Data pipeline & validation

• Automated pre‑filters catch license issues, malformed formats, anomalies, and duplicates before full training.

• Committees are drawn randomly (e.g., VRF) from stakers to resist Sybils; attackers need a majority of stake to sway outcomes.

• Outcomes: accepted → added to training pool; rejected/banned → excluded from future rounds (with audit trail).

Challenge window (future capability) — A fraud‑proof style challenge can reject a bad update and slash offenders; the protocol primarily relies on validator consensus and incentives, with the challenge as a safety hatch.