The charts that did not fit on the main page. Where the texts come from, how EU institutions and the press differ, which words travel with “dependency”, and the checks that left the story standing.
Every paragraph from EU institutions and four news outlets that mentions graphite, batteries, lithium, cobalt or critical raw materials, from Feb 2022 to early Oct 2026.
Kept after the keyword filter and duplicate removal.
Five from EU institutions, four from the press.
24 Feb 2022 to 9 Oct 2026. October 2026 is partial.
1,021 plenary paragraphs were spoken in another language.
Paragraphs kept per source, whole period
Pillar 2 corpus · notebook 06
| Source | Stream | Paragraphs |
|---|---|---|
| Commission press corner | EU institutions | 2,869 |
| EP plenary speeches | EU institutions | 1,519 |
| EP questions and answers | EU institutions | 814 |
| EP adopted texts | EU institutions | 428 |
| Commission documents | EU institutions | 205 |
| The Guardian | The press | 2,968 |
| Reuters | The press | 2,935 |
| Financial Times | The press | 1,798 |
| Euractiv | The press | 1,041 |
Paragraphs kept per month, by stream
EU dips are recesses: August each year and the May–Jun 2024 elections. Plenary speeches end in Apr 2026, when the source corpus stops.
Pillar 2 corpus · notebook 06
Each document counts once: a speech, an article, a question or a press release. A document “makes a claim” if any of its paragraphs says Europe depends on someone.
By stream and period, adjusted for coding errors. Lines show the 95% range.
From Dec 2023–Sep 2025 to after Oct 2025, press claims rise by 19.1 points (range +10.1 to +24.1). EU claims rise by 4.9 (+1.9 to +9.1).
5,489 documents · Claude-coded paragraphs, adjusted with the 150-paragraph check set · 1,000 simulations
| Stream · period | Claims | As coded | Of claims: name a stage | As coded |
|---|---|---|---|---|
| EU institutions · Before Dec 2023 | 39.4% (27.6–47.1) | 49.1% | 29.1% (8.3–46.0) | 35.5% |
| EU institutions · Dec 2023 – Sep 2025 | 35.8% (22.8–44.4) | 46.6% | 22.7% (0.8–39.6) | 28.7% |
| EU institutions · From Oct 2025 | 40.6% (29.6–47.7) | 49.3% | 18.0% (0.5–31.6) | 24.7% |
| The press · Before Dec 2023 | 20.9% (10.7–28.0) | 29.5% | 79.6% (45.4–100.0) | 67.5% |
| The press · Dec 2023 – Sep 2025 | 11.2% (0.0–19.9) | 22.2% | 72.0% (0.0–100.0) | 53.4% |
| The press · From Oct 2025 | 30.3% (16.4–38.5) | 40.4% | 37.2% (6.0–67.5) | 39.1% |
Of the documents that claim a dependency, the share that name a stage of the chain
The ranges are very wide, so read the direction, not the levels. The press fell from about 80% to 37%; EU institutions from 29% to 18%.
Same data as above
Of the documents that claim a dependency, the share that name each stage (whole period)
A document can name several stages, so the bars do not add up to 100%.
Same data as above
| Stage | EU institutions | As coded | The press | As coded |
|---|---|---|---|---|
| Mining | 10.2% (1.7–17.5) | 13.0% | 20.3% (6.3–30.3) | 18.1% |
| Refining | 11.0% (1.9–18.6) | 13.8% | 37.8% (12.1–56.3) | 34.6% |
| Components | 5.3% (0.7–9.2) | 6.9% | 13.5% (4.3–20.4) | 11.8% |
| Packs & vehicles | 6.1% (0.7–10.5) | 7.9% | 15.0% (4.4–22.9) | 13.5% |
This check uses no coding at all. We took every paragraph that contains a dependency word (depend, reliance, vulnerable, chokepoint and so on) and counted which other words appear with it.
Share of paragraphs with a dependency word that also contain each group of words
Bold change labels mark a clear change (the 95% range excludes zero). One caution: geopolitics words rise in all paragraphs, not only near dependency words. Compared with the whole corpus, the tie between dependency and geopolitics does not get stronger. The shift is in the press; EU institutions barely move.
