Industrial Policy Monitoring

Core Terms: Article → Measure

• Article: a single press item from our curated outlets — the raw input. An article is mapped once it is linked to the policy coverage it reports on; article counts indicate how much press attention something received.
• Measure: the unit of analysis everywhere in this dashboard — one concrete government action with a specific instrument and actor (e.g. a purchase premium, a tariff, an R&D grant scheme), extracted from the articles of a policy package. One package often contains several measures; every count, chart, filter, and comparison here is measure-based.
• Funding: the monetary amount the press reports for a measure, shown exactly as stated (value and currency). Most measures have no stated amount — shown as "Not Available" — and reported figures are press claims, not verified budget lines. Amounts in different currencies are never summed together.

Which Press We Read

We don't crawl the open web — coverage comes from a curated set of quality outlets per country, so the dashboard reflects mainstream reporting and public debate rather than every fringe mention:

• All countries: articles come from the LexisNexis press archive — a licensed full-text news archive — restricted to a hand-picked whitelist of leading outlets per country: ~27 German titles (Frankfurter Allgemeine, Der Spiegel, Tagesspiegel, taz, Stern, quality regionals, and the dpa financial wire), ~25 French titles (Le Monde, Le Figaro, Libération, La Tribune, the business magazines, major regional dailies, France 24), and national dailies, business press and newswires for the United Kingdom, United States, Italy, South Korea, the Netherlands, Estonia, Poland, Czechia and Spain. Because the archive is licensed, these articles are cited by source and date rather than linked.
• Not every press market is equally served: whitelists differ in depth — some countries are covered by national dailies and trade press, others almost entirely by newswires, because the archive simply does not carry every country's nationals. That is why cross-country views default to the curated comparable set (step 07) and offer a switch to the full corpus. Read a country's own page on the full corpus; compare countries on the curated set. Even then, absolute counts still partly reflect how deep each country's core-press file goes, so treat them as coverage signals rather than as a census of policy.
• Language: retrieval is always in the country's own language — German, French, Italian, Korean, Dutch, Estonian, Polish, Czech, Spanish and English — then normalized into English, so the dashboard captures domestic framing, not just what reaches the English-language press.
• Topic scope: within those outlets we collect everything that mentions the sector's products or industry. Policy relevance is decided by the AI reading step (Step 02) rather than by policy search terms — the exception being a few policy phrases still grandfathered into the legacy German and French queries — which keeps the system able to discover policy instruments nobody thought to search for.
• Geographic scope: national, regional and (for EU members) EU-level actions are kept; other governments' actions are excluded unless they directly involve the tracked country.

Because the corpus is limited to these sources, developments they don't cover may be missed — the dashboard is a structured signal of public debate, not a complete legal registry. A quarterly audit quantifies this gap (see Step 06).

Dual-Tier LLM Processing Strategy

Throughput and extraction quality are balanced by varying the reasoning budget rather than the model: both tiers run the same open-weights model (DeepSeek V4 Flash), so a measure is never judged by a weaker reader than the one that screened its article.

• FAST tier: high-volume relevance screening and binary yes/no checks, run with the chain-of-thought suppressed so the model returns clean JSON without a reasoning detour.
• SMART tier: the same model at maximum reasoning effort, for per-article measure extraction and within-window deduplication — the judgements where reading one article closely is the whole task.
• Big-batch judgements: consolidation and initiative linking weigh dozens of measures at once and run at medium effort, because at maximum the model reasons straight through its completion budget and returns nothing. Blocks are kept small and every reply is bounded, so a stalled batch costs one retry rather than a lost merge.
• Serving: inference is pinned to 8-bit (fp8) providers only, so quality does not drift with whichever host happens to be cheapest; the consolidation and linking stages prefer the model's first-party endpoint.

Clustering & Deduplication Mathematics

To avoid duplicate packages and measures without manual review, the pipeline relies on local semantic vector alignment:

• Sentence Embeddings: Employs paraphrase-multilingual-MiniLM-L12-v2 to convert policy and measure text into 384-dimensional dense vectors.
• FAISS Vector Database: Computes cosine similarity between incoming article vectors and policy centroids (cosine threshold: 0.85).
• Cross-Window Mapping: Snapshots across different months are matched via cosine similarity (threshold: 0.75), with cross-group measure deduplication (threshold: 0.85) applied before frontend export.
• Consolidation: embeddings only propose merges. Candidates are retrieved at a deliberately low floor (cosine 0.55, plus shared-initiative and lexical signals) and every group is then adjudicated by the model, with any assembled group of four or more members re-checked for over-merging. Recall comes from the vectors; the merge decision is the model's.

Intervention-Side Lens (Demand / Supply / Infrastructure / Trade / Regulatory)

The Analysis & Compare view groups every measure into an intervention side. This is a transparent, rules-based mapping from the existing instrument taxonomy — not a separately labelled data field — so it stays fully auditable and can be refined centrally. A future LLM pass may assign an explicit value-chain stage to supersede it.

