Learning Transferable Predictability Representations

# Learning Transferable Predictability Representations
Source: [https://arxiv.org/html/2605.30592](https://arxiv.org/html/2605.30592)
Diyali Goswami1&Auroop R\. Ganguly1,2,3 1Sustainability and Data Sciences Laboratory \(SDS Lab\), Northeastern University, Boston, MA, USA 2AI4CaS: AI for Climate and Sustainability, Institute for Experiential AI, Northeastern University, Boston, MA, USA 3Pacific Northwest National Laboratory \(PNNL\), Richland, WA, USA \{goswami\.di, a\.ganguly\}@northeastern\.edu

###### Abstract

We study the problem of assigning a scalar score to a short trajectory window that reflects its position on an ordered continuum of predictability regimes, spanning structured deterministic dynamics to unstructured stochastic noise\. Existing methods address deterministic\-versus\-stochastic discrimination within a single system and do not produce scores with a consistent numerical interpretation across systems\. We formalize this as ordinal estimation over a five\-level predictability ladder and identify a structural source of cross\-system ambiguity: ranking supervision alone leaves the score coordinate unfixed up to a monotone reparameterization, which we term the gauge freedom of ordinal scoring\. We propose the Gauge\-Fixed Ordinal Network \(GON\), a temporal convolutional model trained with an anchor\-and\-variance objective that pins level\-wise score means to shared target coordinates\. GON operates on 2\-jet features that expose local trajectory geometry, preserved by smooth flows and disrupted by stochastic surrogate procedures\. On five held\-out dynamical systems, initializing from a pretrained GON checkpoint consistently outperforms training from scratch across all window budgets, with adaptation depth reflecting geometric proximity to the training family\. Zero\-shot scores retain ordinal structure at the stochastic boundary, where surrogate procedures most strongly disrupt nonlinear geometry, and pretrained initialization consistently beats scratch across all window budgets\. Pairwise discrimination and globally coherent ordinal scoring are distinct properties requiring a stable score coordinate for cross\-system transfer, with direct implications for predictability assessment, model selection, and early\-warning diagnostics across natural and engineered dynamical systems\.

††footnotetext:Code will be released upon acceptance and repository cleanup\.## 1Introduction

An observed time series can often be viewed as a projection of an underlying dynamical system, though the generative process may be deterministic, stochastic, or a mixture of both\. The source of predictive structure shapes which modeling strategies are appropriate\. Deterministic chaotic systems are governed by an underlying flow that supports geometry\-aware representations, even when long\-horizon forecasting is limited by sensitive dependence on initial conditions\. Purely stochastic processes are better described through statistical regularities\. Both regimes can share low\-order moments such as autocorrelation and power spectrum, making the distinction difficult to draw from observations alone\. A signal may reflect low\-dimensional deterministic dynamics, a stochastic process, or an intermediate regime that retains some structure while losing others\.

Classical tools approach this distinction within a single system\. Surrogate hypothesis tests\(Theileret al\.,[1992](https://arxiv.org/html/2605.30592#bib.bib10); Schreiber and Schmitz,[1996](https://arxiv.org/html/2605.30592#bib.bib11)\)ask whether a signal departs from a chosen stochastic null\. Lyapunov exponent estimators\(Wolfet al\.,[1985](https://arxiv.org/html/2605.30592#bib.bib8); Rosensteinet al\.,[1993](https://arxiv.org/html/2605.30592#bib.bib9)\)quantify instability under an assumed deterministic model\. Complexity statistics such as permutation entropy\(Bandt and Pompe,[2002](https://arxiv.org/html/2605.30592#bib.bib47)\)and sample entropy\(Richman and Moorman,[2000](https://arxiv.org/html/2605.30592#bib.bib48)\)compress a signal into a scalar intended to separate deterministic from stochastic behavior\. Machine learning approaches have extended work on predictive structure in dynamical systems, including neural differential equation models\(Chenet al\.,[2018](https://arxiv.org/html/2605.30592#bib.bib19)\), neural operators\(Liet al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib25); Luet al\.,[2021](https://arxiv.org/html/2605.30592#bib.bib26)\), and pretrained models for chaotic forecasting\(Laiet al\.,[2025](https://arxiv.org/html/2605.30592#bib.bib27)\)\. These methods characterize structure within a system or family, and a new signal typically requires new calibration, new surrogates, or a fresh thresholding decision\. A detailed comparison is provided in Appendix[A](https://arxiv.org/html/2605.30592#A1)\.

This paper addresses a specific problem: given a trajectory window, assign it a position on an ordered continuum of predictability regimes\. We use predictability to mean the degree to which the underlying dynamics constrain future trajectory evolution, a property of the governing flow rather than of any particular model or observation process\. We instantiate this continuum as a five\-level*predictability ladder*spanning structured deterministic dynamics, weakly chaotic regimes, strongly chaotic regimes, structured stochastic surrogates, and unstructured noise\. Standard deterministic\-versus\-stochastic discrimination corresponds to one boundary in this richer ordinal problem\. The central question is whether a learned ordinal score transfers across systems, rather than merely separating classes within a single one\.

The ordinal setting introduces a structural ambiguity\. If a scalar score correctly orders the ladder levels, any strictly increasing reparameterization preserves the same ordinal decisions after a corresponding threshold adjustment\. Correct ordering alone does not determine the numerical coordinate of the score, an identifiability property of ordered regression models\(McCullagh,[1980](https://arxiv.org/html/2605.30592#bib.bib98)\), leaving scores across systems non\-comparable even when both induce identical ordinal predictions\. We refer to this as the*gauge freedom*of ordinal scoring and identify it as the structural barrier to cross\-system transfer; post\-hoc per\-system recalibration cannot substitute, because the intended use case is zero\-shot transfer where no target\-system data is available to fit a calibration mapping\.

To address this, we introduce the*Gauge\-Fixed Ordinal Network*\(GON\), a temporal convolutional model trained with an anchor\-based ordinal objective\. The anchor fixes target locations for level\-wise score means and penalizes excessive within\-level spread, producing a scalar score with a stable coordinate convention\. GON operates on 2\-jet features \(position, velocity, and acceleration\), which expose local trajectory geometry\(Koenderink and van Doorn,[1987](https://arxiv.org/html/2605.30592#bib.bib97)\)preserved by smooth flows and disrupted by surrogate procedures; this representation has been applied to learned physical dynamics inCranmeret al\.\([2020](https://arxiv.org/html/2605.30592#bib.bib20)\)\. We train on twelve dissipative chaotic ODE systems, chosen for their bounded attractors and varied geometric structure, and evaluate zero\-shot transfer and few\-shot adaptation on five fully held\-out systems\.

The main contributions are as follows\.

