Decision: are metered classifier/decision-model calls in scope for ADF lanes? #967

Decision needed

ADF uses subscription lanes only. Typed decision models (TypeSafe Jev, Levanto Sage) raise a policy question the rule does not answer: are metered classifier or decision-model calls in scope for ADF at all?

Facts

• Jev: metered at 0.042 USD per million input tokens; output free; no spend cap.
• Levanto Sage: flat monthly plans with a hard stop (HTTP 402) when the allowance runs out. That is a spend cap by construction, which helps against runaway loops. Per call it costs 2.5 to 20 times more than Jev.
• Both send agent working state to a US-hosted third party. Neither returns a reasoning trace.
• Eval tooling such as the claude-api hillclimb runner also calls a metered API key.

Options

1. Subscription-only covers classifiers and eval runners too (status quo, strictest).
2. Allow capped plans only (hard stop at the allowance), for named, non-gating uses.
3. Allow metered calls under a per-agent spend ceiling enforced by the proxy.

Acceptance

• A recorded decision, plus an update to the lane policy documentation.

Mirror of Gitea terraphim/terraphim-ai#3420 (source of truth).