Something strange was happening with one of our production monitors. Its score kept coming back the same. The monitor measures procedure instruction adherence: whether an agent actually followed the steps prescribed by the methodology. For weeks, it returned the same low mark, run after run. A steady score reads as a steady system, and that was how we understood it. The agents were falling short of the procedure, but at least they appeared to be falling short consistently. That suggested a bounded problem: something we could locate, diagnose, and repair.
But the consistency really wasn't there. The evaluator could return only three values, and the real variation in how agents followed, skipped, and reordered the procedure was being compressed into a scale too coarse to show it. The flat line was not simply a property of the instrument, it was a property of the relationship between the instrument and the variation in the work. And an instrument that can't express instability will report stability forever, whether or not instability exists.
Our evaluation records gave us no way to separate the two. They carried a score, but almost nothing about the judgment that had produced it: which evaluator was running, which version, which criteriai it had applied, what distinctions its scale could express, or which part of the work it purported to describe. Now, none of those fields would have diagnosed the problem for us. They would, however, have made it possible to locate the source of the apparent stability: to ask whether the work had stopped changing or whether the instrument had simply stopped being able to describe the changes.
When a record carries only the result, the properties of the measuring instrument and the properties of the thing being measured arrive fused together. At scale, that fusion happens silently, thousands of times a day, across every system watched this way. So the score travels, but the judgment that gave it meaning does not.
Judgment is being stripped from the handoff
The temptation is to treat this as a tooling problem, and, honestly, our immediate fix looked like one: identify the evaluator, version it, and retire the three-point scale. But the reason the failure sat undetected for weeks was not only that a field was missing. We had built a record that could not identify the conditions of its own judgment, and then trusted it as evidence about someone else’s work.
I have argued before that the intelligence of expert organizations lives partly in the connections between people: in what gets passed upward, what gets corrected, what gets escalated, and what gets absorbed. Hierarchy isn't only a way of allocating authority, it's also a structure through which people positioned at different scopes of the work pass partial judgments to one another. No one sees the whole system directly. The organization becomes capable of judgment by preserving enough signal between those different views.
As more of the work becomes agentic, some of those handoffs themselves become machine-mediated. Agents perform the work, then evaluators inspect particular properties of it, then records carry the resulting observations into monitoring systems, review processes, and decisions made elsewhere. The record is not the whole channel, but it has become part of it. When it carries only a score, the receiving layer gets the conclusion without the basis for understanding what was judged, what the instrument could see, or how much confidence to place in the result. So the handoff is there, but much of the judgment has been stripped out of it.
Meanwhile, I'm seeing the conversation about evaluating agentic systems largely collapse into a single word: outcome. Did the run get where it was supposed to go? It's an understandable step, because it sounds like accountability and produces a number that can be compared. But an outcome is one judgment made at one scope of the work. A system can reach an acceptable answer while skipping required procedures, mishandling dependencies, concealing uncertainty, or repeatedly requiring human correction. It can succeed at the output while failing in execution, coordination, handoff, or fitness.
Judgment therefore has to follow the structure of the work. Each scope exposes a different object to judgment, to a differently positioned reviewer, with different evidence available. What must move between those scopes is not a universal score. It is a structured trace of judgment being exercised: what was assessed, by whom or what, against which criteria, at what scope, on the basis of which evidence, and with what uncertainty or consequence.
That trace does not contain the expert or reproduce everything the expert knows. It carries enough of the expert’s position in relation to the work for another participant to locate the judgment, interpret it, and respond.
Different scopes expose different properties of the work to differently positioned judges. Verification depends not on one score spanning the hierarchy, but on judgments remaining interpretable as they move between its layers.
A judgment needs coordinates
A record cannot carry expertise, at least not in any complete sense. Expertise includes accumulated experience, tacit distinctions, and the capacity to recognize when the available categories no longer fit the case in front of you. What a record can carry is a trace of expertise being exercised in relation to a particular piece of work. If that trace is going to move beyond the person or instrument that produced it, it needs enough coordinates for the next participant to locate the judgment and decide what to do with it.
