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

# Knowing in Advance When an Evolutionary Outer Loop Will Not Help: A Pre-Registered Cheap-Baseline Screening Rule
Source: [https://arxiv.org/html/2606.29119](https://arxiv.org/html/2606.29119)
###### Abstract

We introduce a pre\-registered screening rule that decides,*before any implementation*, whether an evolutionary / population / lifecycle outer loop over neural\-network parameters or structure is worth building\. Such outer loops cost10210^\{2\}–103×10^\{3\}\\timestheir gradient inner loop, yet whether they beat a cheap single\-shot alternative is usually discovered only after the expense is paid\. Our rule computes, at a Phase\-0 gate, a single number—the recoveryR=s/GR=s/G, the best single\-shot gradient/curvature statistic’s gainssdivided by the best gainGGof any cheap method evaluated—and prescribes skipping the outer loop whenR≥90%R\\geq 90\\%\. We validate the rule on a within\-lab series of pre\-registered outer\-loop bets \(two analyzed cases plus a disclosed file drawer\): in both analyzed cases a static or single\-shot computation captured the effect on the project’s own metric, the gate fired \(R≈1\.0R\\approx 1\.0in both cases;≈0\.95\{\\approx\}\\,0\.95under a stricter metric on one\), and the outer loop was abandoned—including one case where a companion factorial decomposition localizes the apparent win to a static substrate change with the evolutionary lifecycle contributing no detectable gain\. On one project the gate cost∼50\{\\sim\}50–7070GPU\-hours and screened out an estimated∼400\+\{\\sim\}400\{\+\}GPU\-hours \(first cell only\) plus weeks of implementation—a∼6\{\\sim\}6–8×8\\timessaving\. The rule is prospectively falsifiable: a task withR<90%R<90\\%where the outer loop still fails to beat single\-shot would refute it\.

## 1Introduction

Evolutionary, population\-based, and lifecycle*outer loops*that search over network parameters, masks, routes, or structure are an attractive idea: they can optimize objectives a gradient inner loop cannot directly serve\. They are also expensive—commonly10210^\{2\}–103×10^\{3\}\\timesthe cost of the inner loop they wrap\. The practical question a lab faces is therefore not “*can*an outer loop help?” but “how do you know, cheaply and in advance, whether it will?”

This paper answers that question with a pre\-registered discipline and reports the outcome of applying it to our own work\. The contribution ismethodological, not a new algorithm; the individual negative results the discipline produced are, on their own, unsurprising \(see Section[2](https://arxiv.org/html/2606.29119#S2)\)\. What is new is the packaged, evidenced, prospectively\-falsifiable*decision procedure*\.

#### Contributions\.

- •A Cheap\-Baseline\-Falsification gate \(Section[3](https://arxiv.org/html/2606.29119#S3)\):a pre\-registered Phase\-0 protocol that runs the cheapest plausible non\-mechanism explanations \(random, single\-shot heuristic, static/non\-evolutionary\) on the*same metric the paper will use*, before the outer loop is built\.
- •An operational screening rule \(Section[3](https://arxiv.org/html/2606.29119#S3)\):skip the outer loop when recoveryR=s/G≥90%R=s/G\\geq 90\\%, wheressis the best single\-shot gradient/curvature statistic’s gain andGGis the best gain of any cheap method evaluated\. The90%90\\%threshold is a conservative convention placed below the observed kills, and the rule is prospectively falsifiable\.
- •Validation on a within\-lab series \(Sections[4](https://arxiv.org/html/2606.29119#S4)–[5](https://arxiv.org/html/2606.29119#S5)\):two analyzed cases, alongside a disclosed file drawer\. In both analyzed cases the gate screened out the outer loop, including a case whose submitted companion factorial isolates the outer loop as the non\-contributor\. \(We do not claim independence—Section[8](https://arxiv.org/html/2606.29119#S8)\.\)
- •A GPU\-hours\-saved ledger \(Section[6](https://arxiv.org/html/2606.29119#S6)\)quantifying what the gate buys\.

We are explicit about what this paper is*not*: it is not a claim that evolutionary outer loops never help\. Section[7](https://arxiv.org/html/2606.29119#S7)states the converse—the condition under which they*should*help—as a hypothesis with a named, pre\-registered successor design whose first screen is now reported; no claim here depends on its unrun branches\.

## 2Related Work and Positioning

We position against three lines of work; none is contested by this paper, which is why our individual negatives are unsurprising and our contribution is the screening procedure rather than any one result\.

