Methodology
A seven-wave multi-agent pipeline turns Russian, Uzbek, and Kyrgyz primary sources into a typed knowledge graph. Eleven specialised agents work in three parallel research waves, one synthesis wave, one adversarial audit, and one outreach wave. Cross-model verification on a strict twenty-dollar OpenRouter budget keeps every Tier-A claim independently re-checked.
What does each wave do?
Each wave produces a typed JSON artefact under state/. Every record carries a verification tag and a sources array. The merged knowledge graph at state/knowledge_graph.json is the single read view that downstream waves consume.
| Wave | Agent(s) | Purpose |
|---|---|---|
| 0 | blueprint-architect | Strategic plan, target lists, and the constraint inventory that scaffolds every later wave. |
| 1 | legal-cartographer · case-tournament | Catalogue the decree corpus on lex.uz and cbd.minjust.gov.kg; tournament 100 global precedents for transferability to UZ and KG. |
| 2 | institution-mapper · donor-pipeline · procurement-harvester · trend-triangulator | Map state bodies (8 tiers), donor programmes with named TTL, live tenders, sector trends. |
| 3 | people-intelligence | 100+ decision-makers, including the 16 diaspora bridges who shape policy from London, Dubai, and Silicon Valley. |
| 4 | initiative-synthesizer · solopreneur-mvp-synthesizer | 100 institutional B2G initiatives plus 200 solopreneur MVPs, scored on five weighted axes. |
| 5 | reflexion-auditor | Adversarial re-verification with a different model, minimum three HIGH-severity issues, public corrections. |
| 6 | pitch-artificer | Tier-A and Tier-B outreach bundles. Private outreach scripts stay in the contributor vault. |
How is verification enforced?
Every record carries one of six verification tags: VERIFIED (primary source plus one corroboration), L2_VERIFIED (primary source only), L3_VERIFIED (primary plus expert commentary), INFERRED (indirect evidence), UNVERIFIED, or CONTRADICTED. The reflexion-auditor calls a different model than the original agent — Sonar Pro re-checks Sonar Deep Research output, breaking the echo chamber. Twelve content quality gates block any deploy on a single H1 violation, internal ID leak, fabricated decree, or fabricated LinkedIn URL.
What is the source priority?
Uzbekistan: lex.uz, gov.uz, president.uz, norma.uz, then spot.uz, gazeta.uz, kun.uz, daryo.uz. Kyrgyzstan: cbd.minjust.gov.kg, president.kg, gov.kg, kabmin.kg, then 24.kg, kaktus.media, akipress.org. Donors: documents.worldbank.org, projects.worldbank.org, adb.org/projects, ec.europa.eu/international-partnerships. Every country claim cites at least one Russian-language source. Fifty-three percent of all sources in this run are Russian-language.
Reproduce this research yourself
The full pipeline is open-source. Clone github.com/avaluev/ca-b2g-research, copy .env.example to .env and add your OpenRouter key, then run make run. Total runtime: about ten hours wall-clock. Total OpenRouter spend (with the audit wave on paid Sonar Pro): under twenty dollars. Anthropic Claude usage runs on your own subscription.