Improve policy with evidence.

Start with a reproducible baseline

Record the task, sources, policy, model route, usage and acceptance criteria. Preserve an equivalent baseline so a shorter context or cheaper route can be compared with what it replaces.

The public Engine provides deterministic processing, controls and evidence. Optional optimization intelligence works through that same contract; ordinary Engine use does not require a private model or service.

Verify the optional runtime

Load optional intelligence only when its artifact signature, compatibility and entitlement checks pass. Its suggestions stay behind the same source, identity, policy and budget controls as a deterministic decision. An unavailable, incompatible or rejected artifact falls back to the public deterministic path; it cannot expand authority or prevent ordinary Engine use.

Propose, evaluate, authorize

Stage	Required evidence
Propose	The candidate change, affected scope and intended benefit
Evaluate	Results against the baseline, including quality, cost, latency and policy constraints
Authorize	A decision from the host or organizational policy
Apply and observe	A traceable policy version, measured outcome and rollback path

The optimizer proposes. The host authorizes. A candidate cannot grant itself additional source access, choose a forbidden provider or skip an approval.

Measure an accepted outcome

Token reduction is one input. Include model output, retries, recovery work, cache effects and the quality check. Keep observed costs, estimates and unknowns separate.

Reject or roll back a change that crosses a security boundary or harms the required outcome, even when it uses fewer tokens. See measurement and evaluation.

Engine contract & referencesReviewed 19 Sept 2026

Engine concepts describe responsibilities and behavior. Exact commands, package versions and platform contracts belong to their versioned references.

Versions & compatibility