#### Evolution strategies on smooth objectives\.

ES is a competitive black\-box optimizer when gradients are unavailable or the landscape is rugged\(Salimanset al\.,[2017](https://arxiv.org/html/2606.29119#bib.bib1); Suchet al\.,[2017](https://arxiv.org/html/2606.29119#bib.bib2); Lehmanet al\.,[2018](https://arxiv.org/html/2606.29119#bib.bib3)\)\. On smooth, differentiable objectives the inner optimizer already serves, ES is known to be at best competitive and often worse:Zhanget al\.\([2017](https://arxiv.org/html/2606.29119#bib.bib13)\)show ES estimates the gradient of a Gaussian\-smoothed objective, so when the per\-minibatch gradient is already accessible its overhead is unwarranted, andLencet al\.\([2019](https://arxiv.org/html/2606.29119#bib.bib12)\)find ES competes with gradient\-based search only on small problems\. Our cases sit in this regime, and the screening rule operationalizes*detecting*it in advance\. This is live terrain: ES is being actively re\-proposed for LLM fine\-tuning\(Sarkaret al\.,[2025](https://arxiv.org/html/2606.29119#bib.bib22)\)and its failure modes actively studied\(Abdiet al\.,[2026](https://arxiv.org/html/2606.29119#bib.bib23); Hoyet al\.,[2026](https://arxiv.org/html/2606.29119#bib.bib24)\)—which is precisely why a pre\-build gate is useful, since it tells you, before committing to an ES outer loop, whether a cheap baseline already recovers the frontier\.

#### Importance\-based mask selection / pruning\.

Magnitude pruning\(Hanet al\.,[2015](https://arxiv.org/html/2606.29119#bib.bib4)\), iterative magnitude pruning with rewind\(Frankle and Carbin,[2019](https://arxiv.org/html/2606.29119#bib.bib5); Frankleet al\.,[2020](https://arxiv.org/html/2606.29119#bib.bib6)\), connection sensitivity\(Leeet al\.,[2019](https://arxiv.org/html/2606.29119#bib.bib7)\), and Fisher\-information importance\(Kirkpatricket al\.,[2017](https://arxiv.org/html/2606.29119#bib.bib8); Sunget al\.,[2021](https://arxiv.org/html/2606.29119#bib.bib9)\)produce strong masks from single\-shot or cheap iterative statistics\.Suet al\.\([2020](https://arxiv.org/html/2606.29119#bib.bib11)\)go further: randomly reshuffling*which*weights are kept within each layer leaves final accuracy unchanged, so only the per\-layer sparsity ratios carry information and the mask identity does not—the direct analog of our random\-κ\\kappacontrol\. That these rival searched masks is established; our mask\-identity result \(Section[5\.2](https://arxiv.org/html/2606.29119#S5.SS2)\) is a confirmation in a bilevel\-constrained\-fitness setting, not a new claim\.

#### Parameter\-space constraints for ES on LLMs \(concurrent\)\.

Schweighoferet al\.\([2026](https://arxiv.org/html/2606.29119#bib.bib10)\)study parameter\-space constraints \(magnitude anchoring\) for ES on LLM fine\-tuning—concurrent work in the same family, orthogonal to mask\-identity selection\. Our mask\-identity baseline includes thegeneric anchored\-weight\-decay mechanism—a plainL2L\_\{2\}pull of the adapter toward its warm checkpoint \(γ​∑‖W−Wwarm‖2\\gamma\\sum\\\|W\-W\_\{\\text\{warm\}\}\\\|^\{2\}\), which we ran as a baseline, not a reimplementation of their specific formulation; we used their reported range only as a starting estimate for the single strength hyperparameterγ\\gamma\. This generic anchored weight decay ties the best single\-shot mask method within noise—a positioning point, not a contest\.

#### “Is the search worth it?” decision procedures\.

The closest formal relatives decide*whether*an expensive search is worthwhile rather than how to run it\.Weertset al\.\([2020](https://arxiv.org/html/2606.29119#bib.bib19)\)formalize a non\-inferiority test and a “tuning risk”—the performance lost by leaving a hyperparameter at its default instead of tuning it—and find defaults are often non\-inferior;Bahmaniet al\.\([2021](https://arxiv.org/html/2606.29119#bib.bib20)\)likewise recommend which hyperparameters are worth tuning\. In neural architecture search, random search is a competitive baseline\(Li and Talwalkar,[2019](https://arxiv.org/html/2606.29119#bib.bib16)\), the search phase often fails to beat random selection\(Yuet al\.,[2019](https://arxiv.org/html/2606.29119#bib.bib17)\), and relative\-improvement\-over\-random has been proposed as the only honest NAS metric\(Yanget al\.,[2019](https://arxiv.org/html/2606.29119#bib.bib18)\)\. The decisive difference is*when*the cheap baseline is consulted: in all of this work it is a*post\-hoc*comparator used to ground or debunk a method after it is built\. Ours is a*pre\-registered, pre\-build*go/no\-go gate, with a committed recovery\-ratio threshold, that decides whether to construct the outer loop at all\. The closest pre\-build intuition we found—running random search first to gauge a search space’s potential\(Krishnaet al\.,[2021](https://arxiv.org/html/2606.29119#bib.bib21)\)—is stated as informal advice, with no recovery ratio, threshold, or pre\-registration\. We do not claim the underlying intuition is new; we claim its packaging as a falsifiable, pre\-committed decision rule is\.

