How a report is made. And where it can be wrong.

1 · Classify against the knowledge base

Every question is matched to one or two of 13 policy types (a tax, a subsidy, a ban, a cash transfer…). Each type carries a policymaker’s mind map: which of 10 branches matter most, 39 specific effects to examine, the decision questions to answer, stakeholders, risks, design levers, and the views to show. Browse it in the knowledge base.

2 · Read the evidence

Kanven searches its evidence library — news from the country being asked about (in English and in its own languages), government documents read from governments’ own websites (laws, budgets, economic surveys, project and audit reports, PDFs included, cited by page), World Bank and IMF data and policy papers, open research indexed by OpenAlex, and the original sources Wikipedia cites (never Wikipedia’s own text) — then searches the web for new sources and stores what it finds. Retrieval deliberately mixes Nepal’s own evidence, policy papers, and other countries’ experience. Crawling identifies itself and respects each site’s robots.txt; Kanven stores passages with links to the original.

3 · Propose links, compute numbers

AI models read the passages along the mind map and propose cause-and-effect links, each with a plausible range and the passages that support it. Standard economics — price to demand, freight to food to inflation to real income to poverty — comes from Kanven’s curated base model, which wins over the AI where they overlap. Known economic signs override the AI; magnitudes are capped by share where one quantity is part of another.

The AI never writes a number you see. The engine propagates the change through the model over three horizons (immediate, one year, three years) and re-runs it hundreds of times with each link varied across its range to give likely ranges.

4 · Verify

Every quote is checked against the stored passage it cites; paraphrases are removed. Outcomes each kind of policy must reach (a tax must reach revenue) are checked and filled from the evidence if missing. When the main model is rate-limited and a smaller one fills in, the report says so and caps its confidence.

5 · Compare countries

Evidence from abroad only helps if the country is similar where it matters. Kanven scores each country against the one being analysed using World Bank indicators for 217 economies — weighted for the kind of policy (fuel-import dependence for fuel pricing, remittances for migration policy, tax take for taxes) — plus whether it is landlocked, and shows the specific similarities and differences.

6 · Check the model against itself and the past

• Two model families. When capacity allows, two different AI model families build the cause-and-effect model independently. The report shows how far they agreed; poor agreement lowers confidence.
• Guardrails. One unmeasured, unsupported link can’t move a national total (revenue, GDP, inflation, poverty) beyond what a single policy plausibly causes. Large effects that are backed by evidence are kept but flagged for review. Every intervention is listed in the report.
• Measured confidence. The confidence level isn’t the AI’s opinion of itself: it is computed from how much evidence is from the country being analysed, how many links cite a passage, model agreement, how many decision questions the evidence answers, guardrail interventions — and, once enough historical cases are verified, Kanven’s track record for that kind of policy. Click the badge on any report to see the breakdown.

7 · A knowledge base that grows

When a question doesn’t fit any policy type Kanven knows, it designs a new one from the existing mind map, saves it and uses it. New variables are kept for reuse. New cause-and-effect links are learned by repetition: a link joins Kanven’s base model only after it has been proposed independently several times, in a consistent direction, with cited evidence — so one noisy answer can’t change future reports. Learned items are marked as learned and can be reviewed or removed by the Kanven team.

8 · Security of the evidence

Evidence comes from the open web, so it is treated as data, never as instructions: the AI is told to ignore any instructions inside documents. The crawler only fetches public addresses and respects each site’s robots.txt.

Limits you should know

• Baselines are approximate public figures; verify before citing a number formally.
• Much international evidence is from abstracts, not full reports.
• The household simulator uses a synthetic population loosely calibrated to Census 2021 — illustrative, not microdata.
• Precedents not found in the evidence are labelled “general knowledge — verify”.
• Similarity says whether evidence is likely to transfer, not that an outcome will repeat.

How to cite

Cite the primary sources a report lists. If you cite Kanven itself: “Kanven policy analysis, ‘[report title]’, [date], [share link]”.