the U.S. export-control category oscillates between subsidy (2020, 2022, 2023) and retaliatory export control (2018, 2024) depending on the specific U.S. package and where the PRC is in its industrial-policy cycle. That is where doctrinal retrieval has to do real work. The May 2026 talent-curb expansion is the diagnostic case: the lever is novel (talent restriction), but the doctrinal lineage (Xi's 2018 self-reliance keynote plus the 2024 essay on new quality productive forces) was retrievable from the corpus six months in advance.

The first ex-ante scored claim

Bloomberg's 26 May 2026 report on the PRC expansion of AI-talent travel curbs to private firms is the launch event. The corpus treats the event itself as a same-day self-reliance and talent-restriction response. The forward call — visible on the Live forecast tab above — is on the substantive cascade: the state-media codification and subsidy follow-on expected over the next eight months, on the same lag template the 2022 U.S. export-control package produced.

Falsification criterion.Within 257 days of 26 May 2026, official state media (People's Daily, Xinhua, or Qiushi) or a State Council, commerce-ministry, or defence-ministry notice produces a follow-on response framed primarily in self-reliance terms whose dominant policy instrument is talent restriction, with escalation observed at tier 3 or higher.

What would falsify this in 257 days. The call resolves againstthe predictor if any of these dominates the response: (a) the PRC reverses the May 2026 travel curbs, signalling a peaceful-rise pivot; (b) state media reframes the talent-restriction track as a common-prosperity measure; (c) the dominant follow-on lever is a military-posture change or a sanction against a named U.S. firm with no parallel talent-restriction codification; (d) no follow-on Politburo readout, commerce-ministry notice, or front-page People's Daily commentary cites the May 2026 measures by 7 February 2027 at all; (e) the cascade lands but at tier 1 or 2 only.

Five moving parts, each versioned

Corpus. Twelve anchor passages at first release (widened to fifteen in a later update) spanning four eras, Mao through Xi. Each entry carries structured speaker, venue, and date metadata, the working English text, a translation-provenance tier, and a verification-status flag.

Pipeline. Retrieval is theme-overlap scoring: each event carries a category-prior theme set; each passage carries its dominant theme tags; their intersection size, multiplied by a source weight (Xi-era state media 1.0, pre-Xi historical 0.6), with a recency tiebreaker, is the retrieval score. The top-K passages enter the prediction call. The current release uses a deterministic stub model for the structured-output step; the language-model swap is scheduled for a later release, with no API contract change.

Output taxonomies. Seven frames: self-reliance, struggle, common prosperity, peaceful rise, sovereignty, new quality productive forces, and whole-process democracy. Five escalation tiers: rhetorical only, symbolic action, economic countermeasure, sectoral policy lever, and military-posture change. Seven policy levers: rhetorical only, subsidy, sanction, retaliatory export control, military posture, talent restriction, and unspecified. Window is in integer days.

Baselines. Three classical baselines run on every backtest: random over the closed taxonomies (floor), most-common-by-category using the modal training response inside the event's category, and nearest-neighbor copying the most similar prior event by category match and recency. The stub model has to clear all three on framing top-1 accuracy, tier mean absolute error, and lever F1 to count as evidence of doctrinal retrieval lift.

Metrics.Framing top-1 accuracy on the seven-way frame label. Escalation-tier mean absolute error on the 1–5 ordinal. Lever F1 micro-averaged across the seven-way closed lever set. Window calibration: fraction of held-out cases whose realized cascade lands inside the predicted window with a 1.5× upper tolerance. Per-dimension Brier scoring follows in a later release, once the registry carries resolved predictions.

Leave-one-out, three baselines, nine event-response pairs

The most-common-by-category and nearest-neighbor baselines clear two-thirds framing accuracy and about 0.44 tier mean absolute error on a 1–5 ordinal — respectable for a nine-case eval set — but lever F1 sits at the floor across the closed seven-way set, because the modal response inside each event category masks the lever the realized response actually used. Doctrinal retrieval has to beat this on lever F1 first, then on window calibration. The current stub is the staging ground; the language-model swap is the experiment. The full backtest table is the bottom block on the Live forecast tab above.