The screening rule itself—a pre\-registered, metric\-matched, pre\-build decision procedure with a falsifiable threshold—is, to our knowledge, not packaged in this literature\.

## 3The Screening Gate and Rule

#### Gate\.

Before any outer\-loop implementation, enumerate the cheapest plausible non\-mechanism explanations of the expected effect \(random, single\-shot heuristic, static/non\-evolutionary\) and run them on the*same metric the paper will use*\. The outer loop is built only if those baselines fail to capture the effect\. Figure[1](https://arxiv.org/html/2606.29119#S3.F1)summarizes the procedure\.

Warm checkpoint\(no further adaptation\)Run the cheap baselines on the paper’s own metric:single\-shot grad/curvature statistic \(gainss\); random; static / non\-evolutionary; unmasked full adaptation\.FrontierG=maxG=\\maxgain over all of them\.Compute recoveryR=s/G∈\[0,1\]R=s/G\\in\[0,1\]R≥90%R\\geq 90\\%?Skipthe outer loopsingle\-shot already at the frontierBuildthe outer looproom a search could exploityesnoFigure 1:The Phase\-0 Cheap\-Baseline\-Falsification gate\. All gains are measured on the paper’s own metric over the warm \(no\-further\-adaptation\) baseline\. The recoveryR=s/GR=s/Gcompares the best single\-shot gradient/curvature statistic’s gainssto the achievable frontierGG\(the best gain reached by*any*cheap method, single\-shot included\)\. DefiningGGas that maximum keepsRRa proper fraction and makes the kill signal “single\-shot*is*the best cheap method\.”
#### Screening rule\.

All gains are measured on the paper’s metric, over the warm \(no\-further\-adaptation\) baseline\. Let the*achievable frontier*GGbe the best gain reached by any method evaluated at the Phase\-0 gate \(single\-shot statistics, random and static baselines, unmasked full adaptation\), andssthe gain of the best single\-shot gradient/curvature statistic\. DefineR=s/G∈\[0,1\]R=s/G\\in\[0,1\]\.IfR≥90%R\\geq 90\\%, skip the outer loop—the single\-shot statistic already sits at \(or defines\) the frontier\. DefiningGGas the max over all evaluated methods keepsRRa proper fraction and makes the kill signal “single\-shot*is*the best cheap method\.”

#### Threshold anchoring\.

Both corpus kills sit atR≈1\.0R\\approx 1\.0\(the mask\-identity study, Section[5\.2](https://arxiv.org/html/2606.29119#S5.SS2), also reads≈0\.95\{\\approx\}\\,0\.95under a stricterΔ​AccB\\Delta\\text\{Acc\}\_\{B\}view—two views of*one*case, not a spread across cases\); the only case where an outer loop beat single\-shot hadR≈0\.004R\\approx 0\.004\. The data therefore supports*any*cutoff in the wide empty band between≈0\.004\{\\approx\}\\,0\.004and≈0\.95\{\\approx\}\\,0\.95; it doesnotlocate90%90\\%precisely \(Figure[2](https://arxiv.org/html/2606.29119#S3.F2)\)\. We adopt90%90\\%as a deliberately conservative convention—when in doubt, prefer wrongly*building*the outer loop over wrongly*skipping*it—and note that no corpus point falls in the ambiguous band, so the convention affects no decision reported here\. The rule isprospectively falsifiable: a task withR<90%R<90\\%where the outer loop still fails to beat single\-shot would refute it\. Table[1](https://arxiv.org/html/2606.29119#S3.T1)collects the full corpus the rule is validated against\.

![Refer to caption](https://arxiv.org/html/2606.29119v1/x1.png)Figure 2:The recovery axis and the conservative90%90\\%convention\. The only corpus case where an outer loop beat single\-shot search sits atR≈0\.004R\\approx 0\.004\(a synthetic, oracle\-aligned positive control; Section[7](https://arxiv.org/html/2606.29119#S7)\); both real kills sit atR≈1\.0R\\approx 1\.0\(the mask\-identity study also reads≈0\.95\{\\approx\}\\,0\.95under a stricterΔ​AccB\\Delta\\text\{Acc\}\_\{B\}view—a second view of the same kill\)\. No point falls in the empty band\(0\.004,0\.95\)\(0\.004,0\.95\), so the data fixes a cutoff only somewhere inside it;90%90\\%is chosen conservatively, not measured\.Table 1:The within\-lab corpus at a glance\. Every real\-task case screensR≥90%R\\geq 90\\%and the outer loop is skipped or never built; the onlyR<90%R<90\\%instance is an oracle\-aligned synthetic control\.RRis the recoverys/Gs/G\(Section[3](https://arxiv.org/html/2606.29119#S3)\) unless noted\.†A direct cheap\-control\-beats\-outer\-loop kill \(control0\.906\>0\.906\>rescue0\.8050\.805\), not ans/Gs/Gratio\.‡RscreenR\_\{\\text\{screen\}\}, the domain\-appropriate result\-ratio for a no\-gradient discrete objective \(Section[7](https://arxiv.org/html/2606.29119#S7)\)\.§The loneR<90%R<90\\%instance where the outer loop wins—but oracle\-aligned, not a real task\.

