Use Only Observable and Executable Training Targets
Series: Building Auditable LLM Systems · Part 11 of 11 A better result from a stronger planner or teacher is not automatically a valid LLM training target. A result is eligible for training only when it passes two gates: The model can observe the information on which the result depends. The model is authorized and able to execute the corresponding action. Otherwise the result can expose an information gap, an action-space gap, or a hindsight upper bound. It should not be presented as an answer the model ought to have learned. ...