1. 1\.Problem formulation\.We formalize ordinal predictability estimation and introduce the predictability ladder, identifying gauge freedom as the structural source of cross\-system score ambiguity\.
2. 2\.Method\.We propose GON, which combines 2\-jet trajectory features with an anchor\-based objective to learn a shared score coordinate across systems\.
3. 3\.Empirical results\.We show that gauge fixing enables cross\-system ordinal transfer: pretrained GON outperforms training from scratch across all held\-out systems, zero\-shot scores retain ordinal coherence at the surrogate boundary, and adaptation depth reflects geometric proximity to the training family\.

## 2Problem Formulation

We formalize*ordinal predictability estimation*: given a trajectory window, assign it a position on an ordered predictability spectrum that remains meaningful across systems, subsuming binary deterministic\-versus\-stochastic discrimination as a special case\.

### 2\.1The Predictability Ladder

Letx=\(xt\)t=1T∈ℝd×Tx=\(x\_\{t\}\)\_\{t=1\}^\{T\}\\in\\mathbb\{R\}^\{d\\times T\}denote a multivariate time series generated by an unknown process\. We define a five\-level ordered taxonomy

ℛ=\{L0,L1,L2,L3,L4\}\\mathcal\{R\}=\\\{L\_\{0\},L\_\{1\},L\_\{2\},L\_\{3\},L\_\{4\}\\\}called the*predictability ladder*, arranged from more predictable to less predictable regimes\.

###### Definition 1\(Predictability Ladder\)\.

The levels satisfy

L0≺L1≺L2≺L3≺L4,L\_\{0\}\\prec L\_\{1\}\\prec L\_\{2\}\\prec L\_\{3\}\\prec L\_\{4\},with the following interpretation:

- L0L\_\{0\}Stable deterministic\.Fixed points or limit cycles, where the present state strongly constrains the future, e\.g\., deterministic flows with non\-positive largest Lyapunov exponentλmax≤0\\lambda\_\{\\max\}\\leq 0\.
- L1L\_\{1\}Weakly chaotic\.Deterministic dynamics with small positive instability and a longer but finite predictability horizon,0<λmax​τ≪10<\\lambda\_\{\\max\}\\tau\\ll 1\.
- L2L\_\{2\}Strongly chaotic\.Deterministic dynamics with stronger instability and rapid trajectory divergence,λmax​τ≫0\\lambda\_\{\\max\}\\tau\\gg 0\.
- L3L\_\{3\}Structured stochastic\.Processes that preserve selected linear statistics of a deterministic signal while disrupting nonlinear structure, e\.g\., surrogate transformationsxt\(3\)=𝒮​\(xt\(2\)\)x\_\{t\}^\{\(3\)\}=\\mathcal\{S\}\(x\_\{t\}^\{\(2\)\}\)that preserve the power spectrum but destroy deterministic phase\-space geometry\.
- L4L\_\{4\}Unstructured stochastic\.i\.i\.d\. noise with no predictive structure beyond the marginal distribution,xt∼𝒫x\_\{t\}\\sim\\mathcal\{P\}andI​\(xt;xt\+τ\)=0I\(x\_\{t\};x\_\{t\+\\tau\}\)=0forτ\>0\\tau\>0\.

This ladder reflects two axes of limited predictability\. AlongL0L\_\{0\}–L2L\_\{2\}, regimes remain deterministic, with predictability degrading as Lyapunov instability shortens the forecast horizon\. TheL2→L3L\_\{2\}\\\!\\to\\\!L\_\{3\}transition instead destroys deterministic geometry: surrogate procedures preserve linear statistics while removing nonlinear flow structure\. LevelL4L\_\{4\}removes all temporal structure beyond the marginal distribution\. Finer granularity can be introduced within adjacent levels by progressively increasing instability \(L0L\_\{0\}–L2L\_\{2\}\), disrupting deterministic geometry \(L2L\_\{2\}–L3L\_\{3\}\), or removing residual temporal dependence \(L3L\_\{3\}–L4L\_\{4\}\)\. This ordering parallels predictability\-horizon arguments in nonlinear dynamical systems, where increasing trajectory divergence reduces forecastability\.

![Refer to caption](https://arxiv.org/html/2605.30592v1/figures/predictability_ladder.png)Figure 1:Predictability ladder construction\. Top: state\-space projections\. Bottom: corresponding time series\. LevelsL0L\_\{0\}–L2L\_\{2\}arise from the same dynamical system under different regimes\. LevelL3L\_\{3\}is obtained by surrogate generation that preserves selected linear statistics while disrupting deterministic structure\. LevelL4L\_\{4\}corresponds to unstructured noise\.
### 2\.2Ordinal Detectability

###### Definition 2\(Ordinal Detectability\)\.

A scoring functionEθ:ℝd×T→ℝE\_\{\\theta\}:\\mathbb\{R\}^\{d\\times T\}\\rightarrow\\mathbb\{R\}is*ordinally detectable*under a distribution𝒟\\mathcal\{D\}over\(x,y\)∈𝒳×\{0,1,2,3,4\}\(x,y\)\\in\\mathcal\{X\}\\times\\\{0,1,2,3,4\\\}if

𝔼​\[Eθ​\(x\)∣y=i\]<𝔼​\[Eθ​\(x\)∣y=j\]∀i<j\.\\mathbb\{E\}\[E\_\{\\theta\}\(x\)\\mid y=i\]<\\mathbb\{E\}\[E\_\{\\theta\}\(x\)\\mid y=j\]\\qquad\\forall\\,i<j\.

Letμk=𝔼​\[Eθ​\(x\)∣y=k\]\\mu\_\{k\}=\\mathbb\{E\}\[E\_\{\\theta\}\(x\)\\mid y=k\]\. To quantify global ordinal structure beyond pairwise separation, we use the*monotonicity slope*

β=∑k\(k−k¯\)​\(μk−μ¯\)∑k\(k−k¯\)2,\\beta=\\frac\{\\sum\_\{k\}\(k\-\\bar\{k\}\)\(\\mu\_\{k\}\-\\bar\{\\mu\}\)\}\{\\sum\_\{k\}\(k\-\\bar\{k\}\)^\{2\}\},\(1\)the least\-squares slope of level means on ladder indices\. A method may achieve a strong AUROC at a single boundary while producingβ≈0\\beta\\approx 0if the full ladder is not coherently ordered\. We reportβ\\betaboth in\-distribution and under transfer to systems withheld entirely from training\.

### 2\.3Cross\-System Score Ambiguity

A scalar ordinal score is useful across systems only if its numerical values retain a stable interpretation; ranking alone does not guarantee this\. Given a scoreE:𝒳→ℝE:\\mathcal\{X\}\\to\\mathbb\{R\}and thresholdsτ0<τ1<τ2<τ3\\tau\_\{0\}<\\tau\_\{1\}<\\tau\_\{2\}<\\tau\_\{3\}, the induced predictor ishE​\(x\)=∑m=03𝟏​\{E​\(x\)\>τm\}h\_\{E\}\(x\)=\\sum\_\{m=0\}^\{3\}\\mathbf\{1\}\\\{E\(x\)\>\\tau\_\{m\}\\\}, mappingxxto one of five ordered levels\.

