Calibration as a first-class field.
Every verdict object now returns axes_confidence with per-axis confidence scores and named uncertainty drivers. The driver list is human-readable — "single-LLM derivation (no cross-model corroboration)" — so a downstream caller can show users why the score is what it is, not just what the score is.
Also new: verdict_volatility.would_change_at_score and verdict_volatility.would_change_to. If a verdict is one point from flipping, the response now tells you so. This means an agent can ask "is this borderline?" without running a second call.
Non-breaking. Existing clients ignore the new fields silently.