improve-lean-optimums

Verify model training metrics against Lean4 formal proofs by replaying metric certificates to confirm data integrity and adherence to theorem-backed performance gates. This validates that observations satisfy preregistered bands.

Diagnosing model performance when metrics miss preregistered bands involves a hypothesis-driven framework exploring five specific lanes: measurement, training, architecture, model, and assumptions. This ensures evidence-based improvements grounded in formal verification.

Yes, you need a Lean4 theorem prover environment and local Python-based metric verification scripts. These dependencies ensure architectural and mathematical consistency when validating metric certificates for formal research campaigns.

Decide whether to continue, stop, or block model promotion using tiered disposition logic based on formal evidence. This framework evaluates verified metric certificates to determine if performance strictly meets theorem-backed gates.

Formal metric verification fails when evaluation observations deviate from Lean4-calculated ranges because the metric certificates no longer mathematically satisfy the preregistered theorem-backed bands. The skill refines these ranges and diagnoses the underlying cause across five specific lanes.