A self-authored companion to Hephaestus. Hephaestus runs exactly 3 cycles and declares v3 "the strongest final version." That's a real, useful design decision — a bound, chosen on purpose. This is what refinement looks like without one.
v1 → critique → v2 → reflection → v3. Three cycles, then a declared endpoint — "the strongest final version." That bound is the useful part: it turns open-ended polishing into a decision you can actually finish and ship.
But nothing about iterative improvement itself contains that bound. A hill-climbing search — try a change, keep it only if it's measurably better, repeat — has no built-in concept of "done." It stops improving locally, or it doesn't. Nothing inside it ever says final. Someone outside has to say that.
A real (if toy-scale) hill-climber, mutating one character at a time, keeping a change only if it's a strict improvement toward the target below. There is no "declare v3" button. That's not an oversight — the point is that no such button exists inside the mechanism. Only you get to decide when to stop.
Hephaestus's "final" isn't a lie — it's an honest, load-bearing design choice: bound the process at 3 cycles so refinement actually ships instead of running forever. The gap it closes is between "converged" and "stopped on purpose." Ties Argus-in-a-Box (a different kind of externally-imposed stop) and 止まれない · Tomarenai (a mechanism with no pause at all, not even a local-optimum one).