Product
LE Control Plane
Stateful execution reuse for multi-step agent workflows—without hosting inference. LE is an API that governs continuation from execution state (spines) in your own vLLM / sglang / HF environment and produces receipts that prove avoided recompute.
Default: hash-only outputs, no prompt logging, customer-hosted inference.
What you get
Hard constraint profiles
EVAL / LIMITED / FULL policies enforced server-side. Benchmark safely, then promote to production without changing tooling.
Receipts and lineage
Every step emits an auditable receipt: lineage hashes, policy version, engine metadata, and reuse/accounting metrics.
Bring your own engine
Inference stays in your infrastructure. LE integrates with vLLM, sglang, or HF endpoints without becoming an inference vendor.
How it works
Replace full prompt replay with a spine reference plus a small delta. LE verifies and records the lineage.
Keys are scoped by org/project and bound to policy profiles and quotas.
Create a spine once from your stable context and schemas.
Each step sends only leaf deltas against spine_ref (no full prompt replay).
LE returns new spine_ref + receipt (hashes, lineage, metrics).
Policies
One API, multiple non-bypassable constraint profiles.
EVAL
Bounded depth and steps. Hash-only outputs by default. Deterministic receipts for side-by-side benchmarking.
- —Max depth + steps
- —Short TTL
- —Strict validation only
LIMITED_PROD
Early production constraints with conservative invalidation and stronger safeguards for sensitive workflows.
- —Bounded persistence
- —Conservative eligibility
- —Org quotas + audit logs
FULL
Uncapped workflows with enterprise governance and policy-governed LK retention.
- —Long-lived state
- —Cross-session reuse
- —Custom retention + compliance
Start with a bounded eval
Run EVAL in your environment. Collect receipts that quantify avoided recompute and validate fit before production.