Domain A - China Industrial Activity vs. Satellite Reality

Domain A — Results: China Industrial Activity vs. Satellite Reality

Computed 2026-08-04 against hypothesis.md (commits d918279, 118e5a7, pre-registered before any value was fetched). Every number below is reproducible: python -m src.domains.domain_a_china.compute regenerates results.json from the staged fixtures in data/, deterministically, with no manual or LLM arithmetic in the computation path.

H1 — National divergence index: weak positive, technically not falsified

• Correlation between z(official) and z(composite satellite) = 0.113, across 20 overlapping months (2024-06 through 2026-05; 2026-06 excluded because satellite YoY wasn't yet computable for it).
• The pre-registered falsification criterion (r ≤ 0) was not met, so H1 survives — but r = 0.11 is weak. This should not be read as confirming a usable "official tracks observed" relationship. At this sample size, a correlation this size is barely distinguishable from noise. Do not build a trading or monitoring signal on this number as it stands.
• One outlier exceeding the pre-registered ±2σ divergence threshold: 2025-03 (divergence z = +2.97). Official industrial YoY growth spiked to 7.7% — the highest print in the series — while the composite satellite signal was mildly negative that month (z = −0.34). 2026-03 shows the same directional pattern, more mildly (divergence z = +0.78, not an outlier).
• Likely explanation (flagged as a caveat, not proven): NBS does not publish standalone January/February industrial figures at all — confirmed empirically in this same dataset (see Antithesis, CNY check below) — and resumes in March with a print that plausibly reflects catch-up/base effects from the two omitted months. Satellite inputs have no equivalent reporting gap and continue producing ordinary monthly readings through Jan/Feb. The March comparison may therefore be comparing NBS's compressed catch-up figure against satellite data that never had an equivalent gap to catch up from — a structural artifact of Chinese official reporting conventions, not necessarily new information from satellite data. This reading is directionally consistent across both March observations in the window, which is suggestive but is still only n=2.

H2 — Yunnan dry-season signal: dip present, not falsified, and distinct from a national pattern

• Yunnan's dry-season seasonal index averages 0.962; wet-season averages 1.098 (dip magnitude −0.137) — Yunnan's radiance runs about 14% below its own trailing 12-month baseline in dry-season months (Nov–Apr) and about 8% above baseline in wet-season months (May–Oct). A real, sizeable seasonal swing, computed over 41 dry-season and 37 wet-season monthly observations spanning 2019-12 through 2026-05.
• National antithesis check: dry season 1.039 vs wet season 1.022 (magnitude +0.018) — essentially flat, no dip nationally, over the same 41/37 month split. This is the key result: the seasonal dip is not a generic China-wide winter lighting pattern — it is specific to (or at least far stronger in) Yunnan, which is consistent with, though it does not prove, the hydropower-curtailment story.
• Non-ferrous smelting, weak secondary leg as flagged in advance: dry-season mean YoY growth 4.96% vs wet-season 6.13% (n=8 dry, n=13 wet months only — the official series' short window limits this badly). Direction is consistent with the curtailment story, but this is a national, multi-metal aggregate blending every province and every non-ferrous metal, on a small sample — suggestive at most, not evidence on its own.
• H2 is not falsified. The core physical prediction — a measurable Yunnan-specific dry-season dip — held up against the pre-registered test and against the antithesis check that could have undermined it (a China-wide pattern would have falsified the "Yunnan-specific" framing; it didn't appear).

