Calibrate a public structural model to confidential data

The only stage that reads source data. Each subject is reduced to bounded multiplicative corrections against the structural model's own prediction, clipped to public prior ranges, and released through a validated differential-privacy backend.

Usage

.fit_calibrated(
  data,
  roles,
  model,
  design,
  priors,
  epsilon,
  covariates = NULL,
  backend = "opendp",
  public_source = FALSE
)

Arguments

data

Confidential PMX event data.

roles

Column roles from pmx_roles().

model

A public pmx_structural_model().

design

A public pmx_trial_design().

priors

Public pmx_priors() for each released correction.

epsilon

Requested subject-level privacy budget.

covariates

Optional public pmx_covariates(). Each declared covariate is released privately and adds one to the released dimension.

backend

"opendp", or "public" for an explicitly public fixture.

public_source

Logical assertion that the input is already public.

Value

A pmx_calibrated_model, carrying corrected typical parameters, accounting, provenance, and a release ledger. It contains no raw records.