feat(sources): expand curated news coverage — 123 countries have no curated publisher
Severity: P1. The enrichment from #7615 works. It reached roughly half the corpus, and the half it missed is unchanged.
Coverage, measured across 25 sampled country pages
13 of 25 have a "Recent developments" section. 11 of 25 have a full country brief.
| Page | R5 words | R6 words | Δ | Enriched? |
|---|---|---|---|---|
| iran | 678 | 1,377 | +103% | Yes |
| bhutan | 657 | 1,057 | +61% | Yes |
| brazil | 690 | 824 | +19% | No |
| andorra | 746 | 858 | +15% | No |
| palau | 779 | 862 | +11% | No |
Enriched pages run 1,179–1,599 words (ukraine 1,599, nigeria 1,390, iran 1,377, india 1,352, sudan 1,330, germany 1,279, norway 1,245, nauru 1,235, canada 1,188, vietnam 1,179). Unenriched run 728–886.
Skipped in the sample: Brazil, Mexico, Kenya, Comoros, Tuvalu, San Marino, Liechtenstein, Monaco, Andorra, Palau.
The enrichment is genuinely analytical, which is why the gap matters
This is not padding. On Iran, 26% of sentences now carry a dated, sourced, causal country-specific claim, against 2% on Brazil, which received nothing. Round 5's finding was that country prose was 100% generated recitation with zero analytical sentences; the enriched half fixed that properly.
Template share across the sample is 33.2% mean (range 9.9%–57.4%), measured with country names and all digits masked, so that is an upper bound.
External evidence that it works: /countries/ took #1 for "current country instability rankings" this round, with the AI answer quoting World Monitor verbatim. The enrichment and the dated-snapshot pattern are why.
Do not trim depth in response to the Iran ranking loss. "Iran country risk score" went WIN → LOSS this round, but the cause was investigated and is vocabulary mismatch, not burial: the extractable answer appears four times including
<title>and meta description. The query wants a durable sovereign rating; we answer with an index that moves 4 points a day. Enrichment is safe — see the round-6 report.
Fix
- Run the enrichment across the remaining ~184 country pages. The pipeline exists and is proven; this is coverage, not new capability.
- Fix the rendering defects first — #7738, literal
**markdown and ISO codes leaking into headings, affects 7 of 11 sampled briefs. Do not multiply that across 184 more pages. - Watch grounding-source thinness. Bhutan and Nauru receive full 24/48/72h forecast briefs from a single grounding source. A confident multi-horizon forecast off one source is a trust liability on a YMYL page; either surface the source count or suppress the brief below a threshold.
- Guard: assert every indexed country page contains at least one dated, sourced, country-specific sentence, so the unenriched tail cannot silently persist.
Related
#7615 (the enrichment), #7738 (its rendering defects)
From the 2026-09-05 GEO audit, round 6.
Source: koala73/worldmonitor