NEWS

nalanda 0.0.2.3

Added plan_survey_simulation() and run_survey_simulation() for generic multi-turn simulated surveys, with row-wise participant profiles, screen-specific schemas, explicit memory policies, per-turn checkpointing, resume support, and long/wide outputs.

Added tidy_forecast_aggregation(), summarize_forecast_disagreement(), and compare_forecast_aggregations() for reusable diagnostics of structured forecast aggregation results.
Added aggregate_model_forecasts() for explicit equal-weight aggregation of completions within prompts, prompts within models, models within families, and families into a consensus.

Added run_structured_responses() as a neutral row-wise structured-prompt primitive; run_text_analysis() remains the text-annotation use-case wrapper.
Added plan_prompt_grid() and run_prompt_grid() for validated multi-model, multi-prompt structured analyses with model-specific repetitions, smoke runs, call previews, provenance, per-completion checkpointing, resumability, and continue-on-error execution.
Restored compatibility with ellmer 0.4.0 by passing structured prompt batches as lists in run_text_analysis() and run_ai_on_chapters_one_turn().
Fixed the nalanda class order on run_text_analysis() results so current dplyr and vctrs operations work without a downstream class workaround.

Added repair_run_ai_on_chapters() to rerun only failed or fully missing simulation units and return a merged result without overwriting the source.
Fixed repair_run_ai_on_chapters() model selection so fully-qualified Portkey names match their normalized model labels in saved results.
repair_run_ai_on_chapters() now retains Azure content-policy failures without retrying their unchanged prompts, while continuing to repair transient failures in the same result.
Added early validation for gpt-5-mini, which only supports temperature = 1, so invalid simulation configurations fail before API calls are made.
Added fail-fast handling for unrecoverable model/integration route mismatches and unreachable gateway errors, so run_ai_on_chapters(on_error = "skip") no longer repeats the same backend or network failure across every simulation unit.
Added split_book_section_by_headings() to split oversized section text files into chapter-level files using known chapter headings.
Added renumber_chapters_across_folders() to renumber chapter files across ordered folders while preserving title slugs.

Added shared multi-model support across the simulation pipeline so simulate_treatment(), run_ai_on_chapters(), run_ai_on_chapters_one_turn(), and run_ai_cumulative_chapters() can accept length(model) >= 1.
Added a model column to raw simulation outputs and updated downstream summaries to preserve model-specific results instead of collapsing across models.
Added summarize_identity_adherence() and summarize_identity_match_rates() to tally whether models adopt the requested identity and to summarize per-model adoption rates.

Renamed simulate_treatment() output columns to use treatment terminology instead of chapter.
Simplified summarize_treatment_results() to summarize treatment-level outputs only.
Updated README and treatment documentation/examples to match the revised treatment workflow API.

Added BugReports field in DESCRIPTION pointing to the GitHub issue tracker.
Fixed README title grammar: "Do books really change lives?".
Replaced placeholder tests with basic offline tests for prompt builders and chapter summary helpers.
Updated summarize_chapter_scores documentation to match current delta/gap output columns.