Distill Session — Extract What You Can't Yet Explain

The most valuable knowledge in a top-tier model's output is the knowledge it applied without mentioning. This skill makes that invisible reasoning visible, testable, and transferable.

The Core Insight

When a weaker model fails, the gap is almost never "it didn't try hard enough." The gap is that the top-tier model executed steps it doesn't consciously report — constraint checks it ran silently, orderings it chose without deliberation, failure modes it avoided by reflex. These implicit behaviors are the extraction target.

The hard part is not getting the top-tier model to succeed. It's getting it to notice what it did that was non-obvious.