Decision behavior

Uncertainty and abstention: why declining to decide is a required output

Definition

Abstention is an explicit system output that declines to produce a treat or no-treat decision because the evidence or the operating conditions are insufficient. It is not an error and not a silent default. Recording why a decision was withheld — degraded imagery, ambiguous life stage, conditions outside the evaluated envelope, missing field context — is what keeps a confident wrong answer out of the field.

This page explains why abstention is designed in rather than tolerated, and what it costs to leave it out.

Published: August 6, 2026
Last updated: August 6, 2026

What does abstain mean?

It means the platform has declined to produce a treat or no-treat decision for a given area and time, because the evidence or the operating conditions were insufficient to support either. The reason is recorded alongside the abstention.

An abstention is a real output with a real record. It names what was missing or unreliable and, where possible, what would resolve it — a recapture under better conditions, a human review of an ambiguous label, or a threshold policy that covers the crop stage in question.

Why is abstention a required output rather than a failure?

Because the alternative is worse than nothing. A system without abstention converts every hard case into a confident-looking treat or no-treat, and the reader has no way to tell those apart from the cases it actually resolved.

Both directions carry cost. A false treat means product applied where it was not warranted, at expense and with resistance-management consequences. A false no-treat means a developing population left alone. Neither is improved by a system that hides which answers it was sure about.

Abstention keeps error rates interpretable: the decisions the system did make are the ones it could support.
It routes the hard cases to people, who are better at ambiguity than a fixed model is.
It makes gaps visible. A cluster of abstentions is a diagnosis — of capture practice, of policy coverage, or of a model operating outside its envelope.
It is honest about a real limit rather than engineering around it with a default.

When should a decision system abstain?

Whenever a required input is missing, unverifiable, or outside the conditions the model and policy were evaluated for. The trigger is the state of the evidence, not the confidence number a model happens to return.

Conditions and the intended outcome
Condition	Intended outcome
Image quality below the threshold for the task	Abstain, with the failing quality measure recorded.
Life stage visually ambiguous	Abstain pending human adjudication of the label.
Capture metadata or calibration missing	Abstain — the asset is an invalid input, not a weak one.
Data rights do not cover this use	Abstain, and exclude the asset from the decision entirely.
Crop stage or region outside the evaluated envelope	Abstain, with the out-of-envelope condition named.
No applicable threshold policy version	Abstain — there is no rule to apply, so there is no decision to make.
Coverage too sparse to represent the area	Abstain for the uncovered area rather than interpolating across it.

How is uncertainty different from abstention?

Uncertainty is a property carried by the evidence; abstention is a decision taken because of it. Every decision records uncertainty, including the ones that resolve to treat or no-treat.

The distinction matters because a raw model confidence score is not uncertainty in the sense that matters here. A model can be highly confident about an image that should never have been analyzed. Uncertainty as recorded by the platform is intended to combine input validity, calibration state, biological ambiguity, coverage, and whether conditions fell inside the evaluated envelope.

A confidence score that has not been calibrated against observed outcomes tells you how the model feels, not how often it is right.

What happens after an abstain result?

The abstention is recorded with its reason and, where one exists, the action that would resolve it. Nothing is applied and nothing is inferred; the decision simply does not exist for that area and time.

1The reason is recorded against the area and time the abstention covers.
2Where the cause is a capture defect, the requirement for a valid recapture is stated.
3Where the cause is biological ambiguity, the case is routed for adjudication.
4Where the cause is a policy gap, the gap is surfaced for review rather than filled with a default.
5The grower and their advisers decide what to do in the absence of a decision, as they would without the platform.

Does abstention make the system less useful?

It reduces the number of outputs and raises what each one is worth. A decision that arrives with a stated basis can be acted on; one that might be a disguised guess has to be second-guessed every time.

There is a legitimate operational concern behind the question: a system that abstains constantly is not usable. The answer is to reduce abstentions by fixing their causes — capture practice, policy coverage, calibration, validation scope — rather than by lowering the bar at which the system is willing to answer.

How is an abstention recorded?

As a full decision record with an abstain outcome, carrying the same source assets, versions, and review history as a treat or no-treat. Abstentions are evidence about the system and are retained as such.

Discarding abstentions would destroy the most useful signal the platform produces about itself. The pattern in what it declines to decide is where capture problems, policy gaps, and envelope violations show up first.

Inputs, outputs, limitations, and evidence status

Inputs

Measured image quality and asset validity state.
Model confidence together with its calibration state.
Biological ambiguity flags, including unresolved life-stage labels.
Capture coverage relative to the area a decision would cover.
Whether crop, stage, region, and capture conditions fall inside the evaluated envelope.

Outputs

An explicit abstain outcome for a defined area and time.
The recorded reason for abstaining.
The resolving action where one exists: recapture, adjudication, or policy review.
Uncertainty recorded on every decision, not only on abstentions.
A full decision record, retained identically to treat and no-treat outcomes.

Limitations

Abstention prevents an unsupported decision; it does not supply a correct one.
A high abstention rate indicates unresolved capture, policy, or scope problems.
Abstention does not tell the grower what to do instead — that remains their judgment.
Uncertainty as recorded is not a calibrated probability of field outcome.

Evidence status

The abstention behavior described here is the intended platform design.
No abstention rates, error rates, or calibration figures are published.
Confidence calibration against field outcomes requires validation data that is not yet published.
The first validation domain is Colorado potato beetle in potatoes.

Terms used on this page

Definitions are shared site-wide. The full list is in the glossary.

Abstention: An explicit output declining to produce a treat or no-treat decision because the evidence or the operating conditions are insufficient. Abstention is a designed result, not an error and not a silent default.
Uncertainty: The recorded degree to which the inputs to a decision are unreliable or incomplete — image quality, ambiguous life stage, sparse coverage, or a model operating outside its evaluated conditions. Uncertainty is carried into the decision rather than discarded.
Operating envelope: The stated conditions under which a model and policy have been evaluated — crop, pest, growth stage, region, capture geometry, light, and equipment. Results produced outside the envelope are flagged, not silently reported.
Calibration: Two related things: the camera calibration that maps image pixels to field geometry, and the confidence calibration that determines whether a model's stated confidence corresponds to its observed reliability. Both are recorded per decision.
Treatment decision: An explicit output stating treat, no-treat, or abstain for a defined area of a field at a defined time, together with the evidence and policy version that produced it. A detection is not a treatment decision.