Concurrent planning and execution metareasoning model

⚠️ Curated Limitations & Negative Results

• The paper studies a static snapshot with supplied process and deadline distributions. Learning accurate profiles online, updating them during execution and integrating the abstraction into an operating planner are outside this model assessment; there is no calibration or deployment guarantee.
• The physical setting uses deterministic world evolution, positive known base-action durations for the analysis, serialized non-preemptible actions and overlap only between computation and a base action. Arbitrary parallel robots, uncertain action outcomes, preemption and continuous-time extensions do not inherit the stated results.
• Mutual independence of the M_i and induced D_i variables is an analysis premise, not an empirical property established for real search subtrees. Product-failure calculations and the stated reductions must not be transferred to arbitrary dependencies. Independence of X_i and R_i is not needed merely to define their difference.
• With a standard CDF, the background expression 1−D(t) counts deadline>t, whereas the induced-deadline derivation allows suffix start<=X−R. Discrete equality mass needs an explicit convention. This draft preserves the printed boundary difference without a silent off-by-one repair.
• The arXiv Appendix A earliest-time expression adds remaining action time W and one new computation unit even though those activities can overlap. In the retained hand case T=0, W=1, no unstarted head actions and one computation unit remaining, the printed progress updates finish both at time 1 while that expression gives2. A distant deadline 3 isolates this issue from deadline equality. It is a source-interface discrepancy, not a diagnosed code defect or a proof that all model results fail.
• The conditional termination hazard requires 1−M_i(T_i)>0 for an unfinished eligible process. The inherited log-failure utility and LPF/time heuristics also require explicit zero/one-probability and positive-allocation-time conventions. These displayed formulas do not supply a complete numerically guarded implementation.
• The known-induced-deadline structural results and fixed-initiation reduction have their own premises. The source does not claim a polynomial general solver; constant-bounded heads or equal slack yield pseudo-polynomial special cases in numeric horizons/explicit distributions. The inspected arXiv proof sketches are source evidence, not a formally verified proof or an extension to arbitrary dependencies.
• The official paper, author-hosted main paper and arXiv v1 are distinct artifacts of one work. The author PDF lacks the appendix it references; official Algorithms 1–5 and the arXiv IPAE/Refined-Max-LET displays differ. The named algorithm schemas and their literal helper-interface questions remain separate from this formal-model identity. No code was acquired or executed.
• Reported comparisons use simulated instances based on 15-puzzle search histograms and selected process counts/action durations. Ellipses show instance results rather than confidence intervals, and small appendix runtimes are rounded for logarithmic plotting. These experiments establish neither universal policy superiority, profile calibration, statistical significance nor integration overhead in a continuously updating planner.