ingest-training-datalake

ingest-training-datalake manages only non-client training corpus acquisition.

It is designed to improve extractor and learn-datalake quality by:

• measuring corpus coverage by sector and file type
• planning targeted downloads for sector gaps
• acquiring additional documents via fetcher
• storing cycle outcomes to memory with taxonomy tags for graph recall

Guardrails

• training-only root enforcement (default allowed root: /mnt/storage12tb/extractor_corpus)
• no direct client datalake ingestion
• no direct client memory writes
• enforced memory scope prefix: datalake_training_*
• loop defaults to planning mode (--no-execute-fetch); downloads require explicit --execute-fetch