Diagnose Only What the Agent Actually Visited

TL;DR for operators

A quiet agent trace is not necessarily evidence that an underlying component is healthy. It may mean the deployed policy barely visited the states where that component could reveal a fault.

Zhu and Chang’s simulation study¹ treats this as an eligibility problem before treating it as an anomaly-detection problem. A component must first have repeated support under clean reference operation and then comparable support in the current runtime stream. Only after both conditions hold does the system interpret aggregate fault signals.

The strongest empirical result supports that ordering. Across 540 unit-arm observations nested within 20 represented physical-component clusters, a model based on traffic through fault-affected states achieved a negative log likelihood of 0.3055 per row versus 0.4319 for a model based on the fraction of affected cells. That is a 29.3% reduction, with a component-bootstrap interval for the improvement of 0.0593 to 0.1918 nats per row.

For assurance workflows, the practical lesson is to distinguish four decisions: whether a component is observable, whether supported telemetry signals a fault, whether environmental drift requires the reference baseline to be refreshed, and whether downstream localization is justified. The study supports that design principle inside its simulator; it does not establish universal support thresholds or real-world localization accuracy.

A quiet trace cannot clear an unobserved component

Consider an operations team reviewing an agent that appears normal in aggregate. Transactions are completing, aggregate actions remain within familiar ranges, and the monitoring dashboard shows no strong anomaly. The tempting inference is that the underlying workflow has also been exercised sufficiently to support that conclusion.

Closed-loop agents complicate that inference because their own actions influence what states they encounter next. A component can exist in the policy yet receive little runtime traffic. If a fault affects states that the policy almost never visits, a large nominal perturbation can leave only a weak aggregate trace. A smaller perturbation applied to heavily visited states can be much easier to observe.

That distinction changes what “no signal” means. It may be evidence against a fault when the relevant states were well observed. It may instead be an absence of usable evidence when the states were barely visited.

The paper operationalizes this distinction with affected clean traffic: the share of clean-policy visits that fall in cells altered by a hypothetical fault. Nominal cell coverage asks how much of the editable state space changed. Traffic asks how much of the behavior actually generating the trace passed through those changes.

Support comes before anomaly scoring

The study uses a simulated hotel-pricing agent divided into 24 physical policy components and evaluated across three demand regimes, yielding 72 regime-component units.

Diagnosis begins with two support checks.

The reference-map gate asks whether a component is repeatedly observable under clean operation. Each unit needs at least 12 observations in a partition, and at least 14 of 15 clean reference partitions must satisfy that requirement. Fifty-five of 72 regime-component units passed, representing 20 of the 24 physical components.

The matched runtime gate then asks whether the current stream still supplies comparable evidence. Reference and current data must jointly satisfy the support requirement in at least 14 of 15 matched partitions. Of the 55 reference-admitted units, 54 passed this second gate.

The remaining unit was not classified as healthy. It abstained.

That state distinction is central. Historical observability does not authorize interpretation indefinitely. If current policy behavior moves away from a previously covered region, the evidentiary basis for diagnosis can disappear even though the component remains present in the system.

Traffic explains signal better than nominal fault size

The paper’s main structural comparison tests whether fault visibility is better explained by affected clean traffic or by affected cell fraction.

The traffic model produced total negative log likelihood of 164.9866, or 0.3055 per row. The cell-coverage model produced 233.2453, or 0.4319 per row. The difference was 0.1264 nats per row, corresponding to a 29.3% reduction relative to the cell model.

The component-cluster bootstrap interval, 0.0593 to 0.1918, remained above zero under the frozen criterion.

The detection pattern is consistent with that result. At the lowest targeted traffic level, 0.15, neither the uniform nor flow-weighted fault family produced a stable detection among the 54 runtime-admitted units. At the highest target, 0.95, both produced 54 of 54.

