How the system keeps itself honest
Research chain
Observation → surprise → marginal change → decomposition → mechanism → cross-check → research view → asset implication → evidence test. Every interpretation answers ten questions, from "compared with what?" to "what print would prove this wrong?"
Four lenses keep the wording honest. Level: where an indicator stands. Marginal change: which way it is moving now, which cumulative series can hide. Breadth: whether the move is shared. Mechanism: what drives it. A claim that mixes them is rewritten before it is published.
From data to thesis
1 · Observations
Append-only rows with release date, retrieval time and source class. Unknown release dates are recorded as unknown, never guessed.
2 · Claims
Numbers enter prose only as bound references to observations or derived values. Each claim is typed Fact, Derived, External view or Inference, and exists in both languages with the same bindings.
3 · Thesis
19 evidence tests move hypotheses; preregistered terminal rules move the thesis. Probability changes in 5-point steps, always with reasons.
Automatic refresh vs analyst judgment
A scheduled job checks the official release calendar twice a day. When a release is due it fetches the new data from the official source, appends only genuinely new observations (a revised value is kept as a new vintage), re-runs every evidence test against the preregistered case, records which tests flipped, runs the eval below and rebuilds this site. Running it again with nothing new does nothing.
Updated automatically
Observations, derived values, evidence-test outcomes, the terminal-rule check, charts and a factual record of what changed. A preregistered terminal rule is the only automatic way the thesis status can change.
Changed only by the analyst
Probability and confidence. After a release the analyst appends an update saying whether it changes the thesis; any probability change must cite evidence and reasoning or the eval fails. The refresh only offers a mechanical tally of the direction of test flips.
Latest refresh 9 Oct 2026, 10:36 UTC: nbs reachable.
Where AI is used, and where it is forbidden
Three jobs are kept apart. Deterministic software establishes what happened. An AI Research Copilot interprets what it means. The analyst decides what to believe. The Copilot is in shadow mode: every run is recorded for audit and evaluation, and nothing it writes is shown on this site as the research view unless the analyst accepts it.
| Task | Done by | Checked by |
|---|---|---|
| Fetching and calculating data, vintages and as-of dates, evidence-test outcomes, terminal rules, and finding which observations and tests changed | Software | Eval gate E0 to E5, unit tests, historical replay |
| Writing numbers into text | Software. The AI writes only references to evidence; code fills in the values | A closed list of evidence ids for each run; any figure typed by the model fails validation |
| Judging whether a change is material, mapping it to the hypotheses, naming the strongest competing explanation, what the evidence does not prove, and the next research questions | AI Research Copilot: two bounded calls, a draft and then a challenge | Hard validation of every output, then analyst review |
| Probability, conviction and every change to the thesis | Analyst | Thesis discipline check E5; append-only, time-stamped updates |
What the AI never does
- Type a data figure, or recompute an indicator in prose
- Decide whether a terminal rule has fired, or change a preregistered rule or the mechanical direction of an evidence test
- Output or change a probability or a conviction
- Cite evidence outside the list the software supplies, or anything dated after the information set
- Publish. An answer that fails validation is discarded, and a run without a complete record is not used
Copilot v0 is in shadow mode. A historical replay tests interpretation under a given framework, not the framework's predictive power, and is never presented as live evidence.
Deterministic eval
Generated 9 Oct 2026, 10:37 UTC · commit fbb0cb8 · public mode
| Check | What it proves | Status |
|---|---|---|
| E0 Data quality | Schema, duplicates, period formats, provenance or a stated limitation for every row cn_cgb10y: not in public build (private_only); private mode only; cn_gov_fund_rev_ytd_yoy: no observations … |
! warn |
| E1 Citation integrity | Every number in every claim equals the data as of the claim's date; both language versions of a claim bind the same numbers and name the same periods | ✓ pass |
| E2 Derived calculation lineage | Every derived value recomputes from observations that exist | ✓ pass |
| E3 As-of leakage | Nothing is cited before it was released | ✓ pass |
| E4 Preregistration immutability | Baseline, amendments, updates and refresh records were committed once and never edited; amendments precede the first tested release 2026-10-09T1036Z.json: not committed yet |
! warn |
| E5 Thesis update discipline | Probability on a 5-point grid; every change explained; terminal status only via rules | ✓ pass |
These checks are deterministic code. A model grading its own research is not treated as evidence of correctness.
