"""GEMEO-CWM: Causal World Model via Block Diffusion + Classifier-Free Guidance.

Target SOTA (May 2026):
  - First Block Diffusion + CFG on clinical EHR trajectories
  - Rare-disease + PT-BR (no incumbent competing in this niche)
  - TTE-validated against >=5 Brazilian PCDT natural experiments
  - Full Robins-Hernan sensitivity suite (E-values, negative controls, tipping-point)
  - On-device (Apple Silicon) — LGPD-compliant, no cloud inference

Beats:
  - EHRWorld (arXiv 2602.03569) — no rare disease, no real counterfactual
  - medDreamer (arXiv 2505.19785) — ICU only, no CFG
  - TA-G-Transformer (Helsinki) — no diffusion, no PT-BR rare cohort
  - ICOM (TechRxiv 2601) — no released code
  - PROCOVA (Unlearn.ai) — only AD/ALS/IBD covered

Module layout:
  block_diffusion.py  — model architecture (absorbing-state + block-causal)
  train_cwm.py        — training loop with conditional dropout for CFG
  cfg_sample.py       — classifier-free guided sampling + counterfactual rollouts
  tte_validate.py     — target-trial emulation against PCDT natural experiments
  sensitivity.py      — E-values, negative controls, tipping-point analysis
  data.py             — event-stream loader from DATASUS SIH/APAC/SIM JSONs
"""
from .block_diffusion import BlockDiffusionTransformer, CWMConfig
from .train_cwm import train_cwm
from .cfg_sample import cfg_sample, counterfactual_pair
from .tte_validate import emulate_trial, ate_with_ci

__all__ = [
    "BlockDiffusionTransformer", "CWMConfig",
    "train_cwm",
    "cfg_sample", "counterfactual_pair",
    "emulate_trial", "ate_with_ci",
]