###### Proposition 1\(Monotone reparameterization invariance\)\.

Letf:ℝ→ℝf:\\mathbb\{R\}\\to\\mathbb\{R\}be strictly increasing\. The predictor induced byEEis unchanged underE~=f∘E\\tilde\{E\}=f\\circ Ewhen thresholds are transformed toτ~m=f​\(τm\)\\tilde\{\\tau\}\_\{m\}=f\(\\tau\_\{m\}\)\.*\(Proof in Appendix[B](https://arxiv.org/html/2605.30592#A2)\.\)*

Ordinal supervision therefore determines score*ordering*but not its numerical*coordinate*: the score is identifiable only up to a monotone reparameterization\. While this ambiguity is harmless within a single system, it becomes consequential under transfer, where identical score values may correspond to different predictability regimes\. Cross\-system comparison thus requires a shared coordinate convention, which the anchor objective provides by fixing level\-wise score locations globally\.

## 3Method

GON has three components: 2\-jet preprocessing, a temporal convolutional encoder, and a gauge\-fixed ordinal objective\.

### 3\.12\-Jet Preprocessing

Raw state coordinates carry system\-specific scaling, orientation, and units\. Representing each window by its 2\-jet exposes local trajectory geometry independent of these system\-specific factors\. For a smooth trajectoryx​\(t\)x\(t\), the 2\-jet is the tuple

\(x​\(t\),x˙​\(t\),x¨​\(t\)\),\(x\(t\),\\dot\{x\}\(t\),\\ddot\{x\}\(t\)\),consisting of position, velocity, and acceleration\. These quantities summarize local behavior up to second order and provide direct access to geometric features such as tangent direction and curvature\. For trajectories generated by smooth ODEs, these components are linked by the governing flow\. Surrogate procedures, by contrast, preserve selected linear statistics while disrupting the original nonlinear relation among position, velocity, and acceleration\. This makes the 2\-jet a natural representation for distinguishing deterministic trajectories from structured stochastic surrogates\.

##### Implementation\.

Given a windowx∈ℝd×Tx\\in\\mathbb\{R\}^\{d\\times T\}, we first apply a Savitzky–Golay filter \(window length77, polynomial order33; clamped to signal length and kept odd if shorter\) to obtain a smoothed trajectoryx^\\hat\{x\}\. Central finite differences are then used to estimate derivatives, as they provide symmetric second\-order accurate approximations while minimizing phase bias in the local trajectory geometry:

x˙t\\displaystyle\\dot\{x\}\_\{t\}=x^t\+1−x^t−12​Δ​t,\\displaystyle=\\frac\{\\hat\{x\}\_\{t\+1\}\-\\hat\{x\}\_\{t\-1\}\}\{2\\Delta t\},\(2\)x¨t\\displaystyle\\ddot\{x\}\_\{t\}=x^t\+1−2​x^t\+x^t−1Δ​t2\.\\displaystyle=\\frac\{\\hat\{x\}\_\{t\+1\}\-2\\hat\{x\}\_\{t\}\+\\hat\{x\}\_\{t\-1\}\}\{\\Delta t^\{2\}\}\.\(3\)One timestep is removed from each boundary to avoid edge effects\. The resulting representation is

J2​\(x\)=\[x^,x˙,x¨\]∈ℝ3​d×\(T−2\),J^\{2\}\(x\)=\[\\hat\{x\},\\,\\dot\{x\},\\,\\ddot\{x\}\]\\in\\mathbb\{R\}^\{3d\\times\(T\-2\)\},which serves as the network input\.

### 3\.2Architecture

The encoder is a temporal convolutional network \(TCN\) of dilated residual blocks that maps the 2\-jet input to a hidden sequenceh∈ℝH×T′h\\in\\mathbb\{R\}^\{H\\times T^\{\\prime\}\}\. After normalization, multi\-scale average pooling aggregateshhinto a vectorz∈ℝ4​Hz\\in\\mathbb\{R\}^\{4H\}, which a two\-layer MLP maps to a scalar score:

Eθ​\(x\)=𝐖2​GELU​\(LN​\(𝐖1​z\+b1\)\)\+b2\.E\_\{\\theta\}\(x\)=\\mathbf\{W\}\_\{2\}\\,\\mathrm\{GELU\}\\\!\\left\(\\mathrm\{LN\}\(\\mathbf\{W\}\_\{1\}z\+b\_\{1\}\)\\right\)\+b\_\{2\}\.\(4\)Outputs are smoothly clipped viaEθ←c​tanh⁡\(Eθ/c\)E\_\{\\theta\}\\leftarrow c\\,\\tanh\(E\_\{\\theta\}/c\),c=5c=5, to keep scores aligned with the anchor range\. Full details appear in Appendix[C](https://arxiv.org/html/2605.30592#A3)\.

### 3\.3Gauge\-Fixed Ordinal Objective

Ranking\-based supervision does not pin a unique score coordinate \(Section[2\.3](https://arxiv.org/html/2605.30592#S2.SS3)\)\. We fix one by assigning target locations

\{tk\}k=04=\{−4,−2,0,\+2,\+4\}\\\{t\_\{k\}\\\}\_\{k=0\}^\{4\}=\\\{\-4,\-2,0,\+2,\+4\\\}and training the model so that each level concentrates around its anchor\. These numeric targets are arbitrarily ordered anchors chosen to fix a shared coordinate; any monotonic rescaling yields an equivalent ordinal representation\. Equal spacing is chosen for simplicity; the ordinal structure is invariant to any monotone rescaling of the targets, so the specific values affect only the absolute score scale and not relative ordering\. The objective is

ℒ​\(θ\)=λa​∑k=04\(μk−tk\)2\+λv​∑k=04\[σk−σ∗\]\+2,\\mathcal\{L\}\(\\theta\)=\\lambda\_\{a\}\\sum\_\{k=0\}^\{4\}\(\\mu\_\{k\}\-t\_\{k\}\)^\{2\}\+\\lambda\_\{v\}\\sum\_\{k=0\}^\{4\}\[\\sigma\_\{k\}\-\\sigma^\{\*\}\]\_\{\+\}^\{2\},\(5\)whereμk\\mu\_\{k\}andσk\\sigma\_\{k\}are the mini\-batch mean and standard deviation of scores for levelkk,\[⋅\]\+=max⁡\(0,⋅\)\[\\cdot\]\_\{\+\}=\\max\(0,\\cdot\), andσ∗=0\.3\\sigma^\{\*\}=0\.3\. The anchor term fixes level\-wise score locations, selecting a canonical representative from the class of monotone\-equivalent solutions \(Appendix[E](https://arxiv.org/html/2605.30592#A5)\); the variance term limits within\-level dispersion, together defining a shared affine coordinate across systems\.

