About Grejna

Grejna is developing a decision-support layer for organisations that use predictive machine-learning models in real-world workflows. The solution complements the existing model with calibrated information about uncertainty, local explanations and a structured way to determine how each output should be handled.

The starting point is simple: a model can be accurate on average without being equally reliable in every individual case. For an organisation, it is therefore not enough to know the model’s average accuracy or receive a single probability value. It also needs to assess how well supported the individual output is, what influenced it and whether it can be used directly or should be reviewed by a person.

Grejna does not replace the customer’s model and does not make the organisation’s decisions. Instead, Grejna creates a controlled link between model outputs, uncertainty, explanations, decision rules, human handling and subsequent outcomes.

Predictive models are usually evaluated using metrics that describe how the model performs across a larger body of data. These metrics are important, but they do not necessarily indicate how much an organisation should rely on an individual output.

Two cases can, for example, receive the same prediction even though the model has substantially stronger support for one than for the other. The model may have seen many similar cases in its training and calibration data, or it may be operating in a part of the data where the available evidence is limited. A high probability value does not necessarily mean that the probability is well calibrated either. If a model assigns an 80 per cent probability, the corresponding outcome should occur in roughly 80 per cent of comparable cases. Many models do not automatically satisfy this requirement.

This distinction matters when a model output is used to determine what happens next. If all outputs are treated the same, the organisation risks either automating cases that should have been reviewed or spending human time on cases where the model could have been used with sufficient confidence.

When we say that AI merits trust, we do not mean general trust in AI or that an individual output is guaranteed to be correct. We mean that a specific model output, under known conditions and for a defined use, has sufficient support according to the calibrated uncertainty information and the organisation’s established decision policy.

Grejna builds on the research behind Calibrated Explanations, developed by Helena Löfström, Tuwe Löfström, Ulf Johansson and Cecilia Sönströd.

Calibrated Explanations is a model-agnostic method for explaining individual predictions while making uncertainty explicit. The calibration mechanism depends on the type of predictive task.

A conventional local explanation method can show which variables had the greatest influence on a prediction. If the model’s original probabilities are poorly calibrated, however, the explanation can be difficult to use as the basis for a real decision. It is not enough for the explanation to represent the model accurately if the model itself provides a misleading picture of its own certainty.

For classification, the method can use Venn–Abers calibration to provide calibrated probability estimates together with information about uncertainty in those estimates. For regression, conformal methods are used to describe predictions, prediction intervals and probabilistic uncertainty. Calibration draws on a separate dataset with known outcomes.

Across these settings, the common objective is to complement the model output with calibrated uncertainty information and local explanations of what influenced the result. A narrower interval means that the estimate is more precisely bounded, while a wider interval indicates that the available evidence leaves greater room for uncertainty.

The uncertainty information is not a guarantee that an individual prediction is correct. It provides a statistical basis for distinguishing between more and less well-supported outputs.

Calibrated Explanations does not only explain the model’s final output. The method also describes how different features in the individual case influenced the calibrated probability.

Each feature contribution is given a defined meaning in relation to the model’s calibrated output. In simplified terms, the probability for the individual case is compared with the model’s estimate when a particular feature is changed in a systematic way. This makes it possible to see whether that feature increased or decreased the probability and how large its contribution was.

The feature contributions are also accompanied by uncertainty information. This means the user can see both the apparent importance of a feature and how confidently that importance can be estimated. This differs from explanations that present feature importance as exact numbers without showing how sensitive those values are to uncertainty in the model and the available evidence.

The method can provide two types of local explanations:

A local explanation needs to be stable in order to be useful. If the same model, the same data and the same case produce different explanations across repeated runs, it becomes difficult to determine which explanation should form the basis for a decision.

Calibrated Explanations is designed so that the same case produces the same explanation as long as the model and calibration data remain unchanged. The published study evaluated the method’s stability and robustness across 25 datasets for binary classification. The results showed very high stability with an unchanged model and low variation when the model and calibration data were changed in a controlled way.

This does not mean that the explanation remains unchanged when the model or its environment genuinely changes. On the contrary, a correct explanation should reflect relevant changes in the model’s output. Stability here means that variation in the explanation should follow real changes rather than arise randomly between identical runs.

A model and its operating environment are not static. Customer behaviour, processes, data sources and types of cases can change. The model itself can be updated and begin producing different probabilities. Calibration data that were previously representative may become less relevant over time.

It is therefore not enough to calibrate the model once and assume that the uncertainty information will always retain the same meaning.

Grejna’s product direction includes continuous assessment of whether the calibrated signal remains useful for the decision policy chosen by the organisation. The central question is not only whether the data or model have changed, but whether the organisation can still use the uncertainty information to distinguish between direct handling and human review.

If that relationship weakens, the organisation should be able to detect it and respond in a controlled way. This can mean temporarily sending more cases for review, adjusting the decision policy or recalibrating the model. Each change needs to be documented and version-controlled so that it is possible to establish when it was introduced and which decisions it affected.

Human oversight only works when it is connected to clear responsibility and a real operational workflow. Simply allowing a user to open an explanation is not sufficient.

Grejna is intended to help the organisation define when human review is required, what evidence the reviewer should receive and how the human assessment should be documented. This makes it possible to focus limited review capacity on the cases where uncertainty or consequences justify it.

This can both reduce unnecessary manual review and make the review that is carried out more relevant. The objective is not maximum automation, but a balance in which automation, human judgement and risk can be monitored together.

Grejna is intended to work alongside the customer’s existing models, workflows and technical systems. The solution does not replace model development, case management, general operational monitoring or the organisation’s broader governance and compliance processes.

Grejna’s focused role is to connect:

• calibrated uncertainty and local explanations,
• the assessment of whether the signal remains useful,
• the organisation’s version-controlled decision rules,
• automated and human handling,
• decision evidence and traceability,
• subsequent outcomes and monitoring.

This can provide the organisation with stronger evidence for transparency, human oversight and documentation. Grejna does not, however, guarantee that every decision is correct or that an AI system automatically complies with every requirement of the EU AI Act.

Calibrated Explanations forms the research-based core. Grejna is developing the commercial infrastructure required to use the method consistently within an organisation.

This means taking the technology from an individual explanation to a connected decision process: from model outputs and expressed uncertainty to decision policies, human handling, traceability and monitoring against real-world outcomes.

It is this connection between research and practical use that Grejna intends to validate together with its first pilot customers.