Plexus targets typesafe_sdk ~> 0.4.0 and uses its public semantic/runtime surface rather than recreating it.
Prepared contracts
TypeSafeSDK.prepare!/1 and TypeSafeSDK.Prepared.fingerprint/1 back Plexus.Contract.Registry. Named/versioned contracts make the same semantic questions reusable across many actors and experiments.
Batch-first measurement
The framework path resolves a contract fingerprint, derives Contract.memo_key/2, checks replay/cache, and submits misses to a coalescer. Coalescers close on batch size or delay and call:
TypeSafeSDK.evaluate_many(client, states, prepared,
ordered: true,
on_error: :collect,
max_concurrency: max_concurrency,
cancellation: token
)A shared Pristine.Cancellation token belongs to each physical batch. Pruning removes the actor's waiters; if no waiters remain, the batch token is cancelled.
evaluate_stream/4 remains exposed through Plexus.Strategy.Fanout for coordinator-held sweeps that benefit from lazy early results.
OTP server
Plexus.Actor uses TypeSafeSDK.OTP.Server. The SDK callback order is handle_evaluation(result, tag, state). The README/examples use that order for direct evaluations.
Telemetry
Each run attaches to TypeSafeSDK evaluate/answer/batch-cancel telemetry and filters on explicit telemetry_metadata.plexus_run_id. Plexus does not log semantic state, questions, API keys, bodies or transport errors. Token measurements from successful TypeSafe calls update the run's observed token ledger.
Test fixtures
TypeSafeSDK.Test remains the deterministic semantic oracle. The Plexus test suite uses the real serialization/decoding path while replacing only the transport, making scheduler/coalescer tests meaningful without live API calls.
Live transport evidence
The standalone dataset runtime attaches a privacy-safe listener to TypeSafeSDK evaluation telemetry. It distinguishes Plexus logical measurement accounting from confirmed provider responses. A provider response is confirmed only when telemetry includes an HTTP status and non-empty request ID.
The summary exposes endpoint/model selection, returned model, status distribution, request IDs, retries, and token usage without storing semantic state, question text, request/response bodies, credentials, or authorization headers.