##### Relation to margin\-based objectives\.

A natural alternative is margin lossℒmargin=∑k=03max⁡\(0,γ−\(μk\+1−μk\)\)\\mathcal\{L\}\_\{\\text\{margin\}\}=\\sum\_\{k=0\}^\{3\}\\max\(0,\\,\\gamma\-\(\\mu\_\{k\+1\}\-\\mu\_\{k\}\)\), which enforces adjacent\-level separation but leaves the absolute score axis unfixed\. Scores therefore remain non\-comparable across systems even when levels are well\-separated; Section[4\.4](https://arxiv.org/html/2605.30592#S4.SS4)tests this directly\.

### 3\.4Training Details

We setλa=1\.0\\lambda\_\{a\}=1\.0with linear warmup over the first five epochs andλv=0\.01\\lambda\_\{v\}=0\.01\. Optimization uses AdamW with learning rate2×10−32\\times 10^\{\-3\}, weight decay10−410^\{\-4\}, cosine decay over3030epochs, gradient clipping at5\.05\.0, and EMA decay0\.9950\.995\. Training uses a batch size of6464\. Weighted sampling balances ladder levels\. The full experimental suite required approximately 40–50 GPU\-hours on a single A100\.

##### Data Augmentation\.

Augmentation is applied to raw trajectory windowsx∈ℝd×Tx\\in\\mathbb\{R\}^\{d\\times T\}before 2\-jet preprocessing and includes additive Gaussian noise \(σrel=0\.05\\sigma\_\{\\text\{rel\}\}=0\.05\), multiplicative scaling \(×\[0\.9,1\.1\]\\times\[0\.9,1\.1\]\), temporal masking \(5%5\\%\), linear drift \(±0\.05\\pm 0\.05\), and causal time shifts \(≤5\\leq 5steps\)\. Each operation is applied independently with probability0\.50\.5; the full pipeline activates with probability0\.70\.7\. Augmentation precedes smoothing, so the filter attenuates discontinuities introduced by masking or time shifts before finite differences are computed\.

## 4Experiments

The experiments address four questions:

1. 1\.Does the gauge\-fixed coordinate remain calibrated on the source distribution, and does it transfer in zero\-shot to held\-out systems?
2. 2\.Does pretraining provide a better initialization than scratch for adapting to unseen systems under limited labeled windows?
3. 3\.Which objective component is responsible for cross\-system coordinate stability?
4. 4\.Does coherent 2\-jet structure contribute beyond dimensionality\-matched controls?

### 4\.1Experimental Setup

##### Source and target systems\.

Training uses 12 three\-dimensional dissipative chaotic ODE systems spanning a range of attractor geometries: Chen, Chua, Duffing, Finance, Genesio–Tesi, Hastings–Powell, Lorenz\-63, Lorenz\-84, Rössler, Rucklidge, Shimizu–Morioka, and Halvorsen\. Each system is simulated to produce trajectories of lengthN=4096N=4096, with system\-specific timesteps chosen to resolve dominant dynamical timescales;Δ​t\\Delta tis determined from the observed sampling rate and requires no system\-level knowledge\.

Ladder levels follow Definition[1](https://arxiv.org/html/2605.30592#Thmdefinition1): deterministic regimes \(L0L\_\{0\}–L2L\_\{2\}\) are assigned usingλ1\\lambda\_\{1\},L3L\_\{3\}is generated by applying surrogate procedures \(AAFT, IAAFT, WLS, phase\-shuffle\) toL2L\_\{2\}trajectories, andL4L\_\{4\}is i\.i\.d\. noise\. Full simulation and labeling details appear in Appendix[D](https://arxiv.org/html/2605.30592#A4)\.

Cross\-system transfer is evaluated on five held\-out systems: Thomas, Tigan, Newton–Leipnik, Rabinovich–Fabrikant, and the forced pendulum\.

##### Windowing and preprocessing\.

Trajectories are segmented into windows of lengthT=256T=256with stride128128\. Inputs are normalized via per\-trajectory, per\-channel z\-score standardization, requiring no source\-distribution statistics and applying equally to any held\-out system\. After 2\-jet preprocessing, each example is a9×2549\\times 254tensor\. Windows are the sole input unit; system\-level information is unknown throughout\. The adaptation experiment measures how many windows suffice to recalibrate a pretrained model to a new system under limited data\.

##### Metrics\.

The monotonicity slopeβ\\beta\(Definition[2](https://arxiv.org/html/2605.30592#Thmdefinition2)\) measures global ordinal coherence\. Cross\-method zero\-shot comparisons use the scale\-normalizedβnorm\\beta\_\{\\text\{norm\}\}; rawβ\\betais used within the GON family, where the score scale is fixed by construction\. Adjacent\-pair AUROC at each boundary\(L0,L1\)\(L\_\{0\},L\_\{1\}\)through\(L3,L4\)\(L\_\{3\},L\_\{4\}\)serves as a diagnostic, since a method can separate one transition well while failing global coherence\. For in\-distribution calibration, we also report anchor drift

δanchor=15​∑k=04\|μk−tk\|,\\delta\_\{\\text\{anchor\}\}=\\frac\{1\}\{5\}\\sum\_\{k=0\}^\{4\}\|\\mu\_\{k\}\-t\_\{k\}\|,measuring how closely the learned level means match the anchor targets\.

### 4\.2Zero\-Shot Evaluation

Models are trained on the source systems and applied directly to the held\-out systems\.

##### In\-distribution coordinate calibration\.

Before assessing transfer, it is useful to verify that the gauge\-fixed score remains well aligned with its intended coordinate system on held\-out source\-domain data\. Table[1](https://arxiv.org/html/2605.30592#S4.T1)reports level\-wise score statistics on the test split of the training systems\. GON closely matches the anchor targets, with mean anchor driftδanchor=0\.0097\\delta\_\{\\text\{anchor\}\}=0\.0097\. The within\-level standard deviations \(0\.080\.08–0\.320\.32\) remain well below the 2\-unit spacing between adjacent anchors\.

Table 1:In\-distribution level calibration \(mean±\\pmstd\)\. Anchor targets\{−4,−2,0,\+2,\+4\}\\\{\-4,\-2,0,\+2,\+4\\\}; mean anchor driftδanchor=0\.0097\\delta\_\{\\text\{anchor\}\}=0\.0097\. Within\-level standard deviations remain well below the 2\-unit inter\-anchor spacing, confirming well\-separated source\-distribution coordinates and supporting zero\-shot transfer\.
##### Cross\-system zero\-shot transfer\.

Figure[2](https://arxiv.org/html/2605.30592#S4.F2)compares GON with neural baselines on the held\-out systems\. All supervised baselines cluster tightly atβnorm≈0\.46\\beta\_\{\\text\{norm\}\}\\approx 0\.46–0\.480\.48despite using different objectives\. GON achieves the highest value,βnorm=0\.559\\beta\_\{\\text\{norm\}\}=0\.559\. Full per\-boundary AUROC and per\-system neural baseline results are reported in Appendix[G](https://arxiv.org/html/2605.30592#A7)\. The transfer pattern is selective rather than uniform\. The strongest zero\-shot separation occurs at the stochastic boundary \(L2→L3L\_\{2\}\\to L\_\{3\}andL3→L4L\_\{3\}\\to L\_\{4\}\), while deterministic\-side boundaries remain weaker:AUROC01\\text\{AUROC\}\_\{01\}is near chance andAUROC12\\text\{AUROC\}\_\{12\}is only moderate \(Appendix[G](https://arxiv.org/html/2605.30592#A7)\)\. The gauge\-fixed score thus retains meaningful ordinal structure under distribution shift, with transfer concentrated where surrogate procedures most strongly disrupt nonlinear geometry\.

