The product, on screen¶

One product, two seats. The person asking sees answers inside the AI they already use — they never see dst at all. The engineer sees the files those answers stand on. Both views below are the same live demo project: the definition the chat answer cites is the highlighted line in the editor underneath it. Change that line, and the answer changes.

Seat one — the person asking¶

No dashboard, no new tool to learn. Their AI calls dst behind the scenes, and the answer comes back with the SQL that produced it, a confidence grade, and the cost:

the AI your team uses — any of them

Where should we look for growth?

asked dst — how many customers are repeat customers?

asked dst — what was total revenue?

asked dst — revenue by month… (and more)

Your clearest lever is repeat purchases. Of your 100 customers, only 19 have bought more than once — 81 bought exactly once and never came back. Total revenue stands at $1,761, so converting even a fraction of those one-time buyers into second orders is the fastest growth path in this data.

And what is the average value of a customer?

asked dst — clarify came back

Before I answer — 'value' means two different things in your data. Which do you want?

lifetime value — total historical revenue per customer
order amount — a single order's total

The person sees conclusions, not queries. Behind that first answer their AI made several dst calls, and each one was checked against the lens's definitions and logged with its SQL, its confidence grade, and its cost — that record exists for the engineer and the auditor, not on this screen.

Two things to notice. "Bought more than once" is not the model's guess — it is your governed definition of a repeat customer, and you will see the exact file it lives in below. And the second answer is a question back: 'value' means two different things in this data, so dst asks rather than guesses.

▲ the answer · the file it stands on ▼

Seat two — the engineer¶

Everything the chat answer stood on is a file in a folder, in version control. The definition the answer cited — repeat_customer — is open below, its wording highlighted. The lens (who may ask, over what data) and the certified answers (question-and-SQL pairs a person approved) are the other two tabs. Underneath, the terminal: dst plan previews the change, dst apply publishes it, dst test re-checks every certified answer against it. All of it real: the files verbatim (lens.yaml abridged), the terminal from a live capture. The tabs and the highlighted tree files are clickable.

demo — semantic/definitions/examples/repeat-customer.md

EXPLORER

demo/

AGENTS.md

README.md

docker-compose.yml

dst.yaml

fixtures/

jaffle_shop.duckdb

profiles/

jaffle.json

jaffle.probe.json

semantic/

entities/examples/

customers.yaml

orders.yaml

definitions/examples/

lifetime-value.md

repeat-customer.md

value.md

lenses/customer_value/

lens.yaml

queries.yaml

certified_answers.yaml

definitions/

revenue.md

evals/

cases.yaml

compiled.yaml

1 2 3 4 5 6

--- metric: repeat_customer sql: customers.number_of_orders > 1 --- A repeat customer has more than one completed order (number_of_orders > 1).

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

name: customer_value display_name: Customer Value description: Customer lifetime value and order activity over the jaffle dataset. connections: - jaffle select: entities: - name: customers - name: orders definitions: - lifetime_value - repeat_customer - value model: temperature: 0.0 answer_mode: balanced instructions: Select explicit columns. Prefer month buckets for time series. access: allow: [] eval_gate: block auto_review: 'off'

1 2 3 4 5 6 7 8

- question: How many customers are repeat customers? sql: SELECT COUNT(customers.customer_id) AS repeat_customers FROM customers WHERE customers.number_of_orders > 1 source: "dst query req_7ba12c2be21b4d6c — verified against jaffle" verified_by: alex - question: What was total revenue? sql: SELECT SUM(orders.amount) AS revenue FROM orders WHERE orders.status != 'returned' source: "re-certified after ruling rev_81907144f8 — revenue excludes returned orders" verified_by: alex

Terminal

$ dst plan
  ~ semantic/definitions/examples/repeat-customer.md
  ~ customer_value — STALE compile (shared changed: definition/repeat_customer)
  ! customer_value: 1 certified answer(s) need re-verify (--full lists them)

Plan: 2 to change, 4 unchanged. (--full shows diffs and hints)
$ dst apply
connections: ok
semantic: updated definition/repeat_customer
lens customer_value: updated
Apply complete. 4 warning(s), 0 error(s). (--json for the row array)
$ dst test customer_value
PASS  me/customer_value: What was total revenue?                              9.7s
PASS  me/customer_value: How many customers are repeat customers?             37.2s
PASS  me/customer_value: What is the average value of a customer? [expect: clarify]
────────────────────────────────────────────────────────────────
3/3 passed (2 certified + 1 behavioral) in org me

A definition's wording changed (the highlighted line). plan names everything the change touches — the stale lens, the certified answer to re-verify — apply publishes it, and test re-runs every certified answer against live generation: 3/3 passed. The revenue entry's source line carries the correction loop's trace: it was re-certified after a review ruling changed what revenue means.