Where China's next half-trillion in industrial subsidy is going.

The cascade from doctrine to subsidy is legible.

The 14th Five-Year Plan, adopted by the National People's Congress (NPC) on 11 March 2021, named seven frontier fields, designated new-energy vehicles (NEVs) a strategic emerging industry, and committed to building a national integrated big-data center system.¹ Across the five-year window that followed, those plan lines compiled into specific ministerial documents: the third National Integrated Circuit Industry Investment Fund (Big Fund III), registered with RMB 344 billion in capital in May 2024; the NEV purchase-tax extension through 2027, issued by the Ministry of Finance (MoF) in June 2023; the East-Data-West-Compute project, jointly approved by the National Development and Reform Commission (NDRC) with the Ministry of Industry and Information Technology (MIIT) and partner agencies in February 2022; and the world's first model-layer regulatory regime, the Cyberspace Administration of China's Interim Measures on Generative AI, issued in July 2023.² The cascade is densely observable. The reading is not.

MERICS writes careful, slow-cycle long-form on plan implementation.³Trivium China reads ministerial follow-ons daily for paying clients behind a paywall. Stanford's DigiChina hosts academic forums on individual provisions; MacroPolo writes retrospectives. None of them ship a tool that takes a single plan line and emits a structured cascade prediction — ranked beneficiaries, a subsidy band, a ministerial lag, a falsification criterion, and a resolve-by date — with the historical precedents that grounded each call shown alongside, and the entire output committed to a public registry that can be scored against the realized record.

The difference is not the reading. It is the schema. Every prediction in this paper resolves to a single ministerial document at a single primary-source URL by a single date. MERICS publishes the strongest analytical narrative in English; this paper publishes the weakest analytical claim — a structured tuple — and stands or falls on whether the tuple matches the document that lands. The first scoreboard closes in late 2028.

The gap exists because the work falls between two professional cultures. The engineers who can build the retrieval pipeline rarely read the Five-Year Plans in the original language; the analysts who read the plans rarely build retrieval pipelines. The thesis of this paper is that fourteen plans worth of realized history is enough training signal to make the fifteenth a structured prediction problem rather than a commentary problem. The first half — ten realized 14FYP cascades — is documented here. The second half — seven falsifiable predictions on the 15FYP — is committed below, with resolution windows running from 2027 through 2029.

Four moving parts. Each one shipped in v0.

The 14FYP text, released on gov.cn on 13 March 2021, together with the ten documented ministerial follow-ons that make up the 14FYP cascade record. Each cascade carries a plan line, the realized ministerial document with its document number where one exists, the realized subsidy band, the lag in months from plan issuance, and the ranked realized beneficiary sectors. The initial scope is deliberately the 14FYP only; the 13th plan and earlier land in a later release alongside the full historical re-ranking.

Each input excerpt is encoded with a TF-IDF vectorizer fit over the historical-cascade excerpt corpus, using one- and two-word ngrams with sublinear term-frequency weighting. Cosine similarity selects the nearest historical precedents. A keyword-rule predictor — the current stand-in for an eventual long-context language-model call — emits a structured prediction with ranked beneficiary sectors, a subsidy band, a predicted ministerial lag, and a citation pack tying each call back to either the subject excerpt or one of the retrieved precedents. Swapping in the language model is the only architectural change planned between this release and the next.

Three baselines run alongside on every backtest fold. Random: uniform sector draw. Mode: always the highest-empirical-weight sector in the 30-sector taxonomy. Keyword match: count taxonomy-keyword hits inside the excerpt, return the top-N by hit count. None are removable. If the retrieval pipeline is not lifting over the keyword baseline, the scoreboard says so before we do.

Recall at top-3 for the realized beneficiary sector set. Mean absolute error on the lag prediction, in months. Subsidy-band accuracy as a strict equality across a five-class schema (under 1 billion, 1–5 billion, 5–20 billion, more than 20 billion, and non-monetary, in USD-equivalent). All three metrics are evaluated leave-one-out across the ten 14FYP cascades. A calibration plot via reliability diagram is deferred to the next release, once the live predictions begin resolving in late 2028.

What v0 sacrifices, stated up front.

Training set of ten.The retrieval index is exactly ten 14FYP cascades. Leave-one-out cross-validation against an n=10 corpus is statistically thin: any single recall@3 measurement carries a confidence interval wider than the difference between baselines. We report point estimates because that is what is available; v0.5 expands the index to the 13FYP and 12FYP cascades (target n ≈ 30), and v1 expands further with the sub-plan layer. Treat the backtest as a discipline check, not as a calibration claim.

Placeholder text in five plan excerpts. Five of the eleven excerpts in the current prediction registry use an institutional-voice paraphrase rather than the verbatim 14FYP text, and are tagged accordingly. The plan-line semantics are accurate; the surface tokens are not yet a primary-source quote. The next release ingests the verbatim text from gov.cn and replaces the placeholders excerpt by excerpt.

Truncation on the back end.The 14FYP cascade record runs through end-2024 in this snapshot. Ministerial follow-ons that land in 2025–2026 (the final two years of the 14FYP cycle) are excluded from the backtest gold set — not because they don't matter, but because the lag observation is right-censored for cascades that haven't fired yet. The cleanest example is the 14FYP biopharma sub-plan: NRDL inclusion continues year-over-year, and the realized 2025–2026 additions will only enter the gold set on the next refresh. v0.5 reruns the backtest against an end-2026 snapshot.

Confidence translation is hand-tuned. The mapping from qualitative labels (high, medium-high, medium) to numeric point estimates with 80 percent confidence intervals is a five-row lookup table, not a calibrated probability. We default to under-confidence: under-confidence is preferable to miscalibrated certainty. The first reliability diagram lands once the late-2028 resolution window closes for the sovereign-AI and climate-energy predictions.

Sectors excluded from the prediction set. Biotech and defense are excluded from the live 15FYP predictions: biotech because the marginal prediction beyond the 14FYP continuation is hard to falsify cleanly, and defense and dual-use on editorial-guardrail grounds. Reading the cascade record without these sectors is a real omission, not a clean methodological choice, and remains an open question for the program.

What we will and will not publish.

Primary sources. Every cited 14FYP plan line resolves to gov.cn, the issuing ministry, the NPC archive, or a State Council bulletin. Press estimates appear with explicit attribution and an unverified tag where a primary fiscal document does not yet exist.

Falsification, not narration.Every live prediction carries a criterion written so resolution does not require subjective adjudication. We score against the criterion, not against an analyst's memory of what we meant.

De-compilation is data.The 14FYP common-prosperity record documents quiet retraction. A plan line that compiles and then de-compiles is a recoverable signal, not an error in the corpus — the prior should encode it as a distinct outcome class.

No fabricated magnitudes. Where a subsidy quantum is asserted by secondary press but not yet anchored to a primary fiscal document, the field carries the press attribution and the unverified tag. We do not interpolate.

Excluded scope. Defense and dual-use cascades are out of the public prediction set on editorial-guardrail grounds. Biotech is documented as a continuation cascade only.