Verification Bandwidth and the ACE Economy

The bottleneck moves from intelligence to verification.

Catalini, Hui and Wu separate the cost to automate from the cost to verify. As models, compute and accumulated knowledge make autonomous execution cheaper, verification remains constrained by human time, expertise and real-world feedback latency.

The result is a widening gap between what machines can execute and what humans can afford to validate. In that model, raw agent output is not the same as realized productive value. The verifiable share becomes the binding factor.

Verification scarcity is ACE scarcity made operational.

ACE — Attentive Cognitive Evaluation — treats consequential human attention as a scarce production factor rather than generic labour. One ACE-second is focused human attention applied to a defined decision surface with sufficient context, relevant judgment, valid authority and responsibility for consequence.

The external verification-bandwidth argument reaches the same scarcity from another direction: machine execution can expand rapidly while trustworthy human evaluation cannot.

Output is not completion.

The external paper asks how much agentic output can be verified. The Governed Completion Unit sharpens the production question: how much work reached valid, governed completion with the required authority, evidence and outcome state?

ACE and GCU are complementary: ACE measures scarce legitimate human evaluation; GCU measures governed completed outcome. Together they shift the metric away from tokens, agent activity and attempted work.

Verification after the fact is not enough.

The paper is primarily verification-centric: improve observability, provenance, independent checking, human augmentation and liability structures. HEIMEL adds a pre-consequence boundary.

That distinction matters for irreversible actions. A payment may already have settled, data may already have left the boundary, or a physical actuator may already have moved before post-hoc review begins.

Evidence is verification compression.

Catalini, Hui and Wu explicitly argue that cryptographic provenance can reduce verification cost. Process evidence such as model version, tools invoked, permissions, data sources, execution traces, signatures and attestations becomes economically valuable when execution is abundant.

In the VALO model, a receipt is therefore not merely audit history. It compresses the cost of reconstructing what happened.

Veritas can therefore be understood as infrastructure that converts execution history into reusable verification capacity.

Traditional AI ROI can overvalue activity.

ACE-Adjusted Return on Intelligence treats review burden, rework, false completion, authority failures, control cost and scarce human attention as real costs of AI production. A system can produce less raw output and still have higher economic value if it creates more governed completion per ACE-second.

This also extends the Solland Paradox. As intelligence becomes cheaper and more useful, scarcity does not disappear; it migrates. Verification bandwidth is one of the next scarce layers.

The stronger claim is trustworthy consequence capacity.

Verification is necessary, but it is only one component. Trustworthy consequence capacity also requires valid principal intent, bounded delegation, current authority, constraint satisfaction, independent consequence enforcement, completion evidence, exception handling, liability and accountability.

This synthesis should be falsified, not merely repeated.

Status: research synthesis and working economic model. External claims are attributed to the cited source. VALO-specific extensions are hypotheses and architecture interpretations that require continued empirical testing.