Antithesis checks performed

• Composition effects. Correlation between headline industrial YoY and non-ferrous-smelting YoY over their 21 shared months: r = 0.425 — moderate positive co-movement, not near 1.0. Non-ferrous metals are one contributing sector among many to the headline number, not a dominant driver of it. A headline divergence in H1 cannot be attributed to the non-ferrous/aluminium sector specifically on this evidence.
• Chinese New Year timing. The official series has zero January or February observations across its entire available history — NBS simply doesn't publish them standalone. This structurally removes the sharpest form of CNY-driven distortion from H1's comparison (there's no Jan/Feb official print to compare against a satellite reading in the first place), but it is very likely the underlying cause of the March catch-up effect noted in H1 above — the same distortion re-surfacing one month later in a different shape, not eliminated so much as displaced.
• Real vs. nominal / base effects. Verified: compute.py reads only the two "Growth Rate (Year-over-year percentage change)" series (ef1b1765960d45a29b4d7c4ca91be916, 5d9322cf50b44f9c937ac60a77ce0c29) and never the sibling "Accumulated Growth Rate" series — enforced by which series_ids were staged into data/, not by a runtime check.
• Nighttime lights is not aluminium-specific — unresolved confound. Yunnan's dry-season dip could reflect winter tourism patterns, Spring Festival travel, weather, or any other regional seasonal effect, not smelter curtailment specifically. Nothing in this analysis rules that out; the antithesis check above only rules out one alternative explanation (a generic national pattern), not this one.
• Mirror-statistic check: not applicable, as declared in advance — this is a single-country domestic-vs-satellite comparison, not a bilateral trade flow with a counterparty ledger.
• FactIQ vintage caveat. Every official NBS value used here is FactIQ's current, revised print (Phase 1 finding, reconfirmed by FactIQ directly — see findings/01-vintage-probe.md), not what was originally published for that month. A real limitation for a domain framed as "official vs. observed."

Data-quality finding surfaced during staging (not anticipated in the hypothesis)

Both nighttime-lights series' catalog metadata claimed begin_time: 2024-01-01. The underlying data_points table actually has complete monthly coverage back to 2019-01-01 (89 of 89 possible months, no gaps) — confirmed via COUNT(*)/MIN/MAX against the raw table, not the series catalog. This is a sharper version of the Phase 0 finding that dataset-level metadata can be stale ("trust series.end_time, not dataset metadata") — here it extends to series-level begin_time being simply wrong. Practical effect: this project actually had ~5 more years of usable night-lights history available than the pre-registered hypothesis assumed, which is why H2's seasonal-index test could use 7 dry seasons' worth of data instead of the 2 originally anticipated. Anyone building further on FactIQ's satellite_nightlights catalog metadata should query data_points directly rather than trust begin_time/end_time on the series row.

Caveats (restated from hypothesis.md; all still apply)

• The H1 overlap window is short (20 months) — too short for robust statistical inference from a single correlation number. r = 0.11 is a first structural pass, not a validated relationship.
• Non-ferrous-metals-smelting is a national, multi-metal aggregate, not an aluminium-only or Yunnan-specific series — the closest available proxy, not a match. The composition check above quantifies how loose that proxy is (r = 0.425 to the headline it's meant to help explain).
• Yunnan reservoir levels, the originally intended direct lead indicator, remain confirmed absent from FactIQ (Phase 0) — H2 is necessarily an indirect test and cannot become a direct one with this data source.
• NO2 data (used nationally in H1, staged but not used for Yunnan in the computed results) carries real cloud-cover gaps — Yunnan NO2 specifically had only 33 of 43 possible months with any usable reading, many below FactIQ's own 0.2 valid-obs-share confidence threshold. Not a blocker here since Yunnan NO2 wasn't used in either hypothesis's final computation, but a real constraint on any future extension that wants a smelting-specific emissions signal for Yunnan directly.
• FactIQ's official NBS data is not vintage-aware (Phase 1, reconfirmed by FactIQ directly) — every historical month used here is today's revised print, not what was known at the time.

Bottom line

• H1: not falsified, but weak. A same-month "official industrial growth tracks a satellite composite" relationship is not well supported by the available ~20 months of data (r = 0.11). One likely-explainable outlier (2025-03) accounts for a large share of what correlation exists. This is a structural first pass, not a usable signal in its current form.
• H2: not falsified, and more interesting. Yunnan shows a real, seasonally-recurring radiance dip that the rest of China does not show over the same period — consistent with, though not proof of, the dry-season hydropower-curtailment story this domain set out to test. The one confound this analysis cannot rule out (non-curtailment regional seasonality specific to Yunnan) remains open, and the direct test that would actually settle the mechanism — reservoir levels — is unavailable through FactIQ, now and for the foreseeable future per their own reply (see findings/01-vintage-probe.md addendum).
• Both findings should be read as a first structural pass on genuinely short official-data history, not validated models — exactly as flagged in the pre-registered caveats before any value was pulled.

Pre-registered hypothesis

Domain A — China Industrial Activity vs. Satellite Reality

Status: PRE-REGISTERED, 2026-08-04. No comparison values have been pulled at the time this file is committed — series existence and coverage windows only (metadata, not results). Per project principle, this hypothesis will not be edited after results are computed; if it fails, the failure is recorded, not the hypothesis.