14,577 paragraphs · 2,299 with a dependency word · ranges from resampling whole documents
| Stream · word group | Before Dec 2023 | Dec 2023 – Sep 2025 | From Oct 2025 | Change, last two periods |
|---|---|---|---|---|
| All texts · Mining words | 14.8% | 15.5% | 14.4% | −1.0 (−4.9 to +2.9) |
| All texts · Refining words | 10.2% | 9.6% | 12.5% | +2.8 (−0.9 to +6.6) |
| All texts · Component words | 4.9% | 7.7% | 2.3% | −5.4 (−7.8 to −3.0) |
| All texts · Pack & vehicle words | 19.0% | 16.5% | 12.0% | −4.6 (−8.5 to −0.6) |
| All texts · Geopolitics & deals | 36.2% | 40.9% | 47.9% | +7.0 (+0.7 to +12.9) |
| All texts · Country names | 50.1% | 43.5% | 48.2% | +4.7 (−1.8 to +11.3) |
| EU institutions · Mining words | 11.9% | 14.0% | 13.4% | −0.6 (−5.1 to +4.2) |
| EU institutions · Refining words | 9.4% | 9.3% | 10.8% | +1.5 (−3.2 to +6.4) |
| EU institutions · Component words | 4.3% | 6.3% | 2.6% | −3.7 (−6.6 to −0.8) |
| EU institutions · Pack & vehicle words | 13.7% | 15.1% | 7.9% | −7.2 (−11.8 to −2.8) |
| EU institutions · Geopolitics & deals | 46.0% | 49.3% | 51.2% | +1.9 (−6.4 to +10.0) |
| EU institutions · Country names | 39.3% | 31.9% | 33.3% | +1.5 (−6.2 to +8.9) |
| The press · Mining words | 18.8% | 18.4% | 16.2% | −2.2 (−9.1 to +5.5) |
| The press · Refining words | 11.4% | 10.3% | 15.3% | +5.0 (−1.8 to +11.5) |
| The press · Component words | 5.8% | 10.3% | 1.7% | −8.6 (−13.5 to −4.1) |
| The press · Pack & vehicle words | 26.5% | 19.3% | 18.8% | −0.5 (−7.8 to +6.8) |
| The press · Geopolitics & deals | 22.5% | 24.7% | 42.4% | +17.7 (+8.2 to +26.6) |
| The press · Country names | 65.2% | 65.9% | 72.9% | +7.0 (−2.5 to +15.9) |
Each dot is a word; a line joins words that appear together far more often than chance. Hover a word to see its neighbours.
An illustration, not a result. How many lines survive depends on a filter setting, and at a looser setting about half the central words change. Graphite sits with anode, cell and cathode; dependency words sit with the US, export controls and the Inflation Reduction Act.
325 linked words, 520 links · word groups and cluster names are ours
How strongly two words stick together, Dec 2023 – Sep 2025 vs from Oct 2025 (0 = no more than chance)
All pairs shown changed more than chance would allow, after correcting for testing many pairs. Rare earths moved closer to China; the US moved away from Ukraine and from tariffs.
Paragraph co-occurrence · 1,000 shuffles of documents between periods
| Pair | Before | After | Paragraphs |
|---|---|---|---|
| rare earths + China | 0.23 | 0.35 | 105 → 179 |
| critical minerals + ally | 0.04 | 0.24 | 5 → 29 |
| technology + state aid | 0.05 | 0.40 | 8 → 26 |
| manufacturing + state aid | 0.21 | 0.51 | 6 → 16 |
| aid + scheme | 0.52 | 0.79 | 13 → 30 |
| lithium + hydroxide | 0.32 | 0.43 | 30 → 22 |
| battery + storage | 0.22 | 0.31 | 135 → 93 |
| electric vehicle + sales | 0.32 | 0.44 | 82 → 43 |
| US + tariff | 0.39 | 0.25 | 77 → 36 |
| deal + green | 0.58 | 0.32 | 46 → 6 |
| wind + solar panel | 0.64 | 0.38 | 72 → 10 |
| US + Ukraine | 0.30 | 0.07 | 64 → 9 |
| Japan + Korea | 0.71 | 0.41 | 14 → 6 |
| deal + plan | 0.22 | -0.03 | 24 → 7 |
We scored three ways of coding the text against paragraphs we coded by hand, re-ran the main result in different ways, and kept the results that came out flat.