## 4Case 1 — A Separately Reported MoE\-Lifecycle Project: LOO\-Aligned Fitness \(Cited\)

This case\(Kumaresan,[2026](https://arxiv.org/html/2606.29119#bib.bib26)\)is a separate companion preprint, publicly available on arXiv; we cite its reported results and do not reopen them\. Pre\-registered claim: a leave\-one\-out\-aligned fitness signal in an evolutionary MoE lifecycle improves outer\-loop selection over a static mixture\-of\-experts\.

On real text, every clean active\-turnover comparison was a loss or a within\-noise tie against the static\-MoE baseline\. The one comparison reaching significance \(code domain,\+0\.0625\+0\.0625nats,p=0\.009p=0\.009\) is decomposed by that project’s own factorial: the gainlocalizes to a static router\-substrate change\(\+0\.0976\+0\.0976nats,p≤0\.002p\\leq 0\.002\), while the evolutionary lifecycle term contributesno detectable gain\(−0\.028\-0\.028nats,p≈0\.05p\\approx 0\.05atn=3n=3\)\. The eval used deterministic batching, which narrows the variance estimate and therefore inflates the apparent significance of every term; the true uncertainty is wider than reported, making “no detectable gain” theconservativereading \(and the router\-rewrite significance an upper bound\)\. We make no “the lifecycle harmed performance” claim—the term is indistinguishable from zero\. In screening terms: a static \(single\-shot\-class\) computation was the frontier;R≈1\.0R\\approx 1\.0; the gate fires\.

## 5Case 2 — The Mask\-Rescue Arc: Heritability→\\rightarrowκ\\kappa\-Sparsity→\\rightarrowMask Identity \(This Paper\)

A single project pre\-registered two successive outer\-loop bets; both were falsified by a cheap baseline on the project’s own metric\.

### 5\.1Heritability does not rescue \(theκ\\kappa\-sparsity result\)

Pre\-registered claim:*heritable*sparse mutability masks, evolved under selection, rescue gradient\-warmed LoRA adapters from perturbation better than non\-heritable controls\. The decisive comparison is against alifecycle\-free random\-κ\\kappacontrol\(fresh randomκ\\kappa\-sparse masks each epoch; no inheritance, no selection\) and a dense matched\-uniform baseline\.

#### Provenance of the destructive regime\.

Theσ=1​e\-​1\\sigma=1\\text\{e\-\}1scale used here was*not*validated as a reproduction of the prior warm\-start failure boundary: a pre\-registered matched\-σ\\sigmareplication \(one\-shot and faithful\-ES variants\) failed its acceptance band at every sampled scale, and these rescue comparisons proceed under a scoped decision to test rescue at a single confirmed\-destructive scale rather than across a replicated boundary\. This does not affect the screening conclusion—the cheap random\-κ\\kappacontrol is compared against the outer loop at the same scale—but it bounds the external reading: this is rescue at a fixed destructive scale, not across a reproduced failure curve\.

At the only perturbation scale that discriminates \(σ=1​e\-​1\\sigma=1\\text\{e\-\}1\), the random\-κ\\kappacontrolbeatthe heritable outer loop on preservation \(0\.906 vs 0\.805, sign\-consistent across both available seeds\), and the heritable loop also showed wider within\-trial spread \(Figure[3](https://arxiv.org/html/2606.29119#S5.F3)\)\. Atσ≤1​e\-​2\\sigma\\leq 1\\text\{e\-\}2the two sparse protocols tie \(perturbation too small to test the mask\); in the safe regime the dense baseline preserves*more*than the heritable loop\. The heritable loop’s only “wins” are against the dense strawman \(\+52\+52pp atσ=1​e\-​1\\sigma=1\\text\{e\-\}1\) and are fully attributable toκ\\kappa\-sparsity, which the lifecycle\-free control reproduces\. Heritability and selection—the evolutionary content—add nothing; the gate fires on the random\-κ\\kappacontrol\. These cross\-seed results aren=2n=2\(seeds 42, 43\); we report the ordering as sign\-consistent and treat the magnitudes as provisional\. The fullσ\\sigmasweep \(Figure[4](https://arxiv.org/html/2606.29119#S5.F4)\) shows the regime structure directly: the dense baseline collapses first—catastrophic atσ≥1​e\-​2\\sigma\\geq 1\\text\{e\-\}2—while both sparse protocols hold near the warm checkpoint throughσ=1​e\-​2\\sigma=1\\text\{e\-\}2and separate only atσ=1​e\-​1\\sigma=1\\text\{e\-\}1, the single scale where the control overtakes the rescue\.

