Changelog • LLMRpanel

LLMRpanel 0.4.0

Added an optional Shiny GUI, launched with run_panel_studio(): build a persona panel from margins, administer a choice item, calibrate against a benchmark, and read the report with its calibration banner. The GUI’s dependencies (shiny, bslib, DT, and the shared LLMR.shiny substrate) are Suggests; non-GUI users install none of them, and the launcher guards on all four. Keys are read from environment variables only; a demo mode runs offline.

API harmonization (pre-CRAN, no deprecation shims). Generic verbs now carry the panel_ stem:

instrument() -> panel_instrument(); administer() -> panel_administer(); calibrate() -> panel_calibrate(); bias_audit() -> panel_bias_audit(); silicon_power() -> panel_power(). The UNCALIBRATED banner and the design-stage stance are unchanged.
Registers the shared generics LLMR::diagnostics(), LLMR::report() (LLMR >= 0.8.4) and tibble::as_tibble() on panel objects.

conjoint_instrument() and amce(): render a conjoint design into one forced-choice item per task, and estimate average marginal component effects by OLS on treatment-coded dummies with CR1 standard errors clustered by persona. conjoint_design() now retains its attribute list so the instrument and the estimator can read it.
silicon_power(): analytic two-arm sample sizes for the planned human study, with dispersion priors taken from the silicon pilot (Likert standard deviation, choice modal share); the priors inherit the panel’s calibration status.
panel_from_data(): draw personas from microdata rows, preserving the joint distribution of attributes, as the counterpart of panel_from_margins().

First public release of calibrated silicon sampling.

panel_from_margins(): persona panels from user-supplied population margins (no data shipped, no “default” populations), with template-rendered persona text.
item_likert() / item_choice() / item_open() and instrument() with per-respondent item- and option-order randomization, recorded per response; vignette_design() and conjoint_design() for factorial stimuli (conjoint profiles within a task are guaranteed distinct, with a warning when the attribute space is too small to allow it).
administer(): persona-conditioned answering through LLMR’s parallel engine; replies matched to offered options, Likert positions scored, failures kept as NA.
calibrate(): coverage-aware comparison to a human benchmark, restricted to items the benchmark covers, with per-item nonresponse recorded, benchmark shares checked to sum to 1, and a three-state banner (UNCALIBRATED, PARTIALLY CALIBRATED, calibrated). Deviations are reported, not adjusted away.
bias_audit(): option-order effects (chi-squared per item) and nonresponse; panel_report() leads with calibration status.
The design vignette runs fully offline (a simulated respondent through the .runner seam), with a gated live section.