When the plan changes, does an adaptive heuristic choose better work?
Exact work selection from a small queue with prerequisites, named budgets and revision history, beside a test of whether an adaptive heuristic helps.
Finding
In an authored changing simulation, the reactive baseline completed 59 of 64 accepted actions and the adaptive heuristic 43 of 64. The negative result is kept.
In practice
For a small queue, exact selection is cheap. Measure an adaptive layer before trusting it.
Why this question matters
Plans change while work is under way: a budget shrinks, a task arrives, a prerequisite slips. An adaptive layer is often assumed to choose better work when that happens.
What I built
An exact selector over a prerequisite graph within named budgets, replanning that keeps completed work, and counterfactuals for what forcing a task in would displace.
A design choice
On a queue this small, I solve the selection exactly on every revision and keep the work already completed, rather than trust a heuristic to approximate it. The adaptive idea was tested separately, in an authored changing simulation against a simple reactive baseline. It completed 43 of 64 accepted actions to the baseline's 59, and that negative result stays in the release.
Where the evidence stops
At most 20 tasks, four named budgets, capacity 0 to 128 and a one-second planning deadline. The simulation is authored, with no live model, human trial or production use. That the adaptive approach works better is not established, and this is not general AI reasoning.