![Refer to caption](https://arxiv.org/html/2606.29119v1/x2.png)Figure 3:The mask\-rescue project at the only discriminating perturbation scale \(σ=1​e\-​1\\sigma=1\\text\{e\-\}1;n=2n=2seeds, dots are per\-seed values\)\.\(a\)Mean preservation: the lifecycle\-free random\-κ\\kappacontrol \(0\.906\)*beats*the heritable rescue loop \(0\.805\); the dense matched\-uniform baseline is catastrophic \(off scale\)\. The rescue effect isκ\\kappa\-sparsity, not heritability or selection\.\(b\)Within\-trial organism spread: heritable rescue spreads*wider*than the control, consistent with the lifecycle locking early\-unlucky lineages\. Atσ≤1​e\-​2\\sigma\\leq 1\\text\{e\-\}2both sparse protocols preserve≈1\.0\\approx 1\.0and are indistinguishable\.![Refer to caption](https://arxiv.org/html/2606.29119v1/x3.png)Figure 4:Preservation across the fullσhigh\\sigma\_\{\\text\{high\}\}sweep, all three protocols, clean reruns \(n=2n=2seeds 42/43; black dots are per\-seed values\)\. The dense matched\-uniform baseline preserves atσ≤3​e\-​3\\sigma\\leq 3\\text\{e\-\}3but is catastrophically destroyed atσ≥1​e\-​2\\sigma\\geq 1\\text\{e\-\}2\(off scale; values annotated\)\. Both sparse protocols hold near the warm checkpoint throughσ=1​e\-​2\\sigma=1\\text\{e\-\}2and diverge only atσ=1​e\-​1\\sigma=1\\text\{e\-\}1, where the lifecycle\-free random\-κ\\kappacontrol overtakes the heritable rescue—the single discriminating scale of Figure[3](https://arxiv.org/html/2606.29119#S5.F3)is the rightmost point here\. The result substantiates in one view the regime claims made in prose: dense dies first, sparse ties below the discriminating scale, control\>\>rescue only at the top\.

### 5\.2Mask identity does not carry exploitable information \(the mask\-identity study\)

Pre\-registered claim: the*identity*of the optimalκ\\kappa\-sparse mask carries causal information an evolutionary search can exploit, beyond what single\-shot statistics recover, under a bilevel constrained fitnessF​\(M\)F\(M\)\.

On GSM8K→\\rightarrowARC\-Challenge atκ∈\{0\.01,0\.02,0\.05\}\\kappa\\in\\\{0\.01,0\.02,0\.05\\\}, the best single\-shot statistic mask \(gradient\-norm,F\(M\)=\+0\.324/\+0\.343/\+0\.344F\(M\)=\+0\.324/\+0\.343/\+0\.344\) was*itself the frontier*—it exceeded even unmasked full adaptation \(\+0\.302\+0\.302\) at every sparsity \(R≈1\.0R\\approx 1\.0;∼0\.95\{\\sim\}\\,0\.95on the stricterΔ​AccB\\Delta\\text\{Acc\}\_\{B\}view\), leaving≈0\{\\approx\}\\,0headroom \(Figure[5](https://arxiv.org/html/2606.29119#S5.F5)\)\. Baseline spread collapses withκ\\kappa\(0\.370→0\.304→0\.0760\.370\\rightarrow 0\.304\\rightarrow 0\.076\): mask identity matters*less*with more capacity, soκ=0\.01\\kappa=0\.01is the hardest test and showed no headroom even there\. Under the screening rule, the evolved\-mask outer loop \(Protocol 7\) wasnot built\.

![Refer to caption](https://arxiv.org/html/2606.29119v1/x4.png)Figure 5:Mask\-identity baseline\-kill on GSM8K→\\rightarrowARC\-Challenge\.\(a\)Per\-protocol mean constrained fitnessF​\(M\)F\(M\)atκ=0\.01\\kappa=0\.01\(n=3n=3\)\. The two single\-shot statistics \(gradient\-norm, Fisher; hatched\) sit*above*the unmasked full\-adaptation frontier \(\+0\.302\+0\.302, dashed\); magnitude and IMP actively hurt\.\(b\)The single\-shot frontier holds acrossκ\\kappa\(gradient\-norm and Fisher both exceed\+0\.302\+0\.302at everyκ\\kappa\), while the baseline spread—the signal a mask search could exploit—collapses from0\.3700\.370to0\.0760\.076asκ\\kappagrows\. There is no headroom for an evolved mask at any sparsity\.#### Evidence bar\.

The kill was declared on a single task pair \(ARC\-Challenge; the default Far pair, HumanEval, failed the pre\-flight tractability gate and was substituted—disclosed\)\. The pre\-registered rule requires a*positive*result to hold across all three pairs \(no cherry\-picking\); a*negative*is conclusive from any pair where the single\-shot baseline already exceeds the ceiling, since additional pairs cannot create headroom the easiest cell lacks\.

