Skip to contents

Calibrated decision outputs

Source: R/decision.R
decision_types.Rd

[Experimental]

Decision outputs ask the model for probability evidence instead of a bare answer, then decode that evidence locally. They mirror the experimental decision types introduced in DSPy 3.4 (Noul, Score, and Choice):

  • decision_bool() for a type_boolean() output. The model reports P(TRUE), and the output is TRUE when that probability reaches threshold.

  • decision_score() for an ordered type_enum() output (a rubric). The model reports a probability for every level. Their probability-weighted mean level index is a continuous score, and cuts map that score to a returned level.

  • decision_choice() for an unordered type_enum() output. The model reports a probability for every option, and the option with the largest probability * weight is returned.

Attach these specifications to a module with with_decisions(). The signature keeps its ordinary output types, so predictions still contain a native logical or character value that existing metrics can compare. The probability evidence is available through decision_evidence().

The numeric settings (threshold, cuts, and weights) are never sent to the model. Changing them re-reads cached evidence without new provider calls, which is what lets ReAnchor() fit them cheaply against a metric.

Usage

decision_bool(threshold = 0.5, criteria = NULL, description = NULL)

decision_score(cuts = NULL, criteria = NULL, description = NULL)

decision_choice(weights = NULL, criteria = NULL, description = NULL)

Arguments

threshold

Probability in [0, 1] at or above which a Boolean decision is TRUE. Defaults to 0.5.

criteria

Optional descriptions of the outcomes, sent to the model. For decision_bool(), a list or character vector with elements named true and/or false. For decision_score(), one description per level, in level order. For decision_choice(), descriptions named by option.

description

The question the model answers for this field. When NULL, the output type's own description is used. One of the two is required, because a decision needs an explicit question.

cuts

Increasing boundaries on the continuous score, which runs from 0 (first level) to N - 1 (last level). There must be N - 1 cuts, each strictly inside that range. NULL (the default) uses the midpoints 0.5, 1.5, ..., which select the level nearest the score.

weights

Named non-negative multipliers for the option probabilities. Omitted options use 1. A zero weight disables an option. NULL (the default) weights every option equally.

Value

A dsprrr_decision_spec object for use with with_decisions().

See also

Other decisions: ReAnchor(), decision_evidence(), decision_settings(), with_decisions()

Examples

decision_bool(threshold = 0.7, criteria = c(true = "Service blocked"))
#> <dsprrr_decision_spec> bool decision
#> threshold: 0.7
decision_score(criteria = c("Cosmetic", "Degraded", "Outage"))
#> <dsprrr_decision_spec> score decision
decision_choice(weights = c(other = 0.5))
#> <dsprrr_decision_spec> choice decision
#> weights: other = 0.5