Average agreement score (F1, 0 to 1) on 150 hand-coded paragraphs. The word list scored 0.53.
Agreement between our two coders on 40 paragraphs (0.61 on “is there a claim, and does it name a stage”).
Share of check paragraphs where Claude Opus gave exactly the same codes as Claude Sonnet.
Share of kept paragraphs that are on topic. It finds 76–85% of the relevant ones.
Change in points, Dec 2023 – Sep 2025 vs from Oct 2025. Lines show the 95% range; no line means no range was computed.
The rise in vague claims shows up in every cut. Stage-naming claims stay flat overall. The word check, which uses no coding, points the same way.
Paragraph and document adjustments · co-occurrence test
| Measure | Change (points) | 95% range |
|---|---|---|
| Claims that name no stage · Paragraphs, adjusted for coding errors | +10.0 | +7.0 to +12.0 |
| Claims that name no stage · Documents, adjusted | +11.4 | +6.6 to +14.7 |
| Claims that name no stage · Documents, as coded | +9.8 | — |
| Claims that name no stage · EU institutions only (documents) | +5.6 | +2.8 to +9.4 |
| Claims that name no stage · The press only (documents) | +15.6 | +6.3 to +21.0 |
| Claims that name a stage · Paragraphs, adjusted for coding errors | +0.7 | −1.0 to +2.2 |
| Claims that name a stage · Documents, adjusted | +1.6 | −0.1 to +3.3 |
| Claims that name a stage · Documents, as coded | +1.6 | — |
| Claims that name a stage · EU institutions only (documents) | −0.8 | −2.9 to +1.3 |
| Claims that name a stage · The press only (documents) | +3.5 | +0.8 to +5.8 |
| Words next to “dependency” (no coding used) · Component words | −5.4 | −7.8 to −3.0 |
| Words next to “dependency” (no coding used) · Pack & vehicle words | −4.6 | −8.5 to −0.6 |
| Words next to “dependency” (no coding used) · Geopolitics & deals words | +7.0 | +0.7 to +12.9 |
Agreement with our hand coding (F1, 0 = none, 1 = perfect), by code
The check set holds only 8 to 16 paragraphs per stage, so one paragraph moves a stage’s score by several points. Claude’s errors are mostly over-claiming, which the adjustment corrects.
150 hand-coded paragraphs (75 EU institutions, 75 press) · codebook v2
| Code | Hand-coded positives | Claude | Word list | Similarity |
|---|---|---|---|---|
| Mining | 12 | 0.62 | 0.52 | 0.30 |
| Refining | 12 | 0.75 | 0.60 | 0.32 |
| Components | 8 | 0.61 | 0.42 | 0.18 |
| Packs & vehicles | 16 | 0.65 | 0.51 | 0.30 |
| No stage named | 34 | 0.74 | 0.49 | 0.16 |
| No claim | 81 | 0.88 | 0.64 | 0.72 |
| All six (average) | 150 paragraphs | 0.71 | 0.53 | 0.33 |
Claims rose briefly, then fell back from Jan 2024. Documents claiming a dependency dropped 7.9 points between the first two periods.
The shift-finding tool found no confirmed lasting shift in the EU stream. The Oct 2025 rise is a press rise.
No lasting shift for mining, refining or components. The only one is press talk on packs & vehicles in 2023, during the UK–EU dispute over battery rules of origin.
Documents naming a stage: +1.6 points (range −0.1 to +3.3). More talk, not sharper talk.
Grouping neighbouring months by their mix of stages gave no clear periods. The talk drifts more than it jumps.
Only 71 graphite paragraphs after Oct 2025. After Dec 2023, graphite claims became less stage-specific (−9.6 points, range −17.4 to −1.6).