Checked model calls
Catch local type, call, field, and direct classified-flow mistakes before effects start.
PRE-ALPHA POLICY-SAFE WORKFLOWS
Keep sensitive AI
workflows inside
policy.
Kora is an experimental language for checked, replayable document workflows with explicit data boundaries.
$ brew install ImAbhishekTomar/tap/kora
$ kora check examples/03_salary_review.ko
K 03_salary_review.ko (excerpt)
1type Employee:2 name: str3 role: str4 classified salary: int56type Assessment:7 band: str8 rationale: str910agent review(emp: Employee) -> str:11 budget: max_tokens = 400012 declassify emp.salary as pay for local_model:13 result: Assessment = analyze(14 {"role": emp.role, "pay": pay, "market": market_rate(emp.role)},15 "assess whether pay is below, at, or above market; band must be one of below/at/above"16 )1718 match result:19 case Ok(a):20 return f"{emp.name}: {a.band} - {a.rationale}"21 case Uncertain(why):22 return f"human review: {why}"23 case Exhausted(meter):24 return f"budget exhausted: {meter}"25 case Failed(why):26 return f"provider failed: {why}"
$ kora run --replay examples/03_salary_review.ko
Ada: below - The pay of 165 is less than the market rate of 210. Grace: above - The pay of 180 is higher than the market rate of 175.
Output replayed from the committed cassette
Replayable runs
Deterministic execution you can inspect, share, and replay.
Safe data flow
Explicit data boundaries and policies to protect what matters.
Build agent workflows that read like programs.
Use ordinary control flow with typed model calls, explicit budgets, and outcomes you can handle instead of hidden failures.
typeddurablecomposable
Trace every step. Opt into durable execution.
plan 812ms
tool.web 1.23s
provider.retry replayed
return complete
Classified data stays inside its boundary.
The checker catches direct unsafe model flow, and the runtime enforces the configured sink policy again when the call executes.
classified→declassify→approved sink