SCI Degradation Curve

Progressively degrades an evidence table and records how the Submission Confidence Index responds. The curve is the operational form of construct validity: a well-behaved index must fall (never rise) as passing checks are turned into failures.

Usage

sci_degradation_curve(evidence, n_steps = 10L, config = sci_config_default())

Arguments

evidence

A validated evidence data.frame (from r4subcore).

n_steps

Integer number of degradation steps after the baseline. Must be at least 1.

config

An sci_config from sci_config_default().

Value

A tibble with one row per step and columns:

• step: integer step index, starting at 0

• n_degraded: number of passing checks turned into failures

• frac_degraded: n_degraded divided by the baseline passing count

• SCI: the Submission Confidence Index at that step

• band: the decision band at that step

• delta: change in SCI from the previous step (NA at step 0)

Details

Starting from the supplied evidence, the function repeatedly converts a growing number of currently-passing checks to "fail" and recomputes the SCI at each step. Passing checks are degraded in order of how much they contribute to the baseline SCI, largest contribution first, so the curve shows the steepest descent the evidence allows and is reproducible. Checks that already warn or fail are left untouched.

A passing check contributes 1 - severity_weight to its indicator score, so ordering by contribution degrades the lower-severity passing checks earliest. This is a property of the scoring model, not of this function.

Step 0 is the unmodified evidence. The final step degrades every passing check. When the number of passing checks is smaller than n_steps, the number of steps is reduced to the number of passing checks so that each step degrades at least one additional check.

See also

sci_monotone_check() to confirm a curve never rises.

Examples

if (FALSE) { # \dontrun{
curve <- sci_degradation_curve(evidence, n_steps = 8)
curve
} # }