Reading a result

CLARITY produces the explanation of an evaluation within the same calculation as the recommendation: the explanation is never reconstructed afterwards. The same calculation reads at three levels, for three audiences.

Three explanation levels

The operational explanation

The result screen of an evaluated application shows:

• the recommended outcome (approval, conditional approval, review, decline) and, where relevant, the forced reviews that prevented an automatic approval;
• the score placed on the policy's scale, with its zones;
• the key factors: the variables that contributed most, with the direction of their contribution and the points obtained;
• the exclusions triggered, and those that could not be evaluated;
• missing data, with its reason, and the resulting confidence level (high, medium, low); for an excluded application, which has no score, the confidence is "not applicable";
• conditions that could not be computed, distinguishing a missing policy parameter from missing application data;
• levers, under the heading “What would change the score zone”: for a variable, the closest more favourable band that, on its own, would place the score in a better outcome zone, and the corresponding gain in points.

The points obtained by the application are always visible. The full scale of the scorecard is reserved for the people who design, approve or audit the policy.

Communicable reasons

The communicable level proposes reasons worded for the applicant. It always carries an explicit status: draft to be validated by a person, or insufficient reasons, human drafting required. Restricted data may count in the score, but never appears as a communicated reason.

What is still under construction

• Levers only concern the variables that the characteristics dictionary declares changeable, but this filter is not yet covered by a dedicated test, and that qualification remains to be validated with an institution. They are read as an indication of the score's sensitivity, not as advice.
• Explanations are produced only for applications evaluated one by one; no output is sent back to the institution's systems.
• No committee PDF document is produced; the verifiable material of a decision is exported as an evidence file.