![Refer to caption](https://arxiv.org/html/2605.30592v1/figures/ordinal_structure.png)Figure 2:Scale\-normalized monotonicity slopeβnorm\\beta\_\{\\text\{norm\}\}\(rawβ\\betadivided by the inter\-level step size of2\.02\.0, soβnorm=1\\beta\_\{\\text\{norm\}\}=1indicates perfectly monotone level\-wise means\) on five held\-out systems under zero\-shot transfer; higher is better\. All supervised baselines cluster atβnorm≈0\.46\\beta\_\{\\text\{norm\}\}\\approx 0\.46–0\.480\.48regardless of training objective; GON achieves0\.5590\.559via a fixed shared ordinal coordinate\.

### 4\.3Adaptation from Pretrained Initialization

We evaluate whether GON pretraining provides a better initialization than training from scratch acrossk∈\{5,10,…,100\}k\\in\\\{5,10,\\ldots,100\\\}labeled windows per held\-out system, averaged over five random seeds\. Three strategies are compared:scratch\(random initialization\),pre\_head\(frozen encoder, readout fine\-tuned\), andpre\_all\(full model fine\-tuned from the pretrained checkpoint\)\. Table[2](https://arxiv.org/html/2605.30592#S4.T2)reports averages over all five systems; Figure[3](https://arxiv.org/html/2605.30592#S4.F3)and Appendix[H](https://arxiv.org/html/2605.30592#A8)show per\-system results and fine\-tuning details\.

![Refer to caption](https://arxiv.org/html/2605.30592v1/figures/fig_winshot.png)Figure 3:Adaptation from pretrained initialization vs\. scratch per held\-out system \(k≤100k\\leq 100labeled windows, mean±\\pmstd over 5 seeds\)\. Strategies:scratch\(random initialization\),pre\_head\(frozen encoder, readout fine\-tuned\),pre\_all\(full fine\-tune from pretrained checkpoint\)\. Dashed line marks zero\-shot GON performance\.pre\_allconsistently exceeds zero\-shot with lower variance;scratchnever recovers ordinal structure on Rabinovich–Fabrikant even atk=100k=100\.##### Pretrained initialization is faster and less variable than scratch\.

Atk=5k=5,pre\_allleads scratch by\+0\.776\+0\.776inβ\\beta\(1\.3191\.319vs\.0\.5430\.543\) and\+0\.191\+0\.191inAUROC23\\text\{AUROC\}\_\{23\}\(0\.7580\.758vs\.0\.5670\.567\), and this gap does not close atk=100k=100\(Table[2](https://arxiv.org/html/2605.30592#S4.T2)\), indicating that the pretrained encoder encodes geometric structure that random initialization cannot recover within this window budget\.pre\_headreliably exceeds scratch onAUROC23\\text\{AUROC\}\_\{23\}but captures little ofpre\_all’s gain inβ\\beta, confirming that effective adaptation requires updating the encoder\.

Table 2:Adaptation: meanAUROC23\\text\{AUROC\}\_\{23\}andβ\\betaaveraged over 5 held\-out systems and 5 random seeds \(std across systems\)\.scratch: random initialization;pre\_head: frozen encoder, readout fine\-tuned;pre\_all: full fine\-tune from pretrained checkpoint\. Thepre\_alladvantage over scratch does not close atk=100k=100\.
##### Transfer depth scales with geometric proximity\.

Systems with polynomial or exponential vector fields represented in the source distribution \(Rabinovich–Fabrikant, Newton–Leipnik, Tigan\) adapt immediately or are solved at zero\-shot\. Rabinovich–Fabrikant is the clearest case:scratchyieldsβ≈0\.01\\beta\\approx 0\.01–0\.540\.54across allkk, never recovering ordinal structure, whereaspre\_allexceeds zero\-shot fromk=10k=10\. Thomas and the forced pendulum are harder: both contain state\-dependent sinusoidal vector\-field terms absent from training \(e\.g\.,sin⁡θ\\sin\\theta,sin⁡y\\sin y\), producing unseen oscillatory curvature in the 2\-jet\. In both cases,pre\_allrecovers global ordinal structure but fails to resolve theL2→L3L\_\{2\}\\to L\_\{3\}boundary, identifying sinusoidal vector\-field geometry as the primary gap in source\-distribution coverage\.

### 4\.4Ablations

The ablations ask which parts of the method matter for transfer\.

##### Objective ablation: why the anchor matters\.

Table[3](https://arxiv.org/html/2605.30592#S4.T3)shows that coordinate stability, not ordinal coherence alone, is critical for transfer\. Replacing the anchor with a margin objective leaves in\-distribution AUROC unchanged by at most0\.0100\.010across adjacent pairs and produces comparable scale\-normalized transfer \(βnorm=0\.554\\beta\_\{\\text\{norm\}\}=0\.554vs\.0\.5590\.559\), yet anchor drift rises from0\.0050\.005to2\.5562\.556\. A model can therefore retain ordinal coherence while losing coordinate stability, confirming that relative level separation is insufficient for cross\-system comparability: without a fixed coordinate, scores on unseen systems carry no consistent numerical meaning\. The variance term plays a secondary but consistent role: removing it increases within\-level spread from0\.1950\.195to0\.2300\.230and lowers transferβ\\betafrom1\.1181\.118to1\.0551\.055\.

Table 3:Loss ablation\. Anchor drift measures coordinate stability; transferβ\\betameasures ordinal transfer on held\-out systems\. Anchor\-based objectives yield stable coordinates and better transfer\.
##### Input ablation: why the 2\-jet helps\.

Table[4](https://arxiv.org/html/2605.30592#S4.T4)shows that coherent derivative structure improves transfer beyond raw states\. Adding first derivatives improvesβ\\betafrom1\.1321\.132to1\.3361\.336; the full 2\-jet improves further toβ=1\.504\\beta=1\.504andAUROC23=0\.846\\text\{AUROC\}\_\{23\}=0\.846\. Shuffling derivatives across time dropsβ\\betato0\.9830\.983, confirming that temporal coherence among position, velocity, and acceleration drives the gain\. A control replacing derivatives with per\-pass Gaussian noise remains near the raw\-state baseline, ruling out channel count\.

