Input A
A small, controlled example.
Context Budget
Allocate a fixed window among instructions, evidence, history, and output.
Curated simulation—not a live provider comparison
Controlled input
Input B
A competing configuration.
Check
A hand-derived expected result.
Configuration A
Run the first controlled setting against the same input.
- Signal
- One behavior becomes visible
- Risk
- A single run can hide variance
Configuration B
Change one setting and replay the identical input.
- Signal
- The contrast is attributable
- Risk
- The result still needs repeated trials