Bundle schema reference

The bundle schema version is stamped in manifest.json as bundle_schema_version. The current version is 5.

manifest.json schema

{
  "bundle_schema_version": 5,
  "package_version": "1.0.0",
  "recipe_id": "b2b_saas_procurement_v1",
  "seed": 42,
  "generation_timestamp": "2026-05-27T00:00:00Z",
  "exposure_mode": "student_public",
  "difficulty_profile": "intermediate",
  "table_inventory": {
    "accounts": 1500,
    "contacts": 4200,
    "leads": 5000,
    "touches": 38421,
    "sessions": 19847,
    "sales_activities": 12034,
    "opportunities": 1421
  },
  "file_hashes": {
    "tables/leads.parquet": "sha256:abc123..."
  }
}

task_manifest.json schema

{
  "task_id": "converted_within_90_days",
  "label_column": "converted_within_90_days",
  "label_window_days": 90,
  "primary_table": "leads",
  "split": { "train": 0.7, "valid": 0.15, "test": 0.15 },
  "description": "..."
}

Entity ID format

All entity IDs are deterministic, zero-padded strings:

Entity	Format	Example
Account	acct_NNNNNN	acct_000001
Contact	cont_NNNNNN	cont_000042
Lead	lead_NNNNNN	lead_002501
Touch	touch_NNNNNN	touch_019844
Session	sess_NNNNNN	sess_005002
Sales activity	sact_NNNNNN	sact_007311
Opportunity	oppt_NNNNNN	oppt_000893

IDs are stable: the same (recipe, seed, entity index) always produces the same ID.

Parquet conventions

• All tables use snappy compression.
• Timestamps are stored as datetime64[us, UTC].
• Nullable integers use Int64 (pandas nullable dtype), not float64.
• Boolean columns use bool, not int.