• Demand: consumer subsidies, public procurement, and local-content requirements — instruments that move who buys or adopts.
• Supply & Production: fiscal subsidies, tax incentives, credit & finance, equity, industrial funds, R&D, labour & skills, land, and business-environment measures — instruments that lower the cost or raise the capacity of producing.
• Infrastructure: infrastructure investment and industrial clusters — shared physical and spatial enablers.
• Trade & Investment: import measures, export controls and promotion, and investment screening.
• Regulatory & Framework: environmental regulation, market-access standards, and other framework-setting measures.

Goal, Status & Authority

Goal categories capture the public objective a measure pursues:

• Green Transformation · Industrial Competitiveness · Resilience & Security · Defense & Space · Employment & Labor · Regional Development · Digital Transformation · Physical Infrastructure · Trade Agreements · Other Public Goal

Measure status tracks each measure's lifecycle stage:

• In Discussion: debated, proposed, or negotiated without a formal commitment yet.
• Announced: formally committed or decided, but not yet implemented.
• Active: implementation has begun — funds are flowing, construction is underway, or rules are in force.
• Cancelled: previously announced or active, then withdrawn, stopped, or abandoned.
• Expired: previously active, now ended or lapsed.

Originating authority records who decided the measure, following the GTA originator/implementer convention:

• Supranational (EU) · National · Regional/State. Named ministries and agencies are counted under National in all statistics and filters; where the press names the specific body, the measure sheet shows that name directly.

How to Read the Numbers

Every figure on this dashboard is derived from press coverage by an automated pipeline. That makes the numbers a measure of what the press reported, not an official register. The rules below say what each number is and, just as importantly, what it is not.

• A count is a count of canonical measures, not of articles or packages. Every article is processed separately, so one government action typically appears dozens of times in different wordings; the consolidation stage merges those into one measure. When consolidation improves, the headline count falls while the information content rises — a lower number after a recompile is not lost data.
• Counts measure press attention, not policy volume. Corpus size, language, outlet mix and archive depth differ by country: the United States is read through far more outlets than Estonia. Do not rank countries by raw totals. Compare shares (share of measures by instrument or goal), trends over time, and the curated set — never the bare number of rows.
• Curated vs. extended set. A measure is in the curated comparable set when at least one of its sources is a national daily, national business title or newswire (tiers T1–T3). Trade and regional press (T4–T5) vary enormously by country and inflate uncurated counts unevenly. Cross-country views filter to the curated set by default and show a "Curated set: N of M" badge; single-country views show everything. The switch on each badge lets you compare the two.
• Scanned, mapped, links. Scanned is every article retrieved from the archive. Mapped is the distinct articles that back at least one measure (counted once even when the same wire piece sits in two sector corpora). A measure's sources are links: one article routinely supports several measures, so links exceed mapped articles. Per-country totals therefore sum to slightly more than the corpus-wide distinct figure.
• The country × sector heatmap shades each row against its own maximum. A dark cell means "this sector dominates this country's coverage", not "this cell is large in absolute terms". Read across a row, not down a column; the tooltip gives the cell's share of that country's measures.
• Sector totals follow strict boundary rules. The six sectors form a partition: a measure lands in exactly one. Battery cell plants are EV, not minerals; ammonia and fertiliser plants are chemicals, not hydrogen; direct-reduction steelmaking is steel, not hydrogen; only upstream mining, refining and stockpiling is minerals. A "minerals" count is therefore not the whole battery supply chain, and a "hydrogen" count excludes hydrogen use owned by other sectors.
• Status is the last reported stage, not verified implementation. "Active" means the press last reported funds flowing, rules in force or construction under way; it has not been checked against a gazette. The status history in each measure sheet shows how reporting moved over time; a status that never advances past "Announced" may be stalled — or simply no longer newsworthy.
• Amounts are press claims, never budget lines, and are never summed across currencies. Most measures state no amount at all. Where a measure is reported at different scales (£2.5bn in one paper, £2,500m in another) the sheet shows a range on one currency-and-period basis; a per-year figure is labelled as such and must not be read as a one-off total. Non-monetary figures (vehicles, GW, jobs) keep their own unit.
• Dates are reporting dates at mixed precision. "First seen" is the first press observation, not the enactment date, and the press often dates a measure only to a month or a year. Monthly views use article publication month, so the current month is always partial — a dip in the latest month is an artefact of the calendar, not a policy slowdown. The monthly coverage card therefore opens on the last complete month.
• An empty cell is usually real. A country × sector with zero measures means the corpus was read and no industrial-policy measure was found in it, most often because the corpus itself is thin (Czechia's hydrogen coverage, for example). It is not a pipeline failure: a failed stage leaves no registry at all and is reported, never silently shown as zero.
• "Consolidated from N observations" and source counts are attention proxies. They say how often and how widely a measure was reported, which correlates with political salience — not with the size or importance of the measure itself. A small regional grant covered by two papers and a €3bn national premium covered by ninety are both one measure each.
• Initiatives group measures across countries and languages. Roughly four in ten measures belong to a named programme such as an EU act or a national strategy; the rest are standalone actions, which is expected, not missing data. Initiative names are canonical English forms chosen by the pipeline and may differ from any single outlet's wording.
• Everything is traceable, and you should trace it. Each measure sheet carries the extracted evidence quote, the outlet and date it came from, and the full list of mapped articles with a web-search shortcut. All content fields are machine-extracted and carry that uncertainty; before quoting a figure or a status, open the sheet and check the underlying reporting.