## 6What the Gate Bought

For the mask\-identity study \(Section[5\.2](https://arxiv.org/html/2606.29119#S5.SS2)\), the Phase\-0 gate \(pre\-flight \+ budget calibration \+ baseline\-kill across threeκ\\kappa\) cost∼50\{\\sim\}50–7070GPU\-hours on a single consumer GPU and returned aκ\\kappa\-robust kill\. Building and running the screened\-out evolutionary outer loop was estimated at∼400\+\{\\sim\}400\{\+\}GPU\-hours111Estimated as population \(∼\\sim16\)×\\timesgenerations \(∼\\sim50\)×\\timesper\-organism adapt\-and\-evaluate cost \(∼\\sim0\.5 GPU\-hour, the measured per\-trial cost of the baseline\-kill runs on the same hardware\)≈400\\approx 400GPU\-hours\. This is an order\-of\-magnitude estimate of work deliberately never run, not a measurement\.for the first cell alone, before any multi\-pair /κ\\kappa\-robustness sweep, plus weeks of implementation—a∼6\{\\sim\}6–8×8\\timesGPU saving on the first cell, far more once the unbuilt sweep and engineering time are counted \(Table[2](https://arxiv.org/html/2606.29119#S6.T2)\)\. The recurring shape across all cases: a static or single\-shot computation captured the effect on the project’s own metric, and it was identifiable*before*the outer loop was built\.

Table 2:What the Phase\-0 gate bought on the mask\-identity study\. Costs are GPU\-hours on a single consumer GPU; the screened\-out figure is the first\-cell estimate only and excludes the multi\-pair /κ\\kappa\-robustness sweep and weeks of implementation that the kill also foreclosed\.
## 7The Positive Condition — a Hypothesis, Not a Result

We state the organizing thesis explicitly as apost\-hoc synthesis, not a pre\-registered prediction: an evolutionary outer loop earns its keep only when it optimizes structure the inner optimizer cannot reach—a non\-differentiable, discrete, or black\-box objective gradient descent does not serve and no single\-shot statistic recovers \(R<90%R<90\\%at the gate\)\.

We havezero real\-task instancesof this condition in the present corpus\. The only positive evidence is a synthetic mechanism existence\-proof\.222A prior synthetic sandbox in which ES beat SGD on a discrete router by\>3​σ\{\>\}3\\sigma\(closing55\.9%55\.9\\%of the routing gap vs SGD’s0\.2%0\.2\\%\)—but only because the adapter→\\rightarrowdomain mapping was oracle\-pre\-aligned; removing the oracle collapses ES to the no\-skill floor\. Its authors label it a positive control, not a co\-evolution result\.It demonstrates the mechanism under an idealized condition; it is not a task win\. Apre\-registered successor designwas built to*hunt*for exactly this: a real, discrete/non\-differentiable objective that screensR<90%R<90\\%and on which an outer loop then wins\. Its first candidate substrate has now been screened—in\-context\-learning exemplar\-set selection \(choosing and ordering an 8\-shot set from a 64\-candidate pool to maximize held\-out accuracy; AG News, Qwen2\.5\-1\.5B\-Instruct\), the discrete black\-box regime this section names\. The Phase\-0 screen returned an analogous recoveryRscreen≈97%R\_\{\\text\{screen\}\}\\approx 97\\%: the strongest cheap\-tier method \(beam/greedy\-with\-restarts marginal\-gain construction\) recovered97%97\\%of the frontier reached by random search at the outer loop’s own matched budget \(10​Nscreen10\\,N\_\{\\text\{screen\}\}\), so theRscreen≥90%R\_\{\\text\{screen\}\}\\geq 90\\%gate fired andthe outer loop was not built\(Figure[6](https://arxiv.org/html/2606.29119#S7.F6)\)\.333HereRscreenR\_\{\\text\{screen\}\}is the pre\-registered, domain\-appropriate*result*\-ratio \(best swept\-cheap result\)/\(achievable frontier\) for a discrete objective with no gradient statistic—not the gain\-ratioR=s/GR=s/Gof Sections[3](https://arxiv.org/html/2606.29119#S3)–[5](https://arxiv.org/html/2606.29119#S5), though both express that a cheap method already sits at the frontier\. Single seed; the screen sits≈7\{\\approx\}7pp above the gate\.This direction is independently documented—cheap retrieval’s advantage over a naive exemplar set collapses as the shot count grows\(Bertschet al\.,[2024](https://arxiv.org/html/2606.29119#bib.bib14)\), while set\-level optimization helps mainly on tasks where cheap baselines leave headroom\(Yeet al\.,[2023](https://arxiv.org/html/2606.29119#bib.bib15)\), the very kind of task this hunt was designed to find\. This is the first*deliberate hunt*for anR<90%R<90\\%real\-task substrate, and the first candidate came upRscreen≥90%R\_\{\\text\{screen\}\}\\geq 90\\%—a consistent confirmation thatstrengthensthe “zero real\-task instances” finding above, not a refutation \(which would requireR<90%R<90\\%with the outer loop still failing\)\. A second pre\-registered branch \(a circuit\-discovery objective\) is unrun\. We stress that engaging this condition is itself a substantial research effort—a full discrete\-objective evolutionary study with its own multi\-week build and compute budget, deferred to future work—not a cheap confirmation withheld\. The paper stands entirely on the screening rule and the corpus in hand; the successor study is invoked only as the falsifiable test the rule invites, now with its first screen reported\.

