Overview

Cheaper depth via reusable execution state

CLC provides a control-plane protocol that turns repeated prompt replay into stateful continuation. You keep inference in your own vLLM / sglang / HF stack. We provide policy enforcement and receipts.

The loop

Replace full prompt replay with a small delta against an authoritative spine reference.

Bootstrap

Create a session and bootstrap a spine once (your stable context and schemas).

Continue

Each step sends only a leaf delta + spine_ref; no full prompt replays.

Execute

Inference runs in your infrastructure (vLLM / sglang / HF).

Prove

Receive receipts: lineage hashes, policy version, and avoided-recompute metrics.

Constraint profiles

One API, multiple hard policies. Start in EVAL. Promote to production when ready.

EVAL

Bounded depth and steps. Hash-only outputs by default. Deterministic receipts for benchmarking.

  • Max depth (e.g. 3)
  • Max steps (e.g. 25)
  • Short TTL + no cross-session persistence

LIMITED

Early production constraints with conservative invalidation and stronger safeguards.

  • Bounded persistence
  • Strict validation
  • Org-level quotas and audit trails

FULL

Uncapped workflows with enterprise governance and long-lived state (LK) under policy control.

  • Long-lived spines
  • Cross-session reuse
  • Custom governance + retention

Training gets cheaper too

The same spine/leaf structure can be applied to vertical training corpora to reduce tokens and variance.

Vertical Training Compiler

Compile vertical corpora into stable spines (schemas, policies, formats) and minimal leaf deltas. Output standard datasets and metrics without sending raw data to CLC Labs.

Security model

  • Customer-hosted inference (no GPU hosting)
  • Hash-only by default; no prompt logging
  • Policy-governed state retention and deletion