A third-party agent evaluating the MCP surface reached buyer-facing answers carrying terms like recency sentinels, matched cutoffs and created-date resolution. Each was accurate, and each needed an expert to act on. To the person deciding whether to trust a model, a warning they cannot read is the same as no warning.
Every readout now carries warnings_explained beside the unchanged warnings
list: for each warning, the family, the technical term, what it means for your
numbers and what you would do about it, across 25 warning families. Warnings
the translator does not know are counted in untranslated_warning_count
rather than dropped. Stored feature-stability warnings and true-loss notes,
which the readouts used to omit, are included through the same collector, so
warning counts and approval caveats are complete.
One more disagreement is gone. recommend_model labelled a model’s data
provenance unresolved on every call while list_models resolved it correctly,
because the recommender never received the connection list. Both tools now
read provenance through one resolver and cannot give two answers about the
same model.