![Refer to caption](https://arxiv.org/html/2606.29119v1/x5.png)Figure 6:The Direction C in\-context\-learning exemplar\-set screen \(AG News, Qwen2\.5\-1\.5B\-Instruct, single seed\)\. The cheap greedy/beam\-with\-restarts construction \(0\.762 held\-out test accuracy\) recoversRscreen≈97%R\_\{\\text\{screen\}\}\\approx 97\\%of the frontier reached by random search at the outer loop’s own matched budget \(10​Nscreen10\\,N\_\{\\text\{screen\}\}; 0\.785\)\. The2\.32\.3\-pp headroom is far below the10%10\\%the rule requires, so the gate fires and the evolutionary outer loop was not built\. This is the first deliberate hunt for anR<90%R<90\\%real\-task substrate, and its first candidate screenedR≥90%R\\geq 90\\%\.
## 8Limitations

- •Not independent\.Same lab, author, and design instinct; the two cases are not independent replications\. We present a series, and disclosefour further evolutionary efforts that produced no analyzable comparison—an honest file drawer—plusone further evolutionary outer loop that did produce an analyzable negativeand is not developed as a case: a faithful evolution\-strategy boundary replication that failed to reproduce the assumed warm\-start destruction boundary within its pre\-registered band \(it over\- or under\-shot at every sampled scale\)\. It is consistent with the thesis—another outer loop that did not beat its cheap reference—and we keep it out of the corpus count only to avoid conflating a replication failure with the screening\-rule cases \(its destructive\-regime consequence is noted in Section[5\.1](https://arxiv.org/html/2606.29119#S5.SS1)\)\.
- •Post\-hoc thesis\.Each project pre\-registered a*narrower*mechanism \(fitness alignment / heritability / mask identity\), not the reachability law; the law is an after\-the\-fact synthesis, offered as a falsifiable hypothesis \(Section[7](https://arxiv.org/html/2606.29119#S7)\)\.
- •Statistical strength\.Cross\-seed results aren=2n=2–33; we cite only out\-of\-noise effects as confirmations and label within\-noise differences as ties\.
- •Novelty boundary\.The per\-result findings echo known results \(Section[2](https://arxiv.org/html/2606.29119#S2)\)\. The novel claim is the screening rule and its validation as a pre\-build decision procedure\.
- •Single\-substrate screening\.The rule is validated on LoRA/MoE\-scale LLM work; generalization to other model families/scales is untested\.

## 9Discussion and Conclusion

The screening rule reframes a recurring, expensive mistake—building an evolutionary outer loop over a space the inner optimizer already serves—as a*cheap, advance, falsifiable check*\. Its value is decision\-theoretic: a few GPU\-days, run before implementation, that can foreclose a multi\-week build\. Across a within\-lab series, both analyzed cases fired the gate, and a companion factorial \(Case 1\) localizes the failure to the outer loop specifically rather than to the substrate it rode on\.

We deliberately decline the stronger story\. There is no real\-task positive in this corpus; the positive condition is a hypothesis with a named, pre\-registered successor design, whose first screen is reported here and whose circuit branch remains unrun\. What we claim is narrower and, we argue, more useful: a prospectively\-falsifiable rule for*when not to spend*on an evolutionary outer loop, validated where it cost us the most to learn it\.