Table 4:Jet ablation: transferβ\\betaandAUROC23\\text\{AUROC\}\_\{23\}over five held\-out systems \(anchor\+\+variance, no Savitzky–Golay smoothing\)\. Coherent 2\-jet features improve transfer; shuffled and random \(†\\dagger\) controls degrade performance\.
##### Noise robustness\.

GON remains robust under moderate corruption, withAUROC23\\text\{AUROC\}\_\{23\}staying above0\.890\.89at≥20\{\\geq\}\\,20dB SNR and30%30\\%temporal dropout\. Performance degrades gracefully inβ\\beta, indicating reduced transfer strength before loss of discriminability\. Full results appear in Appendix[I](https://arxiv.org/html/2605.30592#A9)\.

## 5Discussion

##### A fixed score coordinate enables transferable predictability estimates\.

Replacing the anchor term with a margin loss leaves in\-distribution discrimination nearly unchanged but increases anchor drift and reduces transferβ\\betaby47%47\\%\(Table[3](https://arxiv.org/html/2605.30592#S4.T3)\)\. Source\-domain separability alone, therefore, does not ensure that a scalar score retains a stable meaning on unseen systems\. Gauge fixing instead learns a shared ordinal coordinate, enabling predictability comparisons that are transferred across heterogeneous dynamics\.

##### Transfer concentrates where predictability is structurally limited\.

Zero\-shot transfer is strongest at the stochastic boundary \(L2→L3L\_\{2\}\\to L\_\{3\}andL3→L4L\_\{3\}\\to L\_\{4\}\); deterministic\-side boundaries remain weaker\. The same asymmetry holds under perturbation:AUROC23\\text\{AUROC\}\_\{23\}is more robust thanAUROC01\\text\{AUROC\}\_\{01\}to noise and dropout\. GON captures geometric structure disrupted by surrogate generation, whereas finer distinctions within deterministic chaos are more system\-specific\.

##### Geometric representations support cross\-system generalization\.

The 2\-jet representation improves transfer by exposing local trajectory geometry\. The benefit depends on coherent relationships among position, velocity, and acceleration; disrupting this alignment reduces the transfer even at a matched dimensionality\. This supports the view that transferable predictability cues arise from shared geometric structure rather than system\-specific statistics\.

##### Pretraining helps, but transfer depth is system\-dependent\.

Systems whose vector fields resemble those seen in training adapt immediately or are solved at zero\-shot\. In contrast, systems with unseen structure \(e\.g\., sinusoidal forcing in Thomas and the forced pendulum\) require more data and may fail to resolve finer boundaries\. Performance degrades smoothly with geometric distance from the source distribution\.

##### Scope, limitations, and broader implications\.

The benchmark covers 12 source and 5 target systems, all low\-dimensional and dissipative; extensions to higher\-dimensional systems, conservative dynamics, or real\-world signals, remain unestablished\. This work takes one of the earliest steps toward a general theory of transferable predictability estimation, establishing gauge\-fixing as a necessary condition for cross\-system ordinal comparability\. Predictive structure in real systems is often limited by chaos, stochastic forcing, or their interaction, arising across weather and climate dynamics, turbulent flows, biological oscillations, financial processes, and engineered systems\. By learning a transferable ordinal predictability coordinate, the framework enables cross\-system comparison without system\-specific calibration, informing model selection, early\-warning diagnostics, and the design of resilient coupled human–natural systems\. The method operates exclusively on synthetic ODE signals, involves no personal data, and has no foreseeable negative societal impacts\.

## 6Conclusion

This paper identifies gauge\-fixing as the missing ingredient for transferable ordinal predictability scores\. The central challenge is not ranking regimes within a single system, but assigning scores whose numerical values retain a consistent interpretation across systems\. The loss ablation makes the point cleanly: replacing the anchor with a margin objective leaves the source\-domain discrimination nearly unchanged while reducing transferβ\\betaby47%47\\%\. Ordinal coherence is robust to 4\-bit quantization and Gaussian noise down to 20 dB SNR, confirming the learned coordinate survives common observational degradations\. Pairwise discrimination and globally coherent ordinal scoring are distinct properties that can dissociate, and a stable coordinate is necessary to bridge them under transfer\. The benchmark is limited to low\-dimensional dissipative systems; this bound does not invalidate the core finding, since the ablation isolates coordinate convention as the operative variable within a fixed system class\. Extension to higher\-dimensional attractors, partial observability, and real\-world signals in weather, climate, and physiology remain the most consequential direction for future work\.

