ReadonlytypeUnique type identifier for this oracle. Used in test configuration to specify which oracles to run.
Evaluate the test output.
Evaluation input (test, expected/actual outputs, target entity, user)
Oracle-specific configuration
Promise resolving to oracle result
response-well-formed— was the output usable at all?This is the oracle that measures the malformed/discard class directly — the issue's headline number. It asks strictly less than AgentDecisionOracle: a turn can be well-formed and wrong, but never malformed and right. Keeping them apart is what lets a comparison say which of "malformed rate fell" and "decision accuracy rose" actually happened.
Configuration is a WellFormedConfig. Note
requireEnvelopedefaults to off: measurement measured that a correct native tool call carries no text at all on Google, so requiring an envelope unconditionally would fail every correct native turn.