Turn your experts' cases into tests

A scenario is a sequence of events and the outcome your expert expects. Product managers and domain experts can read it. Tarski runs it on the same engine as production and compares the result exactly.

Scenarios run directly on the engine

Classic BDD connects a scenario to the system through hand-written step definitions. That code can keep passing after the system has drifted from what the scenario says. A Tarski scenario has no step definitions: the production engine runs it directly.

Classic BDD
scenario (English)
↓ hand-written step definitions
↓ test harness
↓ mocks & adapters
↓ the actual system

Each arrow is code that can drift away from what the scenario says.

Tarski
scenario (input timeline + expected world)
↓ the production evaluator
= exact comparison

One deterministic run, with no adapter code between the scenario and the engine.

A scenario your domain expert can read

A drive-thru voice parser mishears an order as eighteen thousand waters. The scenario states the expected outcome: the order stays staged, the point-of-sale system receives nothing, and the customer is asked to confirm.

A product manager can read every line, and a risk officer can sign off on the expected outcome. Your coding agent changes the rules until the scenario passes.

fixtures/order-water-x18000 · timeline + expected world⊨ holds
scenario "absurd parse stays staged"

given the timeline
  obs mic.audio        "...eighteen waters..."
  obs candidate.order  water × 18_000   ← llm parse
  obs menu.bounds      water ≤ 24 per order

expect the world
  order.state        = staged
  pos.tickets        = none
  prompt.customer    = "confirm: 18,000 waters?"
  # the parse is recorded with provenance,
  # and the POS never receives it

They don't flake

Inputs are fixed and the engine is deterministic, so a scenario cannot pass on one run and fail on the next. A failure means the behavior changed.

They come from real cases

Any timeline in the log, including a mistake made in production, can be promoted to a scenario in one step.

What they cover

Scenarios check behavior on the cases you have recorded. Invariants are checked at every fixpoint, whatever the input. You review both before a release.

Turn your first expert case into a scenario.