Demo seed (manage.py seed_demo)

Seeds one project — Support Copilot at Ledgerline, a fictional payments company — with thirty days of traffic across three capabilities (ticket triage, KB answering, dispute resolution) and every downstream surface filled: an eval matrix and eval sets, live span and trajectory scores with EvalScore history, signal subscriptions and events, sessions, datasets with immutable cell chains, grouped eval runs, optimiser runs (harness, model comparison), training jobs with judge evals and class metrics, deployed models with inference traffic, and the credits ledger.

Run

docker compose exec api python manage.py seed_demo [--owner EMAIL]

~10 min, deterministic (seeded RNG, uuid5 ids for capabilities, jobs and groups, timestamps anchored to NOW) and idempotent: the reset deletes the project by slug (plus the retired Undermind demo slugs), the @ledgerline.dev team users, and the owner's non-free-credit ledger rows before it seeds. The project belongs to --owner (default frey@overmindlab.ai); the account is created with password password when it does not exist. Plaintext API keys print at the end.

Beat-safety invariants — never break these

Celery beat + reconcilers stay running against the seeded DB. Violations cause workers to re-drive seeded rows against real providers and spend credits:

• Every EvalRun / FinetuningJob / OptimizerExperiment / dataset Cell is TERMINAL (reconcilers re-drive non-terminal rows within 10–60s).
• The span sweep re-scores any span whose eval_attributes lacks a score key, and the trajectory sweep any root from the last day with a score but no trajectory. So every span is written with a final eval_attributes (children score: None), every root carries a trajectory block, and every span's received_at is backdated in the same _flush() that inserts it.
• Signal subscriptions are deterministic signals only; plan_drift calls the judge on every run of the signal beat.
• A capability sync fires the rebind hook (identity.enqueue_rebind); the seed patches it to a no-op. The Default eval set upsert is synchronous and calls no model.
• The training monitor syncs each FinetuningJobEval score from its own EvalRun summary, so every job eval links to its own single-target run (bench_runs).
• Connectors keep auto_sync_enabled=False.

Mechanics

• auto_now_add/auto_now columns are backdated via the raw-SQL backdate() helper (executemany), not via the ORM.
• Datasets land through services/datasets/land. Explicit fixture transformations publish measured, provenance-preserving frames directly; no platform agent or script runner exists. Every dataset has a real active frame. Historical demo consumption stamps used_at directly and pins cells, including each training job's eval_cell; it does not fabricate a semantic quality review. Live use reports missing reviews as warnings. The reset deletes jobs, experiments and eval runs before the project.
• An eval run is one target; a comparison is runs sharing a group with one is_baseline. Each run's rows are EvalRunRows with dataset_row EvalScores, and its summary comes from the real innermind.runs.aggregate.
• When building the file in parts, formatter tooling can prune imports that are only used by later parts — restore the import block at the end.

Saved model workflows

seed_model_workflows uses synthetic support inputs, real local exploration/partitioning and the common offline scorer. Its comparison is labelled synthetic, and its saved performance workload explicitly has no measurements. All exploration/run/partition states are terminal; the historical training group is wrapped without dispatch. Reset deletes data explorations, performance runs, training experiments, native evaluations and partition plans before protected cells/jobs. No provider adapter is called.

The training-plan fixture is a static discovery snapshot with no baseline, pilot or selection claim. It schedules no work. Reset removes snapshots newest first, before protected partitions/cells, because a superseding snapshot protects its predecessor.