Known data limitations
- GDP history is the current revised vintage, not first prints; first prints are captured going forward.
- Release dates are official (NBS release agenda) for NBS series. PBOC, FRED and agency series republished on the NBS portal carry no first-release date, so their availability is the time they were retrieved.
- The China 10-year government bond yield has no redistributable public source and is not part of the public build. Its licensed source (Wind) is used only in the researcher's local Private Research Mode; nothing derived from it is published, and this public site runs on official and open sources alone.
- Government fund revenue has no reachable public route (the Ministry of Finance release host refuses automated access), so that diagnostic is absent.
- Exports to the US are missing for 2024-09 and 2024-10, which the customs English bulletin never posted.
- The NBS data portal can trail the NBS press release by several days; the refresh picks a release up once the portal carries it.
- Year-to-date indicators smooth recent changes. Where it matters, the site reads the direction of the cumulative rate to infer the latest single month, and says so.
- Policy events are curated by hand from official sources and are context only. They are not evidence tests, and they do not alter the preregistered baseline.
Data sources and boundary
Private Research Mode consumes Wind and licensed research through a local-only V1.2 Phase 1 integration and runs private diagnostics on top of it (Phase 2, also local-only); it is not required for the public V1. This public site is built only from official and open sources and tracked research artifacts. A claim that depends on licensed data shows as withheld rather than disappearing.
| Source | Status |
|---|---|
| NBS National Bureau of Statistics | active |
| PBOC People's Bank of China | active |
| GACC General Administration of Customs | active |
| FRED Federal Reserve Bank of St. Louis (FRED) | active |
| Wind Wind | private / local only Private Research Mode consumes Wind and licensed research through a local-only integration (private snapshots and private diagnostics on the researcher's machine). Wind data is never published, and the public V1 does not depend on it. |
Metric budget
23 core metrics, evaluated in every run, and 6 diagnostics, pulled only when a core signal needs decomposition.
| Metric | Tier | Serves | Why it is here | Source |
|---|---|---|---|---|
| Nominal GDP, single quarter, current prices 国内生产总值（当季，现价） | core | H1, H2 | Target variable. With real GDP it gives nominal growth and the implied deflator. | National Bureau of Statistics |
| Real GDP YoY, single quarter 国内生产总值不变价当季同比 | core | H1, H2 | Denominator of the implied deflator; real side of nominal growth. | National Bureau of Statistics |
| PPI YoY 工业生产者出厂价格指数同比 | core | H1, H0 | Monthly proxy for the industrial deflator; the main high-frequency price signal. | National Bureau of Statistics |
| PPI MoM 工业生产者出厂价格指数环比 | core | H1, H0 | Momentum free of base effects; distinguishes a real turn from a base-effect turn. | National Bureau of Statistics |
| CPI YoY 居民消费价格指数同比 | core | H1 | Consumer-side price level; part of the GDP deflator. | National Bureau of Statistics |
| Core CPI YoY (ex food and energy) 扣除食品和能源价格后的核心CPI同比 | core | H2, H0 | Demand-driven price pressure; key discriminator between H2 (demand recovery) and H0 (supply-only). | National Bureau of Statistics |
| Manufacturing PMI, output prices 制造业PMI出厂价格指数 | core | H1 | Leads PPI MoM by roughly one release; earliest price read each month. | National Bureau of Statistics |
| M1 YoY M1同比 | core | H2 | Corporate and household transaction balances; a leading read on private activity. | People's Bank of China |