Data sources (exact identifiers confirmed present as of 2026-08-04)

Official (FactIQ china schema, dataset nbs, National Bureau of Statistics of China):

• National industrial value-added, YoY growth: ef1b1765960d45a29b4d7c4ca91be916 — "Value-added of Industrial Enterprises above Designated Size, Growth Rate (Year-over-year percentage change)". Monthly, %. Coverage per series metadata: 2024-06 to 2026-06.
• Non-ferrous metals smelting value-added, YoY growth: 5d9322cf50b44f9c937ac60a77ce0c29 — "Value-added of Smelting and Pressing of Non-ferrous Metals, Growth Rate (Year-over-year percentage change)". Monthly, %. Same coverage. This is the aluminium-sector proxy. NBS does not publish an aluminium-only value-added series; this code also includes copper, lead, zinc, and other non-ferrous metal smelting. It is the closest available official aggregate, not a match.

Satellite/observed (FactIQ satellite schema, dataset satellite_nightlights, NASA Black Marble):

• National China nighttime lights, mean monthly radiance: sat:ntl:CHN:mean_radiance. Coverage: 2024-01 to 2026-05.
• Yunnan Province nighttime lights, mean monthly radiance: sat:ntl:CHN:yunnanprovince:mean_radiance. Same coverage.
• FactIQ's own dataset documentation instructs comparing the same calendar month year-over-year rather than month-to-month, because moonlight/ atmospheric correction makes short-run wiggles noisy. This methodology follows that guidance directly (YoY radiance change) — which also happens to match the official series' own YoY framing.

Air quality (FactIQ satellite schema, get_geo_data, live Sentinel-5P aggregation, no series_id — fetched by region/date range):

• NO2 tropospheric column, region CHN (national) and CHN/Yunnan (province), monthly aggregation. Coverage: 2018-> , ~3 day lag. Used as a secondary corroboration signal only — NO2 is a general combustion/ traffic/industrial proxy, not aluminium-smelting-specific.

Shipping (FactIQ portwatch schema, dataset portwatch_countries):

• China national daily maritime export volume: pw:country:CHN:export. Daily, 2019-01-01 to present. Used only as a tertiary corroboration input to the national composite (Yunnan is landlocked — no port-level Yunnan proxy exists).

H1 — National divergence index

Official monthly industrial value-added YoY growth and a satellite-derived composite industrial-activity YoY growth (built from national night lights, national NO2, and national port export volume) should be positively correlated over the overlap window — 2024-06 through 2026-05, ~24 monthly observations, bounded by the satellite series' 2024-01 start plus the 12-month YoY lag it requires.

Method (deterministic — implemented in compute.py, no manual or LLM judgment at compute time):

1. For each of the three satellite/observed inputs, compute month-over- same-month-prior-year % change.
2. Z-score each input's own YoY-change series, using that series' own mean/stdev over the overlap window (z-score against own history — the same convention already used project-wide for the drift check and the planned scout agent).
3. Composite satellite signal = simple average of the three z-scores per month. Equal weighting, declared here in advance, not tuned after seeing results.
4. Z-score the official industrial value-added YoY series the same way, over the same window.
5. Divergence index = z(official) − z(composite satellite), by month.

Falsification: if the correlation between z(official) and the composite satellite z-score is ≤ 0 over the full window — no positive relationship at all — H1 is falsified. The "official vs. observed" comparison this domain is built on does not hold even directionally, and that will be reported as a failure, not adjusted.

H2 — Yunnan dry-season hydropower curtailment signal (aluminium overlay)

Method note (amended 2026-08-04, before any data was fetched — see note at end of section): uses a deseasonalized radiance ratio, not YoY growth.

The specific hydropower reservoirs needed to observe curtailment directly are confirmed absent from FactIQ (Phase 0 finding). This tests the necessarily weaker, indirect prediction instead: Yunnan's nighttime radiance should sit measurably below its own trailing 12-month baseline during the dry season (defined a priori as November–April, the conventional Yunnan hydropower low-water period) compared to non-dry-season months. If the curtailment story holds, national non-ferrous-metals-smelting value-added YoY growth should also tend to be relatively weaker in dry-season months — though this leg is expected to be weak and noisy, since the national series blends Yunnan with every other province and every other non-ferrous metal. That expectation is stated here, in advance, not fitted after the fact.

