Turn findings into improvements
Review compares the implementation's numerical changes and input assumptions with the task and its evidence. Findings improve the checks, tools, and reference kernels available to later runs.
Review the assumptions behind the speedup
Check whether numerical rewrites and specializations hold throughout the required input domain. Investigate uncovered assumptions, revise the candidate when needed, and validate the result before reuse.
Expand input coverage
Turn an input that exposes an invalid optimization into a regular correctness check. Strong-decay cases, for example, reveal overflow and underflow in a separated decay calculation.
Repair analysis tools
Isolate an analysis tool defect with a small reproducer, validate its repair, and retain the reproducer as a regression test. Later agents then receive more reliable feedback from the same tool.
Reuse reviewed kernels
Keep the implementation with its supported inputs, validation results, timings, and design notes. Later agents can reuse a working kernel or adapt one of its techniques to another operator.