The researchers also tested whether knowing the fault-mask family still materially improved prediction after traffic exposure was included. Adding mask family and its interaction with traffic improved log loss by only 0.0015 nats per row. Its one-sided bootstrap upper bound was 0.0066, below the precommitted 0.01-nat practical-sufficiency margin.

That is a bounded result, not evidence that fault structure never matters. It says that for these two frozen mask families—uniform and flow-weighted—traffic captured most of the predictive distinction relevant to the study’s stable-signal outcome.

Abstention, false admission, and drift are different decisions

The paper deliberately avoids combining every monitoring outcome into one accuracy number.

The false-admission result is especially easy to overread. Observing zero cases among 20 components does not establish a zero false-positive rate. The formal result is the upper bound of 0.1391 under the study’s working assumptions, which passed its precommitted threshold of 0.20.

Drift receives a different treatment. Demand-regime changes alter inventory trajectories and therefore policy visitation. That can change which components are diagnosable even when no component itself is faulty. The frozen shift detector therefore feeds a reference-map recomputation pathway rather than declaring a fault.

For deployed assurance systems, Cognaptus would treat these as different workflow states: interpretable, unsupported, baseline potentially stale, and supported signal requiring investigation. Collapsing them into a binary healthy/faulty dashboard discards information the operator needs.

Better localization cannot repair weak exposure

A separate development audit compared exact minimum hitting set localization with a propagation-aware greedy alternative. They selected identical supports in all 12 audited scenarios and both exactly recovered 9 of 12.

That does not establish that greedy localization is generally equivalent to exact optimization.

Follow-up analysis found that upstream evidence already contained planted-component singleton conflicts. Those singleton observations resolved the relevant ambiguity before the localization optimizer had much opportunity to matter.

This is a useful evaluation warning beyond this simulator. A sophisticated downstream diagnostic method can appear unnecessary—or exceptionally effective—because the benchmark’s evidence construction has already encoded most of the answer. Optimizer comparisons should therefore be interpreted together with the ambiguity remaining at their input.

What an assurance team could change

The paper directly supports an exposure-first interpretation rule inside its simulated environment. Extending that rule to production is an engineering inference rather than an experimentally validated deployment result.

A monitoring system could maintain an exposure map alongside anomaly metrics. Before allowing a component-level health claim, it could verify that reference operation repeatedly exercised the component and that current operation still provides matched support. Unsupported cases could request additional observation, targeted testing, or human review instead of silently appearing healthy.

This also changes where assurance effort is allocated. State-space coverage counts treat visited and unvisited states similarly. Traffic-weighted coverage directs attention toward the states actually carrying operational behavior while still exposing low-traffic regions as evidence gaps rather than clearing them by absence of anomalies.

The design further suggests keeping environmental drift management outside the component-fault escalation path. A shift in usage can invalidate a monitoring baseline without implying an internal failure.

The thresholds are not portable defaults

The evidence comes from one synthetic hotel-pricing environment with 24 physical components. Formal faults shift selected target-field cells by one action bucket, and the primary inferential results rely on approximately 20 represented component clusters.

The clean-versus-clean calibration also uses an exchangeable-independent Bernoulli working assumption at component level despite common random numbers, while cluster-bootstrap intervals based on roughly 20 clusters may have imperfect coverage. The study does not confirm planted-fault localization accuracy, and its fault-family sufficiency result applies only to the two evaluated mask families and the chosen 0.01-nat margin.

Most importantly, the support requirements—12 observations per partition, 14 of 15 qualifying partitions, and the paper’s signal thresholds—are properties of this frozen protocol, not general operating standards.

The transferable proposition is narrower: before interpreting an aggregate trace as evidence about a component, establish that the policy produced enough relevant exposure for that component to have been observable at all.

An agent-monitoring system that cannot make that distinction risks converting unvisited states into false reassurance. In safety and assurance work, knowing when the trace does not support an answer can be as operationally significant as detecting the anomaly itself.

Cognaptus: Automate the Present, Incubate the Future.