Main idea

The decision rule fits in three lines: big gain and low complexity, do it now; medium gain and medium complexity, schedule it; small gain, skip it.

Its fragility is in the second column. The gain is measured on consumption; the effort, though, is estimated from the skill, as if the receiving code were clean and modular. An optimization classed as "low" — switching a synthesis to incremental mode, say — can require tracing what has already been integrated through old code nobody wants to reopen. It is that gap that shifts the ranking, and it justifies marking effort estimates up from the outset.

Why it matters

This stops you reading an audit report as a task list: it proposes an order of magnitude of gains, not a plan.

And it puts the decision back where the information the report lacks sits — with the people who know the codebase.

Nuances and limits

Some optimizations with a low immediate gain change the trajectory: going incremental returns little today and more and more as the data accumulates. A snapshot ranking systematically underrates them.

Open questions

• How do you value an optimization whose gain grows with the volume of data rather than staying constant?