Changes in version 1.0.0 (2026-01-17)                  

First major release.
This version consolidates the API, improves robustness for FactoMineR
outputs, and harmonizes audience-tailored prompts across all functions.

Major Changes & New Features

Unified Architecture (“Hub & Spoke”)

All trainer functions now rely on a centralized core (trainer_core.R)
for consistent prompt building, audience profiling, output capture, and
regex/quoting helpers. Any update to the core instantly benefits all
trainers.

New Trainers

  - trainer_t_test() — Interpret Student’s t-tests (one-sample,
    two-sample, paired, Welch) with audience-aware guidance on p-values
    vs confidence intervals.
  - trainer_cor_test() — Interpret correlation tests (Pearson, Spearman,
    Kendall) with clear distinction between statistical significance and
    practical magnitude.
  - trainer_var_test() — Interpret F-tests for equality of variances.

Robust ANOVA & Linear Model Support

  - trainer_AovSum() and trainer_LinearModel() use centralized
    heuristics (trainer_core_extract_tables_heuristic) to reliably
    extract F-test and T-test tables even when standard headers are
    missing (common with capture.output).
  - Strict, space-safe T-table filtering: factor names with spaces or
    special characters are handled via core quoting (trainer_core_quote)
    and detection helpers.

Gemini Integration

  - gemini_generate() enables direct interaction with Google’s Gemini
    API.
  - Includes automatic retry with exponential backoff and clear error
    surfacing (finish reasons, safety blocks).
  - Adds compile_to = "html" and compile_to = "docx" to directly render
    LLM responses into reports.

Improvements

Pedagogy & Interpretability

  - PCA/MCA: Prompts explicitly guide interpretation of correlation
    signs (PCA) and category positions (MCA) to prevent hallucinations.
  - ANOVA: “How-to-read” blocks clarify deviation coding (sum-to-zero);
    the Intercept is the Grand Mean.
  - t-test: Clear separation between Inference (p-value) and Magnitude
    (effect size/mean difference) in prompts.

Developer Experience

  - Internal helpers for regex, capture, quoting centralized for DRY
    code.
  - Improved error messages when input objects are NULL or of the wrong
    class.
  - Audience sections standardized (BEGINNER/APPLIED/ADVANCED) with
    consistent limits (e.g., Takeaway ≤ 5 bullets).

Bug Fixes

  - Fixed duplication of extract_section/header-loss issues by using the
    shared core heuristic and single re-attachment of T-table headers
    after filtering.
  - Fixed potential regex failures with factor names containing spaces
    in trainer_AovSum.
  - Removed unnecessary stringr dependency in examples.

                        Changes in version 0.9.0                        

Major pre-CRAN update with bug fixes, improved prompts, and Gemini API
support.

Highlights

  - New: gemini_generate() — minimal, robust wrapper for Google Gemini
    (Generative Language API) to programmatically generate LLM responses
    from R.
      - Normalizes model ids (accepts "gemini-2.5-flash" or
        "models/gemini-2.5-flash").
      - Correct endpoint (v1beta/models/{model}:generateContent) and
        query key.
      - Config params: temperature, top_p, top_k, max_output_tokens,
        stop_sequences, system_instruction, seed, timeout, verbose.
      - Adds polite User-Agent header by default: EnTraineR/0.9.0
        (https://github.com/Sebastien-Le/EnTraineR).
      - Safer parsing: handles empty/blocked candidates and reports
        finishReason or safety blocks with clear errors.

Bug fixes & robustness

  - Fixed occasional header-loss in filtered ANOVA T-test tables; always
    re-attaches the column header exactly once when filtering.
  - Resolved errors in trainer_LinearModel() when global-fit lines were
    partially printed (missing df or p). Now defensive against missing
    fields.
  - Stabilized audience sections across trainers
    (BEGINNER/APPLIED/ADVANCED) with consistent wording and limits
    (e.g., “Takeaway ≤ 5 bullets”).
  - Cleaned up duplicate attachment issues in verbatim blocks.
  - Improved error messages when external APIs return 404 or
    safety-filtered results.

Documentation & examples

  - Expanded examples for:
      - trainer_LinearModel() (ham and deforestation case studies).
      - trainer_AovSum() (sensory chocolates and poussin datasets).
  - Added dataset docs for deforestation, ham, poussin with guarded
    examples for optional dependencies.
  - DESCRIPTION improved (clear package blurb) and added Imports: httr2
    to satisfy namespace checks for gemini_generate().

Internal / developer notes

  - Consistent prompt skeleton (context → setup/how-to-read → verbatim →
    output requirements).
  - Audience profiles aligned between ANOVA and LinearModel (hierarchy
    rules, deviation coding, partial vs global tests).
  - Safer environment handling for keys: looks up GEMINI_API_KEY by
    default.

                        Changes in version 0.1.0                        

First CRAN release.

New features

  - Audience-aware prompt builders for common analyses:
      - trainer_AovSum() for ANOVA tables (FactoMineR::AovSum).
      - trainer_LinearModel() for multiple linear regression
        (FactoMineR::LinearModel) with global fit summary + per-term F/T
        sections and optional AIC/BIC selection.
      - trainer_t_test() for stats::t.test.
      - trainer_var_test() for stats::var.test.
      - trainer_prop_test() for stats::prop.test.
      - trainer_cor_test() for stats::cor.test
        (Pearson/Spearman/Kendall).
      - trainer_chisq_test() for stats::chisq.test (GOF and contingency
        tables).
  - All trainers support three audiences: "beginner", "applied",
    "advanced".
      - Consistent language, alpha handling, and no invented numbers.
      - Optional summary_only = TRUE for 3-bullet executive summaries.

Data

  - Add three documented datasets for teaching and examples:
      - deforestation: water/air temperatures before vs after riparian
        clearing.
      - ham: sensory descriptors and overall liking (multiple
        regression).
      - poussin: chick weight by brooding temperature and sex (ANOVA).

Quality and robustness

  - Clear separation of orientation/setup, verbatim output, and output
    requirements.
  - ANOVA T-test section:
      - Correctly re-attaches column header after filtering; clear
        “(filtered)” title and scope message.
  - LinearModel:
      - Includes verbatim model-fit summary (RSE, R², F, p, AIC/BIC).
      - Explains partial (added last) F/T tests and interaction
        hierarchy.
      - Selection notes (AIC/BIC): kept/dropped RHS terms and deltas
        (AIC/BIC/RSE/df).
  - Proportion and correlation tests:
      - Audience-specific guidance, CI wording, and small-sample
        cautions.
  - Chi-squared test:
      - Branches for GOF vs contingency; notes on Yates correction and
        Monte Carlo p-values.

Documentation

  - Roxygen2 docs for all trainers and datasets.
  - Examples guard optional dependencies with
    requireNamespace("FactoMineR", quietly = TRUE).

Internals

  - Shared prompt skeleton via trainer core helpers (audience profile,
    headers, formatting).
  - Consistent phrasing across trainers; improved error checks on
    inputs.