Safe measurement plan · Control and decision

Advance only when the gate opens

Use freeze, cost, maintenance, canary and long stages as independent barriers.

Primary local sources: plan, model matrix, per-cell plan, manifest and sanitized inventory.

The idea in plain language

The recommended next step is not the heavy benchmark. Keep the backlog blocked until an exclusive window, then request authorization only for reversible maintenance and the 14-route canary. Gemini 3.6 stays out until its own preflight.

Think of this measurement as a sealed scale: if you swap the scale or move the object midway, the number loses meaning. The analogy breaks because gateways can also transform the request, not merely weigh it.

Execution stops on drift, ambiguous surface, unsafe retry/failover, identity mismatch, 401, context rejection, two 429/5xx, timeout, budget or human traffic. Restore and a health proof close the window.

Verifiable excerpt

maintenance_approved = false\ncanary_approved = false\nheavy_benchmark_approved = false\nnext = await explicit approval

outputs/BENCHMARK-MEASUREMENT-PLAN.md · outputs/benchmark-cell-plan.csv · outputs/planned-run-manifest.json

The path in one picture

Warp → sidecar → CLIProxyAPI → provider → proof gate

Retrieval check

What is the active decision now?