Iteration Stop Conditions

The bug this exists to prevent

Iterative refinement loops — review, revise, re-review — are written to stop "when it's good enough." That condition is judgement, and judgement is exactly what fails silently in a loop: each pass produces a plausible improvement, so there is never an obvious moment to stop.

Two things then happen. Cost accumulates in small, individually defensible increments until a number arrives that nobody would have approved up front. And an unconverged loop runs forever against a quality bar it will never clear, because nothing was watching iteration count either.

The fix is unglamorous: several independent limits, any of which halts the loop, and each one reporting which fired so the operator learns whether they ran out of money or ran out of progress. Those are different problems with different responses.

The four conditions

• MAX_ITERATIONS — fires when the count reaches the policy cap, and there is a test
• asserting it does not fire one iteration below it. Off-by-one here either wastes a pass or
• cuts one short, and neither is visible in the output.
• BUDGET_EXHAUSTED — fires when cumulative spend reaches the cap. **A null cap means
• uncapped** and never fires, which is deliberate: an unset budget must not be read as zero, or
• every loop halts immediately for a reason the operator never chose.
• STOP_SCORE_REACHED — fires when quality meets the threshold. The success case, and the
• one people implement first.
• An approval gate — a human decision that stops the loop regardless of the others.

Nothing fires when nothing has fired: the default is continue, stated explicitly rather than falling out of an empty branch.

What this does NOT do

No execution, no scoring, no cost measurement, no storage, no clock. It reads counters you supply and returns a verdict. Actually stopping is yours — this makes the decision auditable, not enforced.

It does not detect convergence. "The last three passes changed nothing" is a stronger stop signal than any of these and is not implemented here.

Verified

12 tests: continue-when-nothing-fires, MAX_ITERATIONS at and one below the cap, BUDGET_EXHAUSTED at the cap, a null cap never firing regardless of spend, STOP_SCORE_REACHED at the threshold, and the approval gate.

Not covered: no test of two conditions firing simultaneously — the precedence when budget and iteration cap are both breached is unasserted.