GoalScope

reading the goal heads

Spec

How to read: each spec asks for a function that computes several sub-tasks in order. At the four pivots before the next sub-task's line (coloured in the code), the goal heads are read out in natural language. The green line is what the model then writes. Click a verdict pill for the judge's reasoning. The number is how many of the 4 read-outs the judge accepts. Causality shows, at the same pivots, what the model writes when the same encoding is injected into an information-less C++ stub (the stub, in grey, ends the prompt).

Read-outs at the pivots

Task

Model's code (prefix → its own continuation)