## References

- I\. Abdi, A\. Gupta, M\. Mok, A\. Lu,et al\.\(2026\)Evolutionary strategies lead to catastrophic forgetting in LLMs\.arXiv preprint arXiv:2601\.20861\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Bahmani, R\. El Shawi, N\. Potikyan, and S\. Sakr \(2021\)To tune or not to tune? an approach for recommending important hyperparameters\.arXiv preprint arXiv:2108\.13066\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px4.p1.1)\.
- A\. Bertsch, M\. Ivgi, E\. Xiao, U\. Alon, J\. Berant, M\. R\. Gormley, and G\. Neubig \(2024\)In\-context learning with long\-context models: an in\-depth exploration\.arXiv preprint arXiv:2405\.00200\.Cited by:[§7](https://arxiv.org/html/2606.29119#S7.p2.8)\.
- J\. Frankle and M\. Carbin \(2019\)The lottery ticket hypothesis: finding sparse, trainable neural networks\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Frankle, G\. K\. Dziugaite, D\. M\. Roy, and M\. Carbin \(2020\)Linear mode connectivity and the lottery ticket hypothesis\.InInternational Conference on Machine Learning \(ICML\),Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Han, J\. Pool, J\. Tran, and W\. J\. Dally \(2015\)Learning both weights and connections for efficient neural networks\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Vol\.28\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- W\. Hoy, B\. Wang, and X\. Pan \(2026\)Matching accuracy, different geometry: evolution strategies vs GRPO in LLM post\-training\.arXiv preprint arXiv:2604\.01499\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Kirkpatrick, R\. Pascanu, N\. Rabinowitz, J\. Veness, G\. Desjardins, A\. A\. Rusu, K\. Milan, J\. Quan, T\. Ramalho, A\. Grabska\-Barwinska, D\. Hassabis, C\. Clopath, D\. Kumaran, and R\. Hadsell \(2017\)Overcoming catastrophic forgetting in neural networks\.Proceedings of the National Academy of Sciences \(PNAS\)114\(13\),pp\. 3521–3526\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- R\. Krishna, A\. Kalaiah, B\. Wu, M\. Naumov,et al\.\(2021\)Differentiable NAS framework and application to ads CTR prediction\.arXiv preprint arXiv:2110\.14812\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px4.p1.1)\.
- R\. Kumaresan \(2026\)Decomposing evolutionary mixture\-of\-LoRA architectures: the routing lever, the lifecycle penalty, and a substrate\-conditional boundary\.arXiv preprint arXiv:2605\.11153\.Note:Same\-author companion; cited as Case 1Cited by:[§4](https://arxiv.org/html/2606.29119#S4.p1.1)\.
- N\. Lee, T\. Ajanthan, and P\. H\. S\. Torr \(2019\)SNIP: single\-shot network pruning based on connection sensitivity\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Lehman, J\. Clune, D\. Misevic, C\. Adami, L\. Altenberg, J\. Beaulieu, P\. J\. Bentley, S\. Bernard, G\. Beslon, D\. M\. Bryson, N\. Cheney,et al\.\(2018\)The surprising creativity of digital evolution: a collection of anecdotes from the evolutionary computation and artificial life research communities\.arXiv preprint arXiv:1803\.03453\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- K\. Lenc, E\. Elsen, T\. Schaul, and K\. Simonyan \(2019\)Non\-differentiable supervised learning with evolution strategies and hybrid methods\.arXiv preprint arXiv:1906\.03139\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- L\. Li and A\. Talwalkar \(2019\)Random search and reproducibility for neural architecture search\.arXiv preprint arXiv:1902\.07638\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px4.p1.1)\.
- T\. Salimans, J\. Ho, X\. Chen, S\. Sidor, and I\. Sutskever \(2017\)Evolution strategies as a scalable alternative to reinforcement learning\.arXiv preprint arXiv:1703\.03864\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- B\. Sarkar, M\. Fellows, J\. A\. Duque, A\. Letcher,et al\.\(2025\)Evolution strategies at the hyperscale\.arXiv preprint arXiv:2511\.16652\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- K\. Schweighofer, C\. F\. Hayes, R\. Dailey, R\. Miikkulainen, and X\. Qiu \(2026\)Overcoming forgetting in LLM fine\-tuning with evolution strategies\.arXiv preprint arXiv:2605\.30148\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px3.p1.3)\.
- J\. Su, Y\. Chen, T\. Cai, T\. Wu, R\. Gao, L\. Wang, and J\. D\. Lee \(2020\)Sanity\-checking pruning methods: random tickets can win the jackpot\.arXiv preprint arXiv:2009\.11094\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- F\. P\. Such, V\. Madhavan, E\. Conti, J\. Lehman, K\. O\. Stanley, and J\. Clune \(2017\)Deep neuroevolution: genetic algorithms are a competitive alternative for training deep neural networks for reinforcement learning\.arXiv preprint arXiv:1712\.06567\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Sung, V\. Nair, and C\. Raffel \(2021\)Training neural networks with fixed sparse masks\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Vol\.34\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px2.p1.1)\.
- H\. J\. P\. Weerts, A\. C\. Mueller, and J\. Vanschoren \(2020\)Importance of tuning hyperparameters of machine learning algorithms\.arXiv preprint arXiv:2007\.07588\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px4.p1.1)\.
- A\. Yang, P\. M\. Esperança, and F\. M\. Carlucci \(2019\)NAS evaluation is frustratingly hard\.arXiv preprint arXiv:1912\.12522\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px4.p1.1)\.
- J\. Ye, Z\. Wu, J\. Feng, T\. Yu, and L\. Kong \(2023\)Compositional exemplars for in\-context learning\.arXiv preprint arXiv:2302\.05698\.Cited by:[§7](https://arxiv.org/html/2606.29119#S7.p2.8)\.
- K\. Yu, C\. Sciuto, M\. Jaggi, C\. Musat, and M\. Salzmann \(2019\)Evaluating the search phase of neural architecture search\.arXiv preprint arXiv:1902\.08142\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px4.p1.1)\.
- X\. Zhang, J\. Clune, and K\. O\. Stanley \(2017\)On the relationship between the OpenAI evolution strategy and stochastic gradient descent\.arXiv preprint arXiv:1712\.06564\.Cited by:[§2](https://arxiv.org/html/2606.29119#S2.SS0.SSS0.Px1.p1.1)\.