## References

- Analysis of ordinal categorical data\.John Wiley & Sons\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px4.p1.1)\.
- C\. Bandt and B\. Pompe \(2002\)Permutation entropy: a natural complexity measure for time series\.Physical Review Letters88\(17\),pp\. 174102\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- T\. M\. Bury, C\. T\. Bauch, and M\. Anand \(2020\)Detecting and distinguishing tipping points using spectral early warning signals\.Journal of the Royal Society Interface17\(170\)\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px3.p1.1)\.
- G\. Cai and J\. Huang \(2007\)A new finance chaotic attractor\.International Journal of Nonlinear Science3\(3\),pp\. 213–220\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx4)\.
- W\. Cao, V\. Mirjalili, and S\. Raschka \(2020\)Rank consistent ordinal regression for neural networks with application to age estimation\.Pattern Recognition Letters140,pp\. 325–331\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px4.p1.1),[Appendix G](https://arxiv.org/html/2605.30592#A7.p2.1)\.
- G\. Chen and T\. Ueta \(1999\)Yet another chaotic attractor\.International Journal of Bifurcation and chaos9\(07\),pp\. 1465–1466\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx1)\.
- R\. T\. Chen, Y\. Rubanova, J\. Bettencourt, and D\. K\. Duvenaud \(2018\)Neural ordinary differential equations\.Advances in neural information processing systems31\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px5.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- L\. Chua, M\. Komuro, and T\. Matsumoto \(1986\)The double scroll family\.IEEE transactions on circuits and systems33\(11\),pp\. 1072–1118\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx2)\.
- M\. Cranmer, S\. Greydanus, S\. Hoyer, P\. Battaglia, D\. Spergel, and S\. Ho \(2020\)Lagrangian neural networks\.arXiv preprint arXiv:2003\.04630\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px5.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p5.1)\.
- D\. d’Humieres, M\. Beasley, B\. Huberman, and A\. Libchaber \(1982\)Chaotic states and routes to chaos in the forced pendulum\.Physical Review A26\(6\),pp\. 3483\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx2.SSSx1)\.
- G\. Duffing \(1918\)Erzwungene schwingungen bei veränderlicher eigenfrequenz und ihre technische bedeutung\.Sammlung Vieweg,Vieweg\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx3)\.
- C\. Finn, P\. Abbeel, and S\. Levine \(2017\)Model\-agnostic meta\-learning for fast adaptation of deep networks\.InInternational conference on machine learning,pp\. 1126–1135\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px6.p1.1)\.
- R\. Genesio and A\. Tesi \(1992\)Harmonic balance methods for the analysis of chaotic dynamics in nonlinear systems\.Automatica28\(3\),pp\. 531–548\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx5)\.
- A\. Hastings and T\. Powell \(1991\)Chaos in a three\-species food chain\.Ecology72\(3\),pp\. 896–903\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx7)\.
- D\. T\. Kaplan and L\. Glass \(1992\)Direct test for determinism in a time series\.Physical Review Letters68\(4\),pp\. 427–430\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px1.p1.1)\.
- C\. Keylock \(2006\)Constrained surrogate time series with preservation of the mean and variance structure\.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics73\(3\),pp\. 036707\.Cited by:[§D\.2](https://arxiv.org/html/2605.30592#A4.SS2.p1.9)\.
- S\. Kim, D\. Kim, M\. Cho, and S\. Kwak \(2020\)Proxy anchor loss for deep metric learning\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 3238–3247\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px4.p1.1)\.
- M\. Kirchmeyer, Y\. Yin, J\. Donà, N\. Baskiotis, A\. Rakotomamonjy, and P\. Gallinari \(2022\)Generalizing to new physical systems via context\-informed dynamics model\.InInternational conference on machine learning,pp\. 11283–11301\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px6.p1.1)\.
- J\. J\. Koenderink and A\. J\. van Doorn \(1987\)Representation of local geometry in the visual system\.Biological cybernetics55\(6\),pp\. 367–375\.Cited by:[§1](https://arxiv.org/html/2605.30592#S1.p5.1)\.
- J\. Lai, A\. Bao, and W\. Gilpin \(2025\)Panda: a pretrained forecast model for chaotic dynamics\.arXiv preprint arXiv:2505\.13755\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px6.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- R\.B\. Leipnik and T\.A\. Newton \(1981\)Double strange attractors in rigid body motion with linear feedback control\.Physics Letters A86\(2\),pp\. 63–67\.External Links:ISSN 0375\-9601,[Link](http://dx.doi.org/10.1016/0375-9601(81)90165-1),[Document](https://dx.doi.org/10.1016/0375-9601%2881%2990165-1)Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx2.SSSx2)\.
- Z\. Li, N\. Kovachki, K\. Azizzadenesheli, B\. Liu, K\. Bhattacharya, A\. Stuart, and A\. Anandkumar \(2020\)Fourier neural operator for parametric partial differential equations\.arXiv preprint arXiv:2010\.08895\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px6.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- E\. N\. Lorenz \(1963\)Deterministic nonperiodic flow\.Journal of the Atmospheric Sciences20\(2\),pp\. 130–141\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx8)\.
- E\. N\. Lorenz \(1984\)Irregularity: a fundamental property of the atmosphere\.Tellus A36\(2\),pp\. 98–110\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx9)\.
- L\. Lu, P\. Jin, G\. Pang, Z\. Zhang, and G\. E\. Karniadakis \(2021\)Learning nonlinear operators via deeponet based on the universal approximation theorem of operators\.Nature machine intelligence3\(3\),pp\. 218–229\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px6.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- P\. McCullagh \(1980\)Regression models for ordinal data\.Journal of the Royal Statistical Society: Series B \(Methodological\)42\(2\),pp\. 109–127\.Cited by:[Appendix E](https://arxiv.org/html/2605.30592#A5.p1.3),[§1](https://arxiv.org/html/2605.30592#S1.p4.1)\.
- J\. Pathak, B\. Hunt, M\. Girvan, Z\. Lu, and E\. Ott \(2018\)Model\-free prediction of large spatiotemporally chaotic systems from data: a reservoir computing approach\.Physical review letters120\(2\),pp\. 024102\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px5.p1.1)\.
- M\. I\. Rabinovich and A\. L\. Fabrikant \(1979\)Stochastic self\-modulation of waves in nonequilibrium media\.J\. Exp\. Theor\. Phys77,pp\. 617–629\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx2.SSSx3)\.
- J\. S\. Richman and J\. R\. Moorman \(2000\)Physiological time\-series analysis using approximate entropy and sample entropy\.American Journal of Physiology–Heart and Circulatory Physiology278\(6\),pp\. H2039–H2049\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- M\. T\. Rosenstein, J\. J\. Collins, and C\. J\. De Luca \(1993\)A practical method for calculating largest lyapunov exponents from small data sets\.Physica D: Nonlinear Phenomena65\(1\-2\),pp\. 117–134\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px2.p1.1),[§D\.2](https://arxiv.org/html/2605.30592#A4.SS2.p1.9),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- O\. E\. Rössler \(1976\)An equation for continuous chaos\.Physics Letters A57\(5\),pp\. 397–398\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx10)\.
- A\. M\. Rucklidge \(1992\)Chaos in models of double convection\.Journal of Fluid Mechanics237,pp\. 209–229\.External Links:ISSN 1469\-7645,[Link](http://dx.doi.org/10.1017/s0022112092003392),[Document](https://dx.doi.org/10.1017/s0022112092003392)Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx11)\.
- M\. Scheffer, J\. Bascompte, W\. A\. Brock, V\. Brovkin, S\. R\. Carpenter, V\. Dakos, H\. Höhne, E\. H\. van Nes, M\. Rietkerk, and G\. Sugihara \(2009\)Early\-warning signals for critical transitions\.Nature461\(7260\),pp\. 53–59\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px3.p1.1)\.
- T\. Schreiber and A\. Schmitz \(1996\)Improved surrogate data for nonlinearity tests\.Physical review letters77\(4\),pp\. 635\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px2.p1.1),[§D\.2](https://arxiv.org/html/2605.30592#A4.SS2.p1.9),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- T\. Schreiber and A\. Schmitz \(2000\)Surrogate time series\.Physica D: Nonlinear Phenomena142\(3\-4\),pp\. 346–382\.Cited by:[§D\.2](https://arxiv.org/html/2605.30592#A4.SS2.p1.9)\.
- T\. Shimizu and N\. Morioka \(1980\)On the bifurcation of a symmetric limit cycle to an asymmetric one in a simple model\.Physics Letters A76\(3\-4\),pp\. 201–204\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx12)\.
- J\. C\. Sprott \(2010\)Elegant chaos: algebraically simple chaotic flows\.World Scientific,Singapore\.Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx1.SSSx6)\.
- G\. Sugihara and R\. M\. May \(1990\)Nonlinear forecasting as a way of distinguishing chaos from measurement error in time series\.Nature344,pp\. 734–741\.External Links:[Document](https://dx.doi.org/10.1038/344734a0)Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px1.p1.1)\.
- J\. Theiler, S\. Eubank, A\. Longtin, B\. Galdrikian, and J\. D\. Farmer \(1992\)Testing for nonlinearity in time series: the method of surrogate data\.Physica D: Nonlinear Phenomena58\(1\-4\),pp\. 77–94\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px2.p1.1),[§D\.2](https://arxiv.org/html/2605.30592#A4.SS2.p1.9),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- R\. Thomas \(1999\)Deterministic chaos seen in terms of feedback circuits: analysis, synthesis, “Labyrinth chaos”\.International Journal of Bifurcation and Chaos9\(10\),pp\. 1889–1905\.External Links:[Document](https://dx.doi.org/10.1142/S0218127499001383)Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx2.SSSx4)\.
- G\. Tigan and D\. Opriş \(2008\)Analysis of a 3d chaotic system\.Chaos, Solitons & Fractals36\(5\),pp\. 1315–1319\.External Links:ISSN 0960\-0779,[Link](http://dx.doi.org/10.1016/j.chaos.2006.07.052),[Document](https://dx.doi.org/10.1016/j.chaos.2006.07.052)Cited by:[Appendix F](https://arxiv.org/html/2605.30592#A6.SSx2.SSSx5)\.
- A\. Wolf, J\. B\. Swift, H\. L\. Swinney, and J\. A\. Vastano \(1985\)Determining lyapunov exponents from a time series\.Physica D: nonlinear phenomena16\(3\),pp\. 285–317\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2605.30592#S1.p2.1)\.
- K\. Zha, P\. Cao, J\. Son, Y\. Yang, and D\. Katabi \(2023\)Rank\-n\-contrast: learning continuous representations for regression\.Advances in Neural Information Processing Systems36,pp\. 17882–17903\.Cited by:[Appendix A](https://arxiv.org/html/2605.30592#A1.SS0.SSS0.Px4.p1.1)\.

