Judgment Layer / Execution Layer, Intervention Threshold, and Escalation Rate

The boundary between what agents execute and what the Steward decides is not a management convention — it is an architectural parameter set in code before the system launches. The Judgment Layer / Execution Layer binary names the two sides of that boundary: the Execution Layer is every task that follows deterministic logic and can be owned by agents; the Judgment Layer is every decision that requires genuine human assessment and is owned by the Steward. The Intervention Threshold is the coded parameter that places each task on one side or the other — defined at system design time, not discovered through operational failure. The Escalation Rate is the measurement that confirms whether the boundary was set correctly: the proportion of agentic executions that required Steward escalation in practice. At T1 tasks, the target is 1:100 — 99 percent autonomous execution, one percent genuine judgment calls. A rate above that target means the Steward is operating in the Execution Layer. A rate below it means the boundary is too permissive. Either way, the correct response is architectural, not operational.

When the boundary between the Judgment Layer and the Execution Layer is not made explicit before the system launches, it is discovered through operation — which means discovered through failures. Tasks that should have been handled by the Execution Layer escalate to the Steward because no Intervention Threshold was set to handle the exception class. The Steward resolves the case. The resolution is not encoded. The same exception class escalates again — indefinitely — until the architecture is corrected. The Coordination Tax the autonomous business was designed to eliminate reappears inside the agentic system: the Steward is coordinating the system’s exceptions rather than governing its logic. Administrative Density rises in the Judgment Layer — the proportion of the Steward’s time consumed by tasks the Execution Layer should have handled deterministically. The Revenue Loop that was supposed to operate without proportional human cost is operating with it — because the boundary was never drawn in code at design time.

The internal signature of a mis-calibrated Intervention Threshold is a Steward who is always occupied — not with architectural improvement or exception class design, but with individual task escalations that should have resolved autonomously. The Escalation Rate is above its tier target: at T1 tasks, well above the 1:100 baseline that defines autonomous execution. The Steward is intervening at 1:20 or 1:10 — meaning five to ten percent of T1 transactions require human assessment. At that rate, the Steward is operating in the Execution Layer under the label of the Judgment Layer. The Operational Arbitrage the business was designed to capture at T1 — near-zero marginal cost per additional transaction — is not being captured. The Steward’s time is the marginal cost of each transaction the Escalation Rate diverts from autonomous resolution. MTTI falls as a consequence: the system requires human intervention more frequently not because decisions are genuinely complex but because the threshold that defines complexity was set too conservatively.

The conventional response to a Steward who is too busy is to add operators — to distribute the Judgment Layer workload across a larger team. This addresses the symptom without touching the cause. The Intervention Threshold is still calibrated for the wrong volume of exceptions. More Stewards handle more escalations at the same rate. The aggregate cost of human oversight does not fall — it scales with the number of Stewards added, which is exactly the cost structure the Stewardship Model was designed to avoid. Labor-to-Compute Substitution at the task level does not produce Headcount Decoupling at the business level when the Escalation Rate is high enough that each unit of additional autonomous execution volume requires a proportional unit of additional Steward oversight. The Operational Drag of the mis-calibrated boundary is structural — it requires an architectural response. The Agentic Core carries Intervention Threshold calibration standards as architectural defaults across every Arco portfolio build: the 1:100 T1 target and the 1:5 to 1:10 T2 range are not aspirational benchmarks — they are the design specifications against which the threshold is set before the system launches.

The Escalation Rate converts the Judgment Layer / Execution Layer binary from an architectural claim into an architectural confirmation. A system reporting an Escalation Rate within its tier target is demonstrating — not merely asserting — that the boundary was set correctly and is holding in production. Architectural Certainty — the 72-hour MTTI threshold — is the time-dimension confirmation; the Escalation Rate is the volume-dimension confirmation. Both must be within target for the Intervention Threshold to be confirmed as correctly calibrated. The Continuous Regression Loop that detects Logic Decay in the Execution Layer is the architectural mechanism that maintains the integrity of the Escalation Rate over time: as the system’s input distribution shifts and edge cases accumulate, Ghost Trial evidence confirms whether the threshold that produced the target rate at launch still produces it under the current operating conditions. When the Escalation Rate drifts above target because the threshold is too conservative for the current distribution, the correct response is to adjust the threshold — not to add Stewards. Memo #10 develops the Steward role in full: what the Judgment Layer requires of the human who occupies it, and why a Steward operating in the Execution Layer is not performing the function the architecture requires.

An Autonomous Business with all three components correctly implemented has a human-machine boundary that is architectural rather than operational. The Workforce Arbitrage at T1 — 1:100 escalation, 37 to 50 times throughput advantage, near-zero marginal cost — is captured structurally. The Revenue to Headcount Advantage compounds with scale because the Escalation Rate does not scale with transaction volume: additional T1 volume adds autonomous execution, not additional Steward interventions. The Steward’s time is directed at the Judgment Layer — at novel exception classes, architectural improvements, and system-level decisions requiring genuine human assessment — rather than at individual task escalations the Execution Layer should have handled. Memo #03 develops the overhead argument that the Escalation Rate makes precise: every escalation above the tier target is a unit of overhead the architecture was designed to eliminate, being paid because the Intervention Threshold was set to require it. The boundary between human judgment and agent execution is the most consequential design decision in the build. The Escalation Rate is the evidence — available in production, at every tier — that the decision was made correctly.

Technology changes what is possible. Architecture determines where human judgment is required to run.