config.example.yaml

stitch_version: "0.1"
role: lead

protocol:
  run_id: "mortality_v1_2026q2"     # unique per experiment; also the top folder name
  num_passes: 5                     # number of federated passes
  policy:
    require: all_expected           # wait for EVERY expected site before aggregating
    pass_timeout_minutes: 120
    heartbeat_minutes: 3
  model:
    format: npz                     # EXTENSION ONLY — STITCH never opens the file
  storage:
    backend: localfs                # develop offline; switch to azure_blob for production
    local_dir: "./stitch_store"     # local mirror (paths identical to the blob)
    blob_dir: "./blob"              # stands in for the Azure container while offline
    # azure_blob also needs:
    # account_url: "https://clifstitch.blob.core.windows.net"
    # container: "fedrun"
  transfer:
    max_retries: 5
    retry_backoff_seconds: 2
    poll_seconds: 2                 # raise to ~30 on real infra
  security:
    whitelist: ["model.*", "*.json"]
    forbid_patterns: ["*.parquet", "*.csv", "*.json.gz"]
    max_file_mb: 512
    verify_sha256: true

project:
  train_fn: "project.train:train"
  aggregate_fn: "project.aggregate:aggregate"
  data_dir: "./data"

sites:
  - id: GMU
    lead: true                      # the lead also trains locally, like any participant
  - id: UCMC
  - id: RUSH

minting:
  auth: account_key
  account_key_env: "CLIF_STITCH_ACCOUNT_KEY"
  sas_expiry_days: 30
  embed_sas: true