Knowing in Advance When an Evolutionary Outer Loop Will Not Help: A Pre-Registered Cheap-Baseline Screening Rule

A pre-registered screening rule can determine, before any implementation, whether an evolutionary outer loop over neural-network parameters is worth building, according to a paper posted to arXiv on 28 June 2026 [1]. The rule computes a single metric called recovery and prescribes skipping the outer loop when that metric reaches or exceeds 90% [1]. The rule, introduced by Ramchand Kumaresan, operates at what the paper calls a Phase-0 gate [1]. It calculates recovery R as s/G, where s is the best single-shot gradient or curvature statistic's gain and G is the best gain of any cheap method evaluated [1]. When R is at or above 90%, the rule prescribes abandoning the outer loop [1]. The paper notes that evolutionary outer loops typically cost 10^2 to 10^3 times their gradient inner loop, and whether they outperform a cheap single-shot alternative is usually discovered only after the expense is paid [1][2]. The rule was validated on a within-lab series of pre-registered outer-loop bets comprising two analyzed cases and a disclosed file drawer [1][3]. In both analyzed cases, a static or single-shot computation captured the effect on the project's own metric, the gate fired with R approximately 1.0 in both cases, and the outer loop was abandoned [1][3]. Under a stricter metric on one case, R was approximately 0.95 [1][3]. A companion factorial decomposition in one case localized the apparent win to a static substrate change, with the evolutionary lifecycle contributing no detectable gain [1][3]. On one project, the gate cost about 50 to 70 GPU-hours and screened out an estimated 400-plus GPU-hours for the first cell alone, plus weeks of implementation work, yielding a roughly 6- to 8-fold saving in time and resources [1][2]. The paper states that the rule is prospectively falsifiable: a task with R below 90% where the outer loop still fails to beat single-shot would refute it [1][3]. The only case in the corpus where an outer loop beat single-shot had R approximately 0.004, leaving a wide empty band between roughly 0.004 and 0.95 that supports any cutoff in that range [3]. The authors adopt 90% as a deliberately conservative convention, preferring to wrongly build the outer loop over wrongly skipping it, and note that no corpus point falls in the ambiguous band [3]. This work arrives amid broader scrutiny of whether complex evolutionary pipelines reliably outperform simpler baselines. A separate study found that in code evolution tasks, simple baselines matched or exceeded purpose-built pipelines, and that domain knowledge and search-space formulation mattered more than the evolution pipeline itself [4]. Another line of research has explored early stopping methods to reduce wasted computation in evolutionary search, showing that generic early stopping can cut optimization time in direct policy search tasks without requiring problem-specific knowledge [6].