A complete judgment may draw on many things: the object being assessed, which criteria were applied, which evidence was considered, what uncertainty remained, and the response to the judgment itself. But in our work with evaluation results, three coordinates were repeatedly missing in ways that made the judgments nearly impossible to interpret once they left the systems that produced them: provenance, reference set, and scope.
The first is provenance: who or what made the judgment, by which method, and in which version. The flat line from the opening remained uninterpretable for weeks because the record could not distinguish a deterministic matcher from an LLM evaluator or a human reviewer. Those instruments don't only have quality differences, they can fail in very different ways. A deterministic matcher may be too coarse or too rigid for the variation emerging in the work. An LLM evaluator may change when its underlying model or prompt changes. Human reviewers may apply a standard differently as their experience, instructions, or local conditions change. Once the record identifies the judge, a property of the instrument no longer has to masquerade as a property of the system. We can at least ask the question that took us weeks to formulate: has the work changed, or has the judgment process changed?
The second is the reference set: the cases, examples or the prior population in relation to which the judgment was made. A result is never simply a property of the system. It is a property of the system encountering particular work under particular evaluative conditions. The same evaluator may produce very different distributions over routine cases and unusual ones, or over populations that differ in ways no aggregate score preserves. Naming the reference set does not make two results comparable. It makes it possible to determine whether they can responsibly be compared, and to reconstruct which differences may account for the change.
The third is scope, because judgments travel farther than the work they describe. They leave the run that produced them and appear in dashboards, quality reports, portfolio summaries, and decisions made by people who did not observe the underlying execution. Was the object of judgment a single tool call, a procedural step, an interaction, an entire run, or a pattern across many runs? When scope is ambiguous, a local result is easily stretched to support a broader conclusion. A step-level check becomes evidence of end-to-end quality; a successful run becomes evidence of system fitness. The number stays the same while the claim made from it quietly expands.
Provenance, reference set, and scope do not reproduce the judgment in full. They do something more limited and more useful: they keep the result attached to some of the conditions that give it meaning. They allow a receiving system or reviewer to locate the judgment in relation to the work, rather than encountering a number that does all the talking.
This is where our work in OpenTelemetry fits. The GenAI semantic-conventions effort is developing a shared vocabulary for agentic tasks and evaluation results across tools and systems. Our contribution addresses one narrow part of the larger problem: how an evaluation event can retain enough provenance, reference-set information, and scope to remain locatable after it leaves the system that produced it. A telemetry standard cannot capture expertise, decide what an organization should judge, or ensure that anyone acts on the result. It can keep part of the judgment from being discarded at the handoff.
The point is not that the record will read identically wherever it lands. Different recipients may interpret the evidence differently, challenge the criteria, or disagree with the result. That disagreement is part of the next level of judgment rather than a failure of standardization. The aim is to preserve enough of the original position that the next participant can understand what is being claimed, identify what has changed, and respond without having to reconstruct the entire system that produced it.
A judgment event should not hold the work still. It gives the people and systems around it a common way to say where they are.
The work cannot report its own absences
A judgment event gives the people and systems around the work a common way to say where they are. But it can only describe a place someone reached. It cannot tell us that no one reached a place the work was supposed to go.
Everything so far has been about judgments that exist. An evaluator ran, a reviewer responded, or some part of the work produced enough evidence to be assessed. The problem was whether the resulting record carried enough meaning to remain interpretable after it moved elsewhere. There is another failure that no amount of provenance, reference-set information, or scope can repair: the judgment event that never occurred at all.
A case that never ran emits no result. A required step that was skipped leaves no evaluation of that step. A handoff that never happened produces no disagreement, correction, or escalation. Every event that does arrive may be properly identified, correctly scoped, and entirely defensible, while the whole set of events remains collectively silent about the work that disappeared.
This is why coverage cannot be inferred from the records alone. The observed work has to be compared with some expectation declared outside it: which cases were in scope, which procedures were required, which dependencies should have been resolved, and where review or judgment was expected to occur. The expectation does not need to prescribe every path the work might take. It needs enough structure to distinguish legitimate variation from omission.
