A candidate indicator set for a proposed SDG 18 on culture

Derived from the European Capital of Culture corpus · engine run v1-full (1252c32e-9e63-4644-aef6-a8f6685f0a4b) · promoted 2026-09-11. A cause for debate.

1. Corpus

167 bid books (intent) · 41 outcome documents = 32 evaluation reports + 9 monitoring reports. Excluded (no machine-readable text): Đakovo 2020 — Pre-selection; Prešov 2026 — Pre-selection; Nicosia 2017 — Pre-selection; Nicosia 2017 — Final selection; L'Aquila 2019 — Pre-selection; Plzeň (Pilsen) 2015.

58,775 measures found across 208 documents — every one a quoted span with its page, 51,786 verified verbatim against the source; clustered into 2,647 concepts, of which 152 were explicitly decided by a human reviewer.

2. Headline

Matched pair — for 14 cities holding both their own bid book and their own outcome document: 429 of 1478 promised indicator concepts (29%) are reported against by the same city's evaluation or monitoring. Corpus-wide: 27.3% of concepts promised in any bid book appear in any outcome document.

3. Selection rule

Score = 0.4·delivery (outcome docs/41) + 0.25·prevalence (bid books/167) + 0.15·data presence + 0.2·spread (countries/34); eligible if ≥3 outcome docs, ≥3 countries, ≥1 bid books and not excluded by review; at least 2 per Lab theme, then by score, to 15. Derived from the corpus, offered for debate.

4. The candidate set (15)

5. Cross-check against the Lab's 26 priority indicators

strong = best match reported against by ≥15 of 41 outcome documents; present = matched below that bar; thin = only weak clusters match; absent = no matching concept. Divergence is a finding, not a failure.

6. How this was made — in plain English

What we asked

Should there be an 18th Sustainable Development Goal, for culture? Honestly, we don't know. But if there were one, it would need indicators — things you can actually measure. So we asked a simpler, answerable question: what do European Capitals of Culture already promise to measure, and what do their evaluations actually measure?

What we read

167 bid books with readable text — the documents cities write when they apply for the title, full of promises — and 41 evaluation and monitoring reports — the documents written afterwards, about what actually happened. All of them are public. 6 further documents were left out because they were scanned images or had lost their text; we name them.

What the AI did

It read every page and copied out, word for word, every sentence where a city named something it would measure or an evaluator reported a result — 58,775 of them. For each one it recorded which city, which year, which document, and whether an actual number was attached or just an intention. We checked its work: on a random sample of 500, it had copied the original wording correctly 90 times out of 100; the rest were paraphrases, which we flag and never use as quotes.

What it did not do

It did not judge, rank, or recommend. It did not decide what counts as a cultural indicator. It grouped similar sentences together — "visitor numbers" with "audience figures" — and then stopped. A human expert reviewed those groupings, merging, splitting and discarding, before anything was counted.

How the numbers work

For each indicator we count three things: how many of the 167 bid books mention it (a promise), how many of the 41 evaluations report on it (a result), and how many mentions come with real data. The indicators shown here are chosen by a written rule that favours indicators cities actually evaluate over ones they merely promise, and insists on a spread across themes and countries. Change the rule and the list changes — we show the rule so you can argue with it.

The one finding we'd stand behind

For the 14 cities where we hold both their bid book and their own evaluation, roughly one in three of the measures they promised (29%) was ever reported against. Every number on this page can be traced back to the document it came from. Click any of them and read the original.

What this is not

It is not a verdict on any city. It is not a proposal for what an 18th goal should say. It is a starting point, derived from the record — a cause for debate.

7. Caveats

• Derived from the ECoC corpus by machine extraction and clustering, then human review; offered for debate, not adopted.
• Intent is never outcome: bid-book figures are what cities promised; evaluation and monitoring figures are what was found afterwards.
• Extraction precision 89.8% verbatim on a 500-occurrence sample; only span-verified quotes are shown.
• Concept clustering is a machine pass corrected by reviewer decisions; merges change ranks, not the headline direction.
• Existing-SDG mappings are indicative keyword mappings, not an authoritative concordance.
• Country spread uses a normalisation map applied at query time; joint titles count for every member country.
• Occurrence-level 'has a figure' rates read high because extraction favours quantified statements; the honest data-presence figure is the concept-level share reported against.
• 6 of 173 bid books excluded — no machine-readable text: Đakovo 2020 — Pre-selection, Prešov 2026 — Pre-selection, Nicosia 2017 — Pre-selection, Nicosia 2017 — Final selection, L'Aquila 2019 — Pre-selection, Plzeň (Pilsen) 2015.
• Matched-pair figure pairs each city's own bid book with its own outcome document(s) for the same title (14 cities); small n — quote as a pre-review lower bound.
• Reviewer-confirmed mapping share at scoring time: 61.5%.