| New RMB loans to households, medium and long term, year-to-date 住户中长期贷款新增（年初至今） | core | H2 | Mortgage demand; the cleanest private-credit read on households. Use single() for monthly flow. | People's Bank of China |
| Retail sales of consumer goods YoY 社会消费品零售总额当月同比 | core | H2 | Nominal household consumption. | National Bureau of Statistics |
| Commodity building sales area, YTD YoY 商品房销售面积累计同比 | core | H2 | Property demand; the largest single drag on private demand since 2021. | National Bureau of Statistics |
| Industrial enterprise profits, YTD YoY 规模以上工业企业利润总额累计同比 | core | H1, H2 | Where price and volume meet margins; tests whether reflation reaches corporate income. | National Bureau of Statistics |
| Manufacturing PMI, new orders 制造业PMI新订单指数 | core | H2 | Earliest monthly demand read. | National Bureau of Statistics |
| Total social financing stock YoY 社会融资规模存量同比 | core | H3, H2 | Aggregate credit growth across loans, bonds and government debt. | People's Bank of China |
| TSF, government bond net financing (monthly flow) 社融口径政府债券净融资（当月） | core | H3 | Fiscal credit impulse; separates government-led from private credit growth. | People's Bank of China |
| General public budget expenditure, YTD YoY 一般公共预算支出累计同比 | core | H3 | Spending side of fiscal impulse. | Ministry of Finance (via the NBS portal) |
| Exports in USD, YoY 出口金额（美元计）当月同比 | core | H4 | External demand contribution to nominal growth. | General Administration of Customs |
| Manufacturing PMI, new export orders 制造业PMI新出口订单指数 | core | H4 | Leading read on export momentum. | National Bureau of Statistics |
| Brent crude, monthly average 布伦特原油价格（月均，美元/桶） | core | H0, H1 | Imported cost-push into PPI; a supply-side price driver (H0). | Federal Reserve Bank of St. Louis (FRED) |
| Copper, monthly average 铜价（月均，美元/吨） | core | H0, H1 | Global industrial-metals price driver of upstream PPI. | Federal Reserve Bank of St. Louis (FRED) |
| CNY per USD, monthly average 美元兑人民币汇率（月均） | core | H3 | Fed/USD -> CNY channel; constraint on PBOC easing room. | Federal Reserve Bank of St. Louis (FRED) |
| US 10-year Treasury yield, monthly average 美国10年期国债收益率（月均） | core | H3 | US rate level; with CGB 10Y gives the China-US spread that binds PBOC policy. | Federal Reserve Bank of St. Louis (FRED) |
| China 10-year government bond yield, month-end 10年期国债到期收益率 | core | H1, H2, H3 | Market-implied view of nominal growth; benchmark our thesis against the bond market. | Wind |
| PPI, producer goods YoY PPI生产资料同比 | diagnostic | H0, H1 | Pull when PPI moves; upstream-led moves point to supply or commodity drivers (H0). | National Bureau of Statistics |
| PPI, consumer goods YoY PPI生活资料同比 | diagnostic | H0, H2 | Pull when PPI moves; downstream pass-through suggests demand (H2) rather than supply. | National Bureau of Statistics |
| Industrial value added, real YoY 规模以上工业增加值当月同比 | diagnostic | H0, H2 | Pull when prices rise; rising prices with falling output fits supply discipline (H0). | National Bureau of Statistics |
| New RMB loans to non-financial corporates, medium and long term, YTD 企（事）业单位中长期贷款新增（年初至今） | diagnostic | H2, H3 | Pull when TSF moves; separates capex credit demand from government bond issuance. | People's Bank of China |
| Government fund budget revenue, YTD YoY 政府性基金预算收入累计同比 | diagnostic | H2, H3 | Pull when fiscal or property signals conflict; land sales link property weakness to local fiscal capacity. | Ministry of Finance (via the NBS portal) |
| Exports to the United States in USD, YTD YoY 对美出口（美元计）累计同比 | diagnostic | H4 | Pull when exports move; isolates tariff-exposed demand from rerouting to other markets. | General Administration of Customs |