core-identity 5 tables
Who's who and the guardrails — people & roles, reusable cohorts, the append-only audit trail, and email templates.
profile · cohort · cohort_member · audit_log · email_template
core-experiment 8 tables
The study and its runs — the reusable experiment frame, each launched run as one cell of the comparison grid, its scorecard, who's enrolled, the experiment's context files, and the operator↔enrollee message channel.
experiment · experiment_run · nudge · run_enrollment · experiment_context_file · run_result · message · message_recipient
llm-chat 5 tables
How answers are produced and scored — the model catalog, reusable model configurations, and each scored chat round with its candidate answers. One trusted judge (configured on the experiment) does all scoring.
model_catalog · model_config · model_config_model · chat_round · chat_round_candidate
search + Ask engine 5 tables
Semantic search and the natural-language Ask engine — documents, their embeddings and the embedding-job queue, the registry of what's embedded, and saved NL questions.
document · embedding · embedding_job · embedding_ingest_registry · nl_query