The first row looks ready
A researcher opens a ranked list of compounds for a follow-up assay. The first row has a score of 0.91, a green badge, and an “Approve” button. Before reserving material, the researcher needs to answer a less tidy question: what exactly would approving this row commit the laboratory to doing?
Perhaps the score estimates activity under a particular assay protocol. Perhaps it only orders candidates by resemblance to earlier hits. Neither interpretation says whether the compound is available, whether its measurements are comparable, or whether repeating a familiar candidate serves the next experiment.
Consider a representative research team selecting compounds for experimental follow-up. This is an illustrative product case, with invented scores and workflow details, rather than a client implementation. The model assists prioritization; it does not make treatment recommendations. The design task is to keep the model's estimate, the evidence supporting its use, and the researcher's experimental choice separately visible.
Name what the number estimates
The first change is to replace “Confidence: 91%” with an accurate description of the output. If the model produces a ranking score, call it a ranking score and explain its direction. Converting 0.91 to a percentage does not establish a probability interpretation.
When an output is presented as a probability, calibration becomes relevant. The scikit-learn calibration documentation explains the relationship between predicted probabilities and observed frequencies: among cases assigned probabilities near a given value, the positive outcome should occur at approximately that rate for a well-calibrated classifier. This is an assessment across cases, not a guarantee about the compound currently on screen.
A high-performing model should not acquire a “calibrated” label by assumption. In their 2017 study, Guo and colleagues found calibration problems in the neural networks they evaluated on image and document classification datasets. That study does not establish the performance of a compound-ranking model. It supports the narrower reason to investigate calibration explicitly rather than infer it from other performance claims.
For the representative product, a probability label should link to the applicable evaluation record, including the outcome definition and evaluated data. If that evidence is unavailable, the interface should state the limitation. An extra decimal place cannot supply it.
Put applicability beside the score
Suppose the top-ranked compound was measured using a different protocol from the one covered by the model's evaluation. The application knows the protocol identifiers, but its first design shows that difference only in a metadata drawer. The ranking screen still makes the candidate look ready for selection.
In the revised design, the row says “Protocol outside evaluated scope.” The researcher can inspect the measurement and evaluation scope without losing their place in the shortlist. For this workflow, the candidate moves to a separate review path before it can enter the model-assisted experiment plan.
This check has a defined basis: a comparison against recorded protocol scope. It is not a universal detector of unfamiliar data. If the protocol identifier itself is missing, the product should say that applicability cannot be established. It should not silently treat an unknown protocol as supported.
The model may also use an imputed measurement. Showing the resulting feature as if it had been observed would obscure a distinction the researcher may care about. Keep the observed input, substituted value, and relevant preparation step inspectable. The main screen can summarize the limitation; it need not display the entire processing history beside every score.
Explain the experimental choice
Now consider a second compound with a lower illustrative score of 0.84. Its protocol is within the recorded scope, material is available, and it addresses a gap in the team's planned comparison. The researcher may choose it before the first candidate. That choice is understandable only if the product records the experimental objective alongside the ranking.
Replace the ambiguous “Approve” action with “Add to experiment plan.” The confirmation identifies the intended assay and the material reservation it will request. The selected record retains the model score and version, the input snapshot, and the researcher's reason for selection. It does not rewrite the lower score to make the chosen compound appear model-preferred.
The interface also needs useful actions for other outcomes. “Request additional evidence” can send the candidate to an assigned reviewer. “Defer” should retain a reason and a way to return to it. These labels describe proposed workflow choices; the right set depends on who actually owns the next step.
A team selecting a diverse experimental panel may need a different policy from one confirming the highest-ranked candidates. The screen should name the active objective and selection rule. Otherwise, two researchers can follow different reasonable policies while believing they are performing the same review.
Make disagreement usable
A free-text box beneath an approval button does little if the reviewer cannot inspect inputs, defer the task, or ask someone with the right expertise. In this case, the reviewer needs access to the relevant evidence and authority to change the plan before material is reserved.
The voluntary NIST AI Risk Management Framework 1.0 addresses defined responsibilities for human oversight in GOVERN 3.2 and oversight processes in MAP 3.5. It also includes user feedback and override mechanisms in post-deployment monitoring under MANAGE 4.1. These are framework practices, not a prescription for this particular screen or a claim of regulatory compliance.
For the representative team, disagreement should retain a specific reason: protocol mismatch, unavailable material, conflicting evidence, or a different experimental objective. Allow additional context where those categories are insufficient. Choosing a lower-ranked compound because it is available is different from concluding that the model estimated activity incorrectly.
That distinction matters when the team later evaluates the tool. A reviewer's choice is a workflow event. It should not automatically become a training label stating whether the compound was active. The eventual assay outcome, its measurement conditions, and the selection decision deserve separate records.
Decide how much review the task needs
Detailed review has a cost. Opening evidence for every candidate may overwhelm a team working through a large exploratory screen. A simpler ranked queue can be appropriate when the immediate action is inexpensive, reversible, and explicitly limited to prioritization.
For this case, routine in-scope candidates could share a batch review view, with exceptions brought forward individually. Before adopting that arrangement, the team would need to establish which actions may be grouped, what evidence remains available, and how a reviewer removes a candidate. A batch confirmation should describe the actual commitment and selected set.
A workflow involving scarce material or consequential downstream commitments has a stronger reason to require individual review. Adding the same confirmation step everywhere would avoid making this design decision without resolving it. Review effort should follow the action and the unresolved evidence.
Follow the choice through to the outcome
Test the interface with both illustrative candidates. Ask a researcher to explain the score, identify the protocol limitation, and select a compound for a stated experimental objective. Then ask another colleague to reconstruct that choice from the saved plan. Include a deferred candidate to check that its follow-up reaches an owner.
After deployment, compare model outputs with observed assay outcomes where they become available, keeping model versions and assay context identifiable. Also inspect unresolved reviews and reasons for changing the proposed order. Agreement with the ranking alone would not tell the team whether the resulting experiments were useful.
There is a further limitation: if only selected compounds are tested, outcomes for the rest remain unknown. A monitoring report should show that coverage rather than imply that the entire ranked list has been validated. Evaluating unselected candidates requires its own sampling and experimental plan.
The acceptance question is practical: can a colleague explain why this experiment was selected, using the evidence available at the time? A score contributes to that explanation. The product must leave room for the rest of it.
Sources consulted
scikit-learn — Probability calibration
Guo et al. — On Calibration of Modern Neural Networks (2017)
NIST — Artificial Intelligence Risk Management Framework 1.0 (2023)