china_industrial_reality

Physical validation — two rounds, honest result

Round 1 (Zaragoza): tried the satellite method (Sentinel-1/2, no LiDAR) against the CATL–Stellantis gigafactory, first-stone 2025-11-26, with a success bar fixed before running. It failed — no sustained signal at the site, and the control zone drifted more than the site in one quarter, ruling out the site's isolated spikes as real. Full write-up: zaragoza_b05_conclusion.md.

Round 2 (4-site battery, pre-registered verdict rule): before declaring the method dead, we ran it against 3 more sites — Debrecen (CATL) and Szeged (BYD, the actual candidate assets), plus Tesla's Gigafactory Berlin-Brandenburg (Grünheide) purely as a method sanity check — alongside Intel's cancelled Magdeburg site as a negative control (construction never started there; the method must say so).

But Debrecen and Szeged — the two candidate assets that would actually feed this product — still show no clear sustained signal, same as Zaragoza. The honest read: industrial construction on already-disturbed or agricultural land is a much smaller spectral/radar contrast than clearing a forest, and all three candidate sites still use a synthetic buffer polygon, not the real parcel — likely diluting whatever signal exists. Full matrix and numbers: b05_multisite_conclusion.md.

Consequence: the gate is no longer blocked on "maybe the method is broken" — it's proven capable. But no asset below has a real MEASURED result to publish yet, because none of the actual candidates has shown one. This page stays DOCUMENTED-only until a real parcel boundary is tried on Debrecen or Szeged.

Two more sensors tried, both closed honestly dead

Beyond the optical/SAR method above, two more free public sensors were tested as additional ways to detect real operation or buildout — both came back a clean negative, reported the same way as everything else on this page: pre-registered threshold, no reinterpreting after seeing the data.

New MVP — regulatory/contract signal, blind-tested 5/5, now 13/13

Remote sensing (above) is dead on real candidate parcels for this asset class. A different method isn't: real, dated public signals — municipal/regional interest declarations, strategic-project status, public-consultation periods, resolved regulatory friction, and construction contracts — tracked per asset instead of per pixel. Before trusting it, it was blind-tested against 5 independent cases this method had not yet touched: for each one, a prediction was written and committed to source control before looking at current employment/municipal data, then compared honestly. Full write-ups: b11_alternative_signals_findings.md, b12_calb_sines_blind_result.md, b13_two_more_blind_result.md. With the method holding up 5/5, it was then applied directly (no longer blind) to the remaining 7 assets in the audit table below — all 13 are now covered.

Now populated for all 13 catalog assets — 6 from the original blind tests, 7 from the direct extension:

Code and the 12 automated checks (including the anti-conflation rule that a park-infrastructure contract can never register as a factory-building contract, and the rule that a cancellation is never scored as "regulatory activity"): regulatory_signal_state.py, test_regulatory_signal_state.py, mvp_regulatory_signal.md.

13 projects audited, real sources — not 20

We set out to audit 20 real Chinese industrial projects in Europe. 13 came back with a primary source we could verify live; the other 6 candidate names are known but not yet confirmed against an official source, so they're left out rather than filled in with weak data. Verdict counts as they actually came out: 3 CONVERTIBLE · 6 PARCIAL · 4 NO_CONVERTIBLE.

One structural finding along the way: Hungary has no free national LiDAR program (verified against the EU Joint Research Centre's JRC126223 report), which affects 4 of the highest-value projects here (CATL Debrecen, BYD Szeged, EVE Energy Debrecen, Sunwoda Nyíregyháza) — they'd depend on Sentinel alone, without a LiDAR baseline to anchor a volumetric delta.

CONVERTIBLE = localizable site + free baseline + Sentinel-observable, all three real. PARCIAL = missing one of the three (usually the LiDAR baseline, per the Hungary finding above). NO_CONVERTIBLE = either the project was cancelled before any construction, or it's a conversion of an already-existing plant with no new physical footprint to measure (Chery/EBRO Barcelona, Leapmotor/Stellantis Figueruelas).

Each of these 13 also has its own page with a downloadable machine-readable record: see all audited assets →

Request access

Not sold yet — this page measures real demand first. Leave your email and we'll reach out when there's something to try.

Would you pay for this?

No price set yet — this only records interest, nothing is charged.