Why weren’t Recovery Latency and Rollback Cost part of the original Intervention Dependency axis definition?

The original axis was built around the question of frequency — how often does the system require intervention, measured through MTTI and the Escalation Rate. That question is foundational and correctly came first. Once frequency is established as low, the next honest question — how well does the system perform during the interventions that remain — becomes visible, and these two metrics are the answer to that next question.

Can a business have a strong Intervention Dependency score and still have concerning Recovery Latency or Rollback Cost?

Yes, and this is precisely the gap this memo identifies. A long MTTI and a low Escalation Rate confirm that intervention is rare. They say nothing about what happens during the interventions that do occur. A business can look excellent on the existing axis while having a Deterministic Failure design that is thinner than its quiet uptime suggests.

How is Rollback Cost different from the reversibility window already discussed in the Authorization Trap’s promotion criteria?

The reversibility window is a design precondition evaluated before a decision class is promoted to autonomous operation, part of the [Authorization Trap](https://arcoventure.studio/lexicon/authorization-trap)’s promotion criteria. Rollback Cost is an operational measurement taken after an actual error occurs — what did it cost, this specific time, to reverse this specific incorrect action. The reversibility window is a gate a decision class must pass before promotion; Rollback Cost is a running measurement of what passing that gate has actually cost in practice.

Should Recovery Latency and Rollback Cost be reported for every intervention, or only for consequential ones?

Tracking both for every intervention gives the most complete picture, but the metrics matter most for consequential failures. A business with limited measurement capacity should prioritise tracking both specifically for failures in decision classes with a meaningful Rollback Cost exposure, since that is where an unmeasured gap between strong MTTI and poor actual recovery performance is most consequential.

Does a high Rollback Cost mean a decision class was promoted to autonomous operation too early?

Not necessarily on its own, but a pattern of high Rollback Cost in a specific decision class is a strong signal worth investigating. A single high-cost rollback can be a genuine edge case. A recurring pattern suggests either the reversibility window was set too optimistically at promotion time, or the Deterministic Failure design for that class needs to catch the error earlier, before the cost of reversal accumulates.

What MTTI Doesn’t Measure

MTTI (Mean Time to Intervention) is the average time between required human interventions in an agentic system, with Arco’s target set above 72 hours on core revenue loops. It is the primary operational test that distinguishes an autonomous business from an automated one, and it is one of the two instruments — alongside the Escalation Rate — that the Intervention Dependency (ID) axis of the Autonomy Spectrum Framework is scored from. What MTTI has never measured, and was never designed to measure, is what happens during the intervention itself: how long it takes to actually resolve, and what undoing an incorrect action costs once the system is fixed. A business can have an excellent MTTI — long, quiet stretches between interventions — and still recover badly and expensively the moment one is actually required.

Two gaps, not one

Recovery Latency is the time from failure detection to restored autonomous operation — a distinct measurement from MTTI, which tracks the interval between interventions rather than the duration of resolving any one of them. Deterministic Failure — a failure mode that is predictable, fully logged, and recoverable by design, the architectural standard Arco engineers into every autonomous system so that when the system breaks, it breaks safely — already establishes the design standard this measurement depends on. What Deterministic Failure specifies as an architectural requirement, Recovery Latency measures as an operational outcome — confirming that the design standard is actually producing fast recovery in practice, not simply that the failure was logged correctly when it happened.

Rollback Cost is the financial and operational cost of reversing an incorrect autonomous action once it is identified — a measurement distinct from both MTTI and Recovery Latency, because an intervention can be resolved quickly and still leave behind a costly correction. A payment sent to the wrong account, a contract generated with an error, a customer-facing change that needs to be walked back: each of these can have a short Recovery Latency, in the sense that the system stops making the error quickly once flagged, while the cost of undoing what already happened remains substantial. Rollback Cost is not the same measurement as Rebuild Tax, which is the one-time, deferred cost of re-architecting a system built under MVP shortcuts once a business reaches scale. Rollback Cost is a recurring, per-incident operational cost — how expensive it is, this specific time, to reverse this specific error — not a structural liability accumulated from how the system was originally built.

Why both belong to the existing framework, not a new one

Neither of these metrics describes a new dimension of autonomy. Both describe how well an already-scored business performs on the dimension Intervention Dependency already names. A business with a strong ID score — long MTTI, low Escalation Rate — has demonstrated that it rarely needs a person. Recovery Latency and Rollback Cost ask the next honest question the existing framework leaves open: when it does need one, how well does that actually go. This is why both belong as supplementary measurements to the Intervention Dependency axis and to the Deterministic Failure design standard, rather than as the foundation of a separate framework. The Autonomy Spectrum Framework already asks how autonomous a business is. These two metrics ask how gracefully it handles the moments it isn’t.

What good and bad look like in practice

A business with excellent MTTI, a fast Recovery Latency, and a low Rollback Cost has genuinely achieved what Architectural Certainty is meant to describe: not just infrequent need for a human, but competent, cheap resolution on the rare occasions a human is actually required. A business with excellent MTTI but a slow Recovery Latency and a high Rollback Cost has a different and more concerning profile — the system runs quietly for a long time, which looks identical to genuine reliability, right up until something goes wrong, at which point the business discovers its Deterministic Failure design was thinner than its MTTI suggested. This is a related but distinct honesty check from the one Nominal MTTI already performs. Nominal MTTI asks whether a long MTTI reflects genuine reliability or an unmonitored system. Recovery Latency and Rollback Cost ask a question that remains even when MTTI is genuinely earned: how well does the business actually perform at the one thing MTTI cannot tell you anything about — the intervention itself.

The Operator’s Verdict

A long MTTI is necessary. It has never been sufficient. Measure how fast the system actually recovers once an intervention is triggered, and what it costs to undo whatever went wrong before the recovery completed. A business that scores well on Intervention Dependency but has never measured either number does not yet know the full shape of its own reliability — only the frequency with which that reliability is tested.

Technology changes how rarely the system needs a person. Recovery determines what happens to the business in the moments it does.

KEY TAKEAWAY

What are Recovery Latency and Rollback Cost, and how do they extend the Autonomy Spectrum Framework’s existing metrics?

Recovery Latency is the time from failure detection to restored autonomous operation, distinct from MTTI, which measures the average time between required interventions rather than the duration of resolving any single one. Rollback Cost is the financial and operational cost of reversing an incorrect autonomous action once identified, distinct from both MTTI and Recovery Latency, because an intervention can be resolved quickly while the cost of undoing what already happened remains substantial — and distinct from Rebuild Tax, which is a one-time, structural re-architecture cost rather than a recurring per-incident operational one. Both metrics extend the existing Intervention Dependency axis of the Autonomy Spectrum Framework and the Deterministic Failure design standard, rather than introducing a new dimension of autonomy. Source: Arco Venture Studio.