What v0 cannot honestly claim

Quote provenance is incomplete at first release.Every one of the twelve anchor passages ships with a placeholder tag on the English text. Speaker, venue, and date are independently verified; the working English is editorial paraphrase grounded in the dossier's reading of the primary record. The placeholder gate is deliberate — a falsifiable-prediction project cannot ship fabricated quotation — but it is a real limitation until the verbatim-verification pass lands.

The eval set is small. The leave-one-out backtest runs on nine event-response pairs. That is enough to separate framing top-1 accuracy from a uniform-random floor and enough to expose lever F1 as the hard frontier, but it is not enough to lock in a confidence interval on any metric. The eval set will widen to roughly thirty pairs in a future release. Until that lands, treat every single number on the backtest table as a directional signal, not a calibrated rate.

The taxonomies are closed.Seven frames, five tiers, seven levers, and a window in integer days. That is the entire output space. The closure is a design choice — it makes the predictor scoreable — but it costs the predictor any frame the corpus has not named yet. If a 2027 response invents a new register (the way new quality productive forces emerged in 2023–2024), the current taxonomy will mis-map it.

The corpus is time-asymmetric.Six of the twelve anchors are Xi-era (2012–2024). Two are Mao, two are Deng, one is Jiang, one is Hu. That weighting reflects the empirical observation that current state-media editorials cite Xi much more than they cite Deng or Hu — but it also means the prior systematically over-weights the post-2012 frame distribution.

Three anchors are flagged but not loaded at first release.Xi's 2017 19th Congress political report, the August 2021 common-prosperity speech, and the 2024 Third Plenum communique are flagged as known gaps but are not in the first-release retrieval index. Predictions in the interim do not draw on those anchors. They are loaded in a later update.

The registry is append-only. Every prediction is timestamped at write time and never edited. If a prediction is found to be wrong about the event rather than the response, it is annotated, not deleted. The May 2026 live call resolves on 7 February 2027 either way.

Seven choices the v0 gate does not foreclose

i. Translation of record.When PRC English diverges from the Western analyst translations (CSIS, MERICS, Brookings), which translation does the public-facing corpus show by default? The working recommendation is the PRC text with a translator's-note toggle exposing the divergence.

ii. Common-prosperity coverage.The current eval set is scoped to externally triggered responses and underweights common prosperity as a frame. Adding two or three internally triggered events — the 2021 platform crackdown, the real-estate intervention — as a separate sub-set is the natural extension.

iii. New quality productive forces versus classical self-reliance. From 2024 forward, state media use the new-quality-productive-forces register and the self-reliance register interchangeably. Keeping them as two frames is the right call: the lever implications differ (the productive-forces register clusters to subsidy responses, classical self-reliance to retaliatory export controls plus talent restriction).

iv. Hu Jintao corpus weight. Hu-era doctrine is rarely invoked in current state media but is diagnostically valuable as the peaceful-rise counter-anchor. The working recommendation is to keep it in the corpus with low retrieval weight.

v. Real-time corpus ingest cadence.Daily ingest of People's Daily, Xinhua, and Qiushi with weekly editorial-vector recomputation, matching the auto-research loop cadence.

vi. Passage source weighting.The current weighting puts Xi-era state media at 1.0 and historical anchors at 0.6 — modest, not aggressive. The auto-research loop is the better mechanism for empirically learned weights.

vii. Live-case publication policy. Publish live ex-ante predictions before the realized response, even at the cost of being publicly wrong. The May 2026 talent-curb prediction is the proof of concept; publishing only post-hoc would compromise the falsifiability claim.

The audit pass, then the schema, then the corpus widening

Three things follow this release immediately. First, a verbatim-verification audit pass: walk every anchor passage back to its primary source, flip eleven of twelve from placeholder to verified, surface the translation discrepancies the audit turns up. Second, a schema extension: make translation variants queryable from code rather than buried in editorial notes, and surface the doctrinally heaviest discrepancies on the live page. Third, load the three deferred Xi-era anchors: the corpus widens from twelve to fifteen, closing the post-2017 gap this release ships with. The live prediction resolves on 7 February 2027 regardless.