Attribution After Grounding

What the previous note established

Continuous Generation, Inference Loops, and the Return of Real Grounding argued that a model which continues to generate is no longer evaluated as a finished artifact. Competence is persistence inside a place: objects, tools, hands, and the intervals in which nothing decisive happens. A stock corpus is not that place. Where the environment is absent, it has to be commissioned, under a rubric written before capture, with consent and chain of custody bound to the file.

That argument settles a condition. It does not settle a result. An organization can acquire a faithful environment and still be unable to say whether the agent improved. Acquisition and attribution are different claims. Conflating them is how a data programme reports volume and treats the report as evidence of capability.

The question is now being posed beyond embodied systems. micro1’s October index lists “Can a company tell whether its AI agent is getting better?”, inside a research programme concerned with where human time should be spent. For a grounded loop the coordination problem is specific. Which contributor, in which environment, against which pre-registered failure — and did that contribution move an outcome that can be attributed?

Why proximal scores fail

A continuous policy is a trajectory. It enters a state, misses a transition, recovers, and terminates or does not. A score defined on a frame, or on a short resemblance to a reference clip, is insensitive to recovery. Measured that way, a policy can improve on the sample and degrade at remaining in the environment.

Three identification failures are common enough to treat as the default.

Saturation. Short-horizon actions are already near ceiling. New capture, unless the rubric names the long horizon, over-samples those actions. The mean rises. The failures that determine deployment do not.

Leakage. The environment used for training is the environment used for test. The policy memorizes a site, a hand, a lighting regime. On a second site the gain vanishes. The report still records improvement, because the report never left the first site.

Confounding. Parameters, harness, and environment change in the same interval. Some measured quantity moves. The design does not identify which factor moved it. A later reader cannot reconstruct the interval. They can only accept the narrative.

Additional capture does not repair these failures. A pre-registered definition of improvement does.

Three factors, one manipulation

A grounded loop has three factors. An attributable comparison manipulates one and holds the other two fixed.

Environment. New sites, tools, and failure modes, captured under a rubric that specifies the long horizon: the missed transition, the wrong object, the uneventful interval, the recovery. This is the factor a capture programme controls.

Parameters. A training run on that capture, with the previous checkpoint retained as the baseline.

Harness. The instruction, the tool schema, the retry rule, the timeout. The harness changes the score without changing what the policy has learned about the environment.

A result that supports the word “improved” changes one factor. If all three move, the run is uncontrolled and should be reported as such.

The held-out set must itself be an environment: a second site, a second contributor, or a later interval, designated before training and excluded from the training manifest. The outcome is whether the trajectory completes the task there, recoveries included. It is not whether a sampled frame is familiar.

Benches built for text and for browser-mediated errands answer a different question. A legal-reasoning bench, a contract-review bench, or a personal-agent bench can measure judgment in language and competence at tasks whose state lives in software. They do not measure whether a policy remained inside an environment it had not been shown. The instrument has to be permitted to record that the environment was new.

Identity as an identification variable

Duration is not a unit of improvement. An hour from a practitioner and an hour from someone performing the appearance of the practice are different observations. In the file they can be indistinguishable.

Persistent contributor identity is therefore part of the measurement, not a credential attached after the fact. It is what makes it possible to ask whether a gain came from a different class of contributor or from more of the same distribution. If graders are drawn from the same pool as collectors, the rubric evaluates itself. The two roles have to be separated. The person who writes the rubric should not be rewarded for the volume shipped against it.

The allocation of expert time follows from the same point. The next hour belongs on the failure the previous evaluation could not see. It does not belong on another easy episode because a volume target is short. The rubric is the assignment. A rubric that names no failure spends the hour on a count.

This is the operational content of the primitives in From Models to Reality: context, provenance, and persistent identity. They are not annotations for a catalog. They are what make an improvement claim identifiable after the fact.

The minimum record

An organization can claim that an agent improved when the record contains four items.

1. The failing behavior, named before capture, inside an environment excluded from the training run.
2. The baseline: the same harness, the previous parameters, that held-out environment.
3. A single manipulation: the environment, the parameters, or the harness.
4. The chain on the new capture: who collected it, under what consent, against which rubric, and who graded the held-out run.

Absent any one of these, the supported statement is that more was collected. It is not that the agent improved.

That standard is stricter than a leaderboard of short tasks. A leaderboard is an appropriate instrument when the task is the product. It is a weak instrument when the product is persistence in a place the leaderboard has never contained.

What Harbor will and will not assert

Harbor commissions the environment and keeps the chain on the asset. That is the environment, and the record of the environment. It is not the training run, and it is not the harness. Delivery of hours does not license a claim that a customer’s agent improved.

The delivery can carry the portion of the record that attribution requires from the capture side: the rubric written first, the identity of the person in the environment, the consent, and a split that keeps a held-out environment out of the training manifest. The laboratory still has to run the baseline. A harness change in the same interval is their confound. The manifest should expose it, rather than absorb it into an undifferentiated drop.

Prompt-to-Earn remains a separate instrument. A conversation submitted by choice is neither this environment nor this outcome. The funds are separate. The work is separate.

Conclusion

The previous note held that a stock set is not an environment. This note holds that an environment is not an attribution. Commission the place. Then withhold the claim of improvement until one manipulation, on an environment held out in advance, moves a failure named in advance. The contributor in the environment belongs in that claim. So does the grader who was not permitted to score their own hour.

References

[1] Harbor Research. Continuous Generation, Inference Loops, and the Return of Real Grounding. September 2026. <https://harborml.com/research/continuous-generation-and-real-grounding>

[2] Harbor Research. From Models to Reality. August 2026. <https://harborml.com/research/from-models-to-reality>

[3] micro1. “Can a company tell whether its AI agent is getting better?” October 2026, listed on the micro1 research index. <https://www.micro1.ai/research>