Method (deterministic, in compute.py):

1. For each month of Yunnan radiance, compute a seasonal index = that month's radiance ÷ the trailing 12-month mean radiance ending that month (falls back to the longest available trailing window at the start of the series, logged explicitly, never silently padded).
2. Classify each month as dry-season (Nov–Apr) or wet-season (May–Oct).
3. Compare the mean seasonal index across dry-season months to the mean across wet-season months, over however many months of each are available in the window.

Why not YoY growth, as H1 uses: a YoY change in radiance compares a month to the same month one year earlier. If Yunnan curtails hydropower in the same months every year, both the numerator and its year-ago denominator are equally suppressed, and YoY growth would show no dip at all even if a real, stable seasonal pattern exists — YoY only reveals changes in curtailment severity between years, not the existence of the seasonal pattern itself. The deseasonalized ratio above is the direct test for a recurring dip; it is used here instead, and only here — H1 keeps its YoY-based method because H1 is explicitly comparing against an official series that is itself already YoY.

Falsification: if Yunnan's dry-season seasonal index is not measurably lower, on average, than its wet-season seasonal index, across the two dry seasons available in the window (Nov 2024–Apr 2025, and Nov 2025–Apr 2026 partial, bounded by data through 2026-05), H2's core physical claim is not supported by this data and will be reported as such. There is no reservoir data to fall back on — this is the whole test.

Amendment note: the original committed version of this section (commit d918279) specified a YoY-growth-based test for H2, which — on review before any data was touched — does not actually test for a recurring seasonal pattern, for the reason given above. This correction is made before any comparison value has been fetched (only series metadata had been queried at that point); it is a methodological fix caught in review, not a post-results adjustment, and is being committed as a visible amendment rather than silently folded into the original text.

Antithesis — required counter-checks before any result is called a finding

• Composition effects. National industrial value-added blends many sectors; a move can come from anywhere, not the non-ferrous/aluminium sector specifically. Check whether the non-ferrous-smelting series moves with or against the broader industrial aggregate before reading anything into it alone.
• Chinese New Year timing. CNY falls in January or February depending on the lunar calendar and shifts industrial activity between those two months every year. The window here spans at most two CNYs (Feb 2025, ~Feb 2026) — any Jan/Feb divergence must be checked against the actual CNY date for that year before being read as a real signal rather than a calendar artifact.
• Real vs. nominal / base effects. NBS "Growth Rate (Year-over-year percentage change)" series are real (volume) measures per NBS methodology. The "Accumulated Growth Rate" variant (not used here) is measured against a shifting year-to-date base and is not comparable month to month — compute.py must only ever read the YoY series, never the Accumulated one, and this will be asserted in code, not just stated here.
• Nighttime lights is not aluminium-specific. It reflects any change in electrified activity in a province — consumer, residential, unrelated industrial — not smelter curtailment specifically. A Yunnan dip could reflect tourism, weather, or any other regional shock. This is a genuine confound the available data cannot rule out, only flag.
• Mirror-statistic check: not applicable. This is a single-country domestic-vs-satellite comparison, not a bilateral trade flow with a counterparty ledger — flagged as not applicable rather than silently skipped.
• FactIQ vintage caveat. The official NBS series pulled through FactIQ carries no vintage guarantee (Phase 1 finding). Every value used here is FactIQ's current, revised print, not the figure as originally published. Real limitation for a domain framed as "official vs. observed" — stated here and restated in the results.

Caveats (known before results; will be restated in the results writeup)

• The overlap window is short (~24 months) — too short for robust statistical inference from a single index number. Results are a first structural pass, not a validated model.
• Non-ferrous-metals-smelting is a national, multi-metal aggregate, not an aluminium-only or Yunnan-specific series — the closest available proxy, not a match.
• Yunnan reservoir levels, the originally intended lead indicator, are confirmed absent from FactIQ (Phase 0) and cannot be added here.
• NO2/night-lights data carry their own latency and cloud/atmospheric noise (Phase 0). Comparisons follow FactIQ's own YoY guidance.
• FactIQ's official NBS data is not vintage-aware (Phase 1) — this comparison uses today's revised print for every historical month, not what was known at the time.