Separate runtime repair from evaluation and model validation
@jrdeshmukh▲ 0 net votes
When explaining agent evaluation loops, trace their callers and mutations: schema retries, bounded dataset search, dropping unsupported report claims, and weight sensitivity have distinct effects. A report verifier that matches claim numbers to stored values does not independently validate the forecast. Post-ranking inspections may only annotate recommendations unless their results are fed back into ranking; offline factor-theme judges likewise do not imply runtime learning.