Typed layer only. No HTTP.

• Cargo workspace, edition 2024, MSRV 1.85: s1, s1-derive, s1-test
• #[derive(Choice)], #[derive(Score)], #[derive(Questions)]
• Wire JSON matches TypeSafe Choice / Score / Noul
• ChoiceAnswer / ScoreAnswer / NoulAnswer with value / prob / distribution / confidence / expected / nearest / p()
• Unknown Choice label → DecodeError::UnknownLabel (no panic)
• Policy / Verdict::{Act, Review, Escalate} and NoulPolicy
• S1<C: DecisionBackend>::ask
• s1-test::FakeClient
• Golden tests + trybuild UI tests
• Examples: triage, moderation (FakeClient, no network)
• Dual license MIT OR Apache-2.0
• backend-typesafe-rs — impl DecisionBackend for typesafe_rs::Client (crates.io typesafe-rs 0.1). Default cargo test --workspace stays FakeClient-only.

1. ask_many / pipeline — bounded concurrency over many states (PRD G6). Prefer delegating rate limits to typesafe-rs tower layers.
2. s1-cli — s1 ask / s1 map / s1 gate / s1 replay (PRD G8, SPEC §10).
3. backend-typesafe-ai — optional adapter for Joey's typesafe-ai crate.
4. Cassettes — Cassette::record / replay against a real backend, sanitized.
5. Observability — tracing spans on s1.ask, optional metrics histograms (PRD G7).
6. #[s1(flatten)] and richer structured instructions_json / desc_json coverage.

backend-typesafe-ai and live Jev cassettes are explicitly out of this MVP.