But that distinction is itself a judgment, and not necessarily one that can be made from the level where the work occurred. A skipped step may be a failure, an appropriate adaptation to the case, or evidence that the prescribed procedure no longer fits the work. Recognizing the difference requires someone positioned at a broad enough scope, with enough knowledge of the purpose, risks, and alternatives, to understand what the missing work would have contributed. The higher level does not merely count what is present. It judges the relationship between what happened and what the work required.
Expert organizations have always made this distinction, even when exhaustive verification was impossible. An audit reviewer does not infer the completeness of an engagement solely from the quality of the workpapers present. The engagement defines a population, required procedures, relevant risks, and review responsibilities against which the assembled evidence can be assessed. Sampling and risk-based selection do not eliminate the expectation; they make explicit which portion of it will be examined and why. And the reviewer must still judge whether the work performed addresses the underlying risk, not merely whether the prescribed artifacts appear in the file. “Everything we examined was fine” is a different claim from “we examined everything required.”
The same distinction applies to agentic systems. A run can reach the expected output while omitting required procedures, avoiding difficult cases, or failing to invoke a review that should have occurred. An outcome evaluator may score everything it sees correctly and still certify only the surviving path. In this case, the system did not necessarily pass every relevant check, it just passed every check that happened.
A judgment event can tell us where an observer stood in relation to the work. It cannot tell us that no observer ever reached a place where judgment was needed. For that, we need an expected structure outside the events themselves and a judge capable of interpreting the difference between the route and the journey. The structure need not freeze how every run must unfold. It must preserve enough of the work’s purpose for someone at the next scope to recognize when variation is informative, when it is defensible, and when part of the work has quietly disappeared.
The work does not need to stand still
An expected route gives us something against which to recognize omission. It does not mean that every run has to follow the same path, or that the route itself should remain fixed. Expert work changes as people encounter new cases, revise procedures, discover better distinctions, and learn that something they had treated as variation is actually a different kind of problem. A verification system has to preserve that capacity to move in different directions.
Benchmarks solve a different problem. They make comparison possible by defining in advance what will count as the same evaluation: the task, cases, criteria, and conditions under which systems will be measured. This is useful when we already know what object we want to compare. But in deployed systems, the object itself is often still being discovered. The model changes, the workflow changes, the cases change, and experts begin noticing distinctions that no existing evaluator was designed to express. Holding the evaluation regime still may make the numbers cleaner while making the work less visible.
The usual answer is to determine more carefully what must remain constant. But some of that demand for constancy comes from the limitations of our instruments. We held the task, population, and measurement regime still because once several things moved at the same time, we could no longer tell where the difference came from. Was the system behaving differently? Had the work changed? Had the evaluator drifted? Were the people receiving its output applying a different standard? When the result arrived without those relationships intact, variation became indistinguishable from noise.
A judgment system offers another possibility. We do not need to decide every relevant comparison in advance if we preserve judgment consistently enough to reconstruct it later: what was observed, from which position, against which evidence, under what conditions, and what happened in response. That does not make every result comparable. It allows us to discover which results belong together, which differences can be set aside, and which differences reveal that the work or our understanding of it has changed.
The important stability, then, may not reside in the task or the metric. It may reside in the communication structure around them. We can standardize the coordinates without standardizing the conclusion. Different experts and instruments can make different judgments at different scopes, provided that those judgments remain locatable to one another. Disagreement does not have to be removed from the system. It has to remain visible enough to be interpreted.
This is also the deeper problem identified by work on benchmark monocultures. A dominant measurement does more than rank what already exists, it helps determine which forms of value become legible at all (see Lotfi et al. 2026). Work that fits its categories appears as progress, but work whose importance lies along a dimension the benchmark cannot express has to survive long enough for that dimension to become recognizable. The answer is not necessarily a larger or better benchmark, because no fixed measure can anticipate every distinction that will matter under conditions that have not yet occurred. What we need is a system through which new distinctions can enter the record, be communicated across scopes, and eventually become stable enough to measure.
A cockpit provides a better model for that kind of system than a scoreboard does. A cockpit does not hold the aircraft, the weather, the terrain, or even the route constant. Its instruments provide partial views of a changing situation: position, direction, speed, conditions, deviations, and the status of the systems needed to respond. No instrument contains the flight, and no reading determines the right action on its own. Their value lies in allowing differently positioned participants to remain oriented to the same moving system and to communicate when their views no longer agree.
