The project requested parallel, professional completion work on open issues. Issue #22 asks whether constraint-guided open-ended search can outperform simpler baselines when goals are vague or change.
Before claiming the issue, the live tracker was checked for assignees, active claims, open pull requests, and matching remote branches. None were present.
Merged work already covered the initial E011 simulator and 100-seed sweep, correlated verification, independence-aware aggregation, verification backpressure, a panel phase diagram, a failed live-verifier experiment, partial oracle evidence, corrected dependence models, group-independence limits, and a quorum frontier.
The strongest small unresolved confound remained the one recorded in E011: strategy evaluation budgets were unequal. Scalar evolution could retry until it filled an accepted batch, while QD had a separate initialization allowance.
E024 is a separate benchmark over the unchanged E011 landscape. Random, fixed-scalar, and QD search each receive exactly 2,500 proposals and 2,500 panel verification attempts per seed. Initialization is charged to that budget and acceptance retries are disabled.
The experiment adds fixed-horizon post-change utility AUC and regret AUC. A 100-seed synthetic run preserves the original ordering: QD has higher utility AUC than both baselines in 100/100 seed indices.
This is synthetic mechanism evidence only. It does not demonstrate emergence in real software projects or communities. Equal evaluation count is not equal wall-clock, energy, or human-attention cost. QD is still supplied the plausible goal set, including the changed goal.
The budget confound is removed for the original three-strategy comparison, so the QD hypothesis remains worth testing. Issue #22 remains open because mutable Goal Graph learning, the full baseline set, churn/adversaries, stigmergic memory, explicit novelty/information gain, post-integration defects, catastrophic failure, human attention, and real bounded tasks remain unimplemented.
The new CLI, compact result artifact, tests, and explicit budget contract make the comparison independently reproducible. Keeping the limitations adjacent to the result reduces the risk that a contributor mistakes a toy benchmark for a production architecture decision.
Codex implemented and tested this bounded experiment using repository context. The committed deterministic artifact is generated by the documented command; CI checks the implementation and artifact contract. No synthetic score is relabeled as external or real-world evidence.