Invocations to actions describe an operational funnel

DX Terminal Pro's approximate 7.5 million invocations and 300,000 onchain actions imply an aggregate action-to-invocation ratio of about 4%. That ratio describes the path from model turns to chain activity. It does not mean the other 96% failed. Observe was an ordinary typed action in the historical loop, and the record includes policy checks before submission.

We derive the ratio by dividing 300,000 by 7,500,000. Both inputs are rounded scale summaries in the controls paper companion, so 4% should remain approximate. The ratio averages across the full deployment, rather than describing a particular activity setting, owner or day.

Give each transition its own denominator

A useful operational funnel distinguishes invoked turns, completed model outputs, parsed actions, policy-valid proposed submissions, chain submissions and settled actions. Observe decisions need their own branch because they can complete successfully without a transaction. Otherwise a chart with transactions at the end makes deliberate inactivity look like an execution defect.

The paper reports 99.9% settlement success for policy-valid submissions. That denominator sits downstream of the invocation count. Multiplying 7.5 million by 99.9% to estimate successful chain actions would apply the settlement rate to the wrong population. Equally, subtracting 4% from 99.9% would compare unlike measures rather than uncovering a reliability gap.

An illustrative funnel makes the distinction concrete. Suppose 1,000 turns produce 40 policy-valid submissions and the remainder resolve to observe or another recorded outcome. If all 40 settle, the submission settlement rate is 100% while the action-to-turn ratio is 4%. These values can coexist with perfect execution of the submitted population. This is a hypothetical arithmetic fixture, not a reconstruction of Terminal Pro.

We would save transition counts with reason codes and the exact eligibility definition at every stage. A timeout, a parse rejection and an intentional observe would remain separate. A submitted transaction awaiting confirmation would remain pending until reconciled. That receipt would let operators find the failing transition without treating the entire loop as one undifferentiated conversion rate.

Behavior also changes the funnel

The historical study reports activity-slider trade frequencies from 2.8% to 16.8% of invocations, a sixfold span. Those subgroup rates differ from the approximate 4% aggregate calculated here. Their coexistence is expected when populations have different settings and weights. An aggregate action ratio therefore cannot diagnose compliance with an individual owner's cadence.

The deployment covered 21 days in a twelve-token Base market under one frozen prompt and harness on one model family. The continuous record companion describes a later fleet with finalized turns, fills and a different tool manifest. Its fill-to-turn ratio would answer a related question under a different contract.

For a platform review, the useful deliverable is a transition ledger whose counts reconcile. It should show both deliberate no-trade outcomes and execution completion, allowing the reader to assess behavior and reliability separately.