The instruments for agentic work should serve the same purpose. A procedural evaluator may show that a required method was not followed. A human correction may reveal that the method itself did not fit the case. A pattern across many runs may show that what looked like a local exception is becoming a population-level change. These are not competing attempts to produce the one true score. They are bearings taken at different scopes, each partial, each dependent on the judgment available at that position, and each capable of changing what the organization understands about the route.
This is why the goal is not to make the work stand still long enough to measure it. The goal is to preserve the organization’s capacity to locate itself while the work moves: to know what changed, who could see it, what distinction they made, and how that judgment altered what happened next. We do not need to hold everything constant. We need to hold the communication open.
The expert moves with the loops
Holding the communication open is not the same as producing good judgment. Instruments only help if they make the relevant parts of the work observable, if their readings remain interpretable, and if someone capable of understanding the larger situation receives them. None of that happens automatically because a system is agentic.
But agentic systems do change what is possible. Their work unfolds through interactions, tool calls, intermediate decisions, corrections, and responses that can leave traces at a granularity earlier forms of automation rarely made economical to capture. The same underlying technology can also help inspect those traces: checking procedural adherence, comparing evidence, identifying inconsistencies, clustering recurring corrections, or recognizing when a case no longer resembles the population against which the system was designed. These instruments can extend the reach of judgment across more of the work than human review alone could sustain.
They do not arrive already capable of expert judgment. Evaluators are unreliable in ways we are still learning to characterize, and they have not acquired their standards through years inside the review hierarchies that teach practitioners which distinctions matter, which deviations are defensible, and which apparently small failures change the meaning of the whole. A model can apply a criterion without understanding why the organization chose it. It can detect a departure from a route without knowing whether the route or the departure is wrong.
This is why the evaluators have to enter the same verification structure as the systems they inspect. Their judgments need provenance, scope, reference sets, and evidence. Their coverage has to be compared with what the work required. Their disagreements and failures have to remain visible to people positioned to interpret them. The instruments themselves become objects of judgment, not exceptions to it.
That recursion need not become machines vouching for machines all the way down. The loop remains anchored outside itself. Experts define the purpose of the work, recognize which distinctions matter, judge whether an adaptation is defensible, and revise both the route and the instruments when experience reveals something the existing system could not see. Their corrections are not merely exceptions to be cleared from a queue. They are signals through which the organization learns what its evaluation system should become capable of recognizing.
The opportunity, then, is not to replace expert judgment with sufficiently numerous scores. It is to give judgment a wider field of view and a more durable path through the organization. A local correction can become evidence of a recurring failure. A disagreement between reviewers can reveal an unstable standard. A pattern across runs can show that the population, procedure, or purpose has shifted. Each instrument remains partial, but the communication among them allows the organization to revise its understanding while the work continues.
A cockpit is valuable for the same reason. It does not eliminate uncertainty or make every instrument correct. It creates an environment in which partial readings can be located, compared, challenged, and turned into coordinated action. Reliability comes not from choosing the one score that finally contains the truth, or from holding the world still long enough to measure it. It comes from building a system in which judgment can continue to move, and the organization can continue to learn where it is.`
References:
Lotfi, S., Iranmanesh, A., Naghashyar, L., Shirali, A., Nateghi Haredasht, F., Koyejo, S., Torr, P., Lee, Y. S., Barez, F., Lehman, J., Norvig, P., and Narayanan, A. "Position: There Are Futures That Benchmark-Driven AI Cannot See." ICML 2026, oral. https://icml.cc/virtual/2026/oral/71151 (paper PDF on OpenReview: https://openreview.net/pdf/c0833e1fef52e998ef8d65944caaa3aae0eaa35c.pdf)
OpenTelemetry GenAI Semantic Conventions, gen_ai.evaluation.result event. https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-events.md
Srivats, M. "Add evaluator provenance attributes to gen_ai.evaluation.*," open-telemetry/semantic-conventions-genai, PR #359. https://github.com/open-telemetry/semantic-conventions-genai/pull/359