## Appendix

## Appendix ARelated Work

##### Determinism versus stochasticity\.

Geometric diagnostics\[Sugihara and May,[1990](https://arxiv.org/html/2605.30592#bib.bib56), Kaplan and Glass,[1992](https://arxiv.org/html/2605.30592#bib.bib61)\]and prediction\-based tests study rejection of the stochastic null hypothesis within a single system\. Permutation entropy\[Bandt and Pompe,[2002](https://arxiv.org/html/2605.30592#bib.bib47)\]and sample entropy\[Richman and Moorman,[2000](https://arxiv.org/html/2605.30592#bib.bib48)\]provide scalar complexity measures widely used as binary indicators of dynamical complexity\. The present work places this binary separation within a broader ordinal hierarchy, where pairwise discriminability at a single boundary and global ordinal coherence across levels need not coincide\.

##### Chaos detection and surrogate testing\.

The classical toolbox for the binary problem is surrogate hypothesis testing\[Theileret al\.,[1992](https://arxiv.org/html/2605.30592#bib.bib10), Schreiber and Schmitz,[1996](https://arxiv.org/html/2605.30592#bib.bib11)\], which constructs a null distribution from the signal and tests whether its nonlinear statistics exceed those of matched stochastic alternatives\. Complementary approaches estimate the maximal Lyapunov exponent directly\[Wolfet al\.,[1985](https://arxiv.org/html/2605.30592#bib.bib8), Rosensteinet al\.,[1993](https://arxiv.org/html/2605.30592#bib.bib9)\]\. Both are designed for single\-system analysis: each new signal requires new surrogates or freshly tuned estimators, and both address a binary question\. The present work considers an ordinal formulation in which predictability is represented on a continuum, with scores intended to remain comparable across systems without per\-signal calibration\.

##### Early warning signals\.

Critical slowing down theory\[Schefferet al\.,[2009](https://arxiv.org/html/2605.30592#bib.bib58)\]and spectral precursor methods\[Buryet al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib29)\]study statistical signatures of approaching bifurcations\. The predictability ladder subsumes early\-warning detection as a special case:L1L\_\{1\}is precisely the transition regime where such signatures emerge\. The present benchmark does not directly evaluate early\-warning performance\. Transfer is strongest at the stochastic boundary rather than the deterministic\-side transition, and calibrated cross\-system early\-warning detection remains a natural extension\.

##### Ordinal learning\.

Cumulative link models\[Agresti,[2010](https://arxiv.org/html/2605.30592#bib.bib1)\]and neural extensions such as CORAL\[Caoet al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib2)\]enforce a consistent rank structure across ordered labels\. Ordinal contrastive methods\[Zhaet al\.,[2023](https://arxiv.org/html/2605.30592#bib.bib3)\]encode ordering in representation space\. Proxy\-based metric learning\[Kimet al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib99)\]uses class\-level anchor prototypes to stabilize embedding geometry\. These formulations study ordering within a single system or dataset, while the setting considered here involves scores that are intended to remain comparable across heterogeneous systems\.

##### Physics\-informed and geometric representations\.

Neural differential equation models\[Chenet al\.,[2018](https://arxiv.org/html/2605.30592#bib.bib19)\]and Lagrangian networks\[Cranmeret al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib20)\]encode geometric and physics\-aware inductive biases in learned representations\. Reservoir computing has been applied to chaotic forecasting\[Pathaket al\.,[2018](https://arxiv.org/html/2605.30592#bib.bib28)\]\. The 2\-jet representation used here is grounded in differential geometry: position, velocity, and acceleration jointly encode local trajectory curvature in a way that is invariant to system\-specific scaling and is disrupted by surrogate generation procedures\.

##### Transfer and meta\-learning for dynamical systems\.

Gradient\-based meta\-learning\[Finnet al\.,[2017](https://arxiv.org/html/2605.30592#bib.bib23)\]has been applied to dynamical systems forecasting\[Kirchmeyeret al\.,[2022](https://arxiv.org/html/2605.30592#bib.bib24)\]\. Neural operators\[Liet al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib25), Luet al\.,[2021](https://arxiv.org/html/2605.30592#bib.bib26)\]learn solution operators across parameter families\.Laiet al\.\[[2025](https://arxiv.org/html/2605.30592#bib.bib27)\]trained a transformer on a large synthetic corpus of chaotic systems and reported zero\-shot forecasting on unseen ODEs and PDEs\. These approaches study predictive generalization across systems under supervised forecasting objectives\. The setting considered here focuses on learning representations of predictability regimes, where the goal is to assign comparable ordinal scores rather than to predict future states\.

## Appendix BProof of Proposition[1](https://arxiv.org/html/2605.30592#Thmproposition1)

Letf:ℝ→ℝf:\\mathbb\{R\}\\to\\mathbb\{R\}be strictly increasing\. For any thresholdτm\\tau\_\{m\}and inputxx,

E​\(x\)\>τm⇔f​\(E​\(x\)\)\>f​\(τm\),E\(x\)\>\\tau\_\{m\}\\iff f\(E\(x\)\)\>f\(\\tau\_\{m\}\),since strictly increasing functions preserve the direction of all inequalities\. The induced ordinal predictorhE​\(x\)=∑m=03𝟏​\{E​\(x\)\>τm\}h\_\{E\}\(x\)=\\sum\_\{m=0\}^\{3\}\\mathbf\{1\}\\\{E\(x\)\>\\tau\_\{m\}\\\}depends only on these threshold comparisons\. Under the transformed scoreE~=f∘E\\tilde\{E\}=f\\circ Ewith transformed thresholdsτ~m=f​\(τm