Imported from Centaurioun/osteogenesis_imperfecta (AGENTS.md). Install upstream with npx skills add Centaurioun/osteogenesis_imperfecta. Copyright stays with the author.

🤖 Multi-Agent & System Rules (Osteogenesis Imperfecta Analysis)

This workspace is a pediatric disease reporting ecosystem with a small sample size (n=34). Copilot, Claude, and all autonomous assistants must strictly follow the analytical and architectural protocols below.

1. Planning and Execution Architecture

• Multi-Step Planning: Do not rush user requests directly into code. First design a step-by-step analysis strategy (e.g., Discovery -> Hypothesis building -> Coding -> Error handling -> Verification). If needed, use a switch_agent style exploration mode.
• Full Transparency: Any cases (hasta_kodu) excluded for analytical reasons, merged categories, or ignored anomalies must not stay only in code. They must be clearly reported in user-facing Markdown under a "Decision Rationale" section and also written to 01_data/derived/issue_log_post_advisor_round2_v1_2026-04-18.csv. For legacy compatibility, keep an equivalent copy in issue_log_v3.csv when required.
• Authority-Domain Discipline: Manuscript-facing authority is always under 04_manuscript/; 03_outputs/active/ and Manuscript_Data/ are baseline/provenance lanes and cannot be used as manuscript authority.
• Canonical Analysis Entry Point: Primary entry for current reproduction is 02_analysis/notebooks/active/oi_oro_dental_master_FINAL_1_2.ipynb; use 02_analysis/scripts/active/oi_oro_dental_master_FINAL_1_2.py as the script counterpart.
• Clinical Hallucination Ban: Clinical thresholds, phenotype rules, and diagnostic interpretations must be derived only from workspace sources (01_data/reference/, codebook/SAP/metadata docs); do not introduce external clinical rules.

2. Strict Analytical Standards (N=34 Constraints)

• Small-Cell Prohibition (Categorical Data): If expected cell count is < 5 in contingency tables or hypothesis tests, abandon standard Pearson chi-square immediately. Use permutation testing (preferably >10k iterations) or Fisher-Freeman-Halton / exact methods instead.
• Continuous Data Standards: Most distributions must be treated as non-parametric/non-normal. For two-group comparisons use Mann-Whitney U; for three or more groups use Kruskal-Wallis only.
• Occlusion Red Line (occl_tip): Only 1/2/3 are Angle classes. If occl_tip == 4, flag infraokluzyon_var = 1 (analysis-facing alias: infraokluzyon_var_clean) and exclude from Angle analyses as NaN.
• DMFT Interpretation Rule (dmft_dmft): This field is not a decomposed official DMFT index; it is a count variable. If binary conversion is needed, derive caries_any only via dmft_dmft > 0.

3. Model Reliability and Penalty/Correction Rules

• Overfitting Control Principle: Because of low n, prefer penalized Ridge/L2 or Firth logistic models over classical MLE logistic regression to prevent coefficient explosion.
• FDR / Multiple-Comparison Corrections: When listing p-values, apply Holm-Bonferroni (Holm adjustment) by default across all families and include corrected values in tables.
• Effect Size Requirement: Results cannot be reported with p-values only. Include Cramer's V (Cramer's phi) for chi-square/permutation families and $\varepsilon^2$ (epsilon-squared) for Kruskal-Wallis analyses.
• Deterministic Reproducibility: In all stochastic procedures (permutation/bootstrap/CV), fix SEED = 20260228; identical input must produce identical statistical outputs.

4. Multicollinearity (Dependent Data) Alert

• In no multivariate regression/statistical model should yas and dentisyon_donemi_kod be included at the same time. One is effectively derived from the other and destabilizes the model. If you run model-based analyses, the primary chronological covariate must be yas.
• Fail-Fast Requirement: If unexpected data types/values are detected (e.g., value 2 in a binary field, negative age), the analysis must not continue silently; stop with assert, raise the error, and log it to 01_data/derived/issue_log_post_advisor_round2_v1_2026-04-18.csv (and issue_log_v3.csv if needed).