Seven practical ways into Verisyntax.

1. Select customers

dataset danish_customers:
  from customers as customer
  select customer.id as id
  select customer.name as name
  where customer.country = "DK"

The typed dataset returns only rows matching the declared filter.

2. Join customers and orders

dataset paid_customer_orders:
  from customers as customer
  inner join orders as order_item on customer.id = order_item.customer_id
  select customer.name as customer_name
  select order_item.total as total
  where order_item.status = "paid"

The compiler rejects unknown aliases, incompatible column types, and cross-database joins.

3. Estimate a proportion

p ~ beta(alpha = 2, beta = 2)
successes ~ binomial(trials = 20, probability = p)
observe successes = 14
ask p
  interval 0.89 hdi

The result explains posterior mean, HDI, effective sample size, R-hat, acceptance, and seed.

4. Track article claims over time

Import raw TXT articles individually or in a batch. A local model may propose verbatim logic-bearing passages from the complete text; one editable claim is then structured from the selected evidence.

Passages are verified against the stored source snapshot. Claim Library shows version, status, evidence, and hash, and only a separate approval creates an active fact.

5. Handle aggregated health data carefully

The synthetic example joins 14 aggregate rows across two periods and three hospitals to organization metadata. It filters a declared diagnosis group while retaining contact type and final/provisional reporting status.

Grain, revisions, and denominators matter: raw contact counts are not patients, prevalence, risk, or quality.

6. Find indirect dependencies

depends_on(report, dataset)
depends_on(dataset, source)

indirect_dependency(X, Z)
  if depends_on(X, Y)
  and depends_on(Y, Z)

ask indirect_dependency(report, What)

The open-world result binds What = source and shows both supporting facts in the proof.

7. Explore associations before modeling

The local C++ runtime can scan up to 1,000 complete imported rows and report effect size, p-value, Benjamini-Hochberg-adjusted q-value, a concrete pattern, and data-quality warnings.

The result is exploratory, not causality, a fact, or row-level evidence. Selected candidates are added only to an editable local-AI request.