Five-Year Plan Cascade Analyzer.
China's Five-Year Plans are the most consequential industrial-policy documents on Earth — but the plan text itself only designates broad sectors. The actual money flows in afterwards, through NDRC capital pools, MIIT specialized funds, MoF tax notices, and provincial supplements that land sequentially across the plan period. This analyzer predicts, in advance, which sectors will receive that ministerial follow-on, at what magnitude, and how soon. We trained on the 14th Five-Year Plan (2021–2025) record — ten documented cascades worth hundreds of billions of dollars in cumulative subsidy — and apply the model forward to the 15th (2026–2030, adopted at the NPC on 12 March 2026). Seven live predictions are on the board: sovereign AI compute, semiconductor self-sufficiency, EV consolidation, embodied AI, common-prosperity v2, climate / energy transition, and quantum continuation. Each carries a falsification criterion and a resolve-by date in 2027–2029. The investable signal sits in the gap between plan text and ministerial follow-on.
What we've actually found.
Keyword overlap, not the fancy retriever, is the bar to beat on a ten-cascade corpus.
A plain keyword baseline recovers more than half of the realized beneficiary sectors at top-3. A richer character-level retriever recovers almost none. The honest reading: on n=10, the next lift has to come from a model that understands meaning across languages, not from richer surface text.
Quantum belongs on the prediction list. The audit reopened a cascade we had previously left off.
v0 had treated quantum as a documented counter-example because no Big-Fund analog exists. The v0.1.1 audit re-examined the case and concluded that the 13FYP-to-14FYP chapter-level continuity is itself a falsifiable bet for the 15th plan at state-lab magnitude.
Ten documented 14FYP cascades; seven falsifiable 15FYP predictions logged in public.
From Big Fund III (RMB 344bn for semiconductors) to the NEV purchase-tax extension to East-Data-West-Compute, we trace each cascade from a single plan line to its ministerial follow-on. Then we apply the same machinery forward to the 15th plan with resolve-by dates in 2027-2029.
The ten documented 14FYP cascades, the seven live 15FYP predictions with their falsification criteria, the held-out backtest against four baselines, the pipeline architecture, the editorial guardrails, and the open questions all live inside the v0 founding paper.
Published pieces.
Each card is a standalone read — a substantive finding or a release. Methodology hygiene, propagation fixes, and audit notes live in the Changelog below.
Why a richer English-keyword model made Chinese-policy retrieval worse.
We taught the analyzer to look at policy excerpts at the character level — and confirmed it didn't help. The honest negative result tells us that the real lift, when it comes, has to come from a model that understands meaning across English and Chinese, not richer surface text.
The first published null result in the program. It rules out a whole class of cheap fixes and locks the v0.5 roadmap onto bilingual semantic retrieval — which is where the policy-reading edge actually lives.
We were a year wrong on the most important date in Chinese industrial policy.
A primary-source audit caught a 12-month error in the anchor date for every live prediction, and a second pass reopened the quantum cascade we had previously left off the board as a continuation bet.
Both fixes change what's on the prediction list — by twelve months on six predictions, and by one whole sector. A forecasting program that can't get its own anchor date right is unreadable; we caught it before the first resolution window opens.
Where China's next half-trillion in industrial subsidy is going.
The founding paper of the analyzer. We document ten realized cascades from the 14th Five-Year Plan, log seven falsifiable predictions on the 15th, and ship the pipeline that produced both — with every plan line, ministerial follow-on, and source URL visible on the page.
China's Five-Year Plans move hundreds of billions of dollars across sectors. The plan text designates the sectors; the money lands afterwards through ministerial programs. We forecast the second step before each program is announced — and we will be marked right or wrong at named resolution dates in 2027-2029.
Methodology, propagation, audits.
Smaller fixes — page/code reconciliation, calendar arithmetic, engineering — that aren't publishable on their own but are recorded here for traceability.
- 31 May 2026Details →Calendar arithmetic, fixedThe pipeline was computing resolution dates by adding 30-day chunks, which drifted by a day or two at the two-year horizon. Switched to true calendar-month arithmetic so every resolve-by date lands cleanly on the NPC adoption day.
- 31 May 2026Details →Pipeline caught up to the pageThe v0.1.1 audit added a new prediction and shifted six others by twelve months on the page. v0.1.2 propagates those edits into the Python pipeline so a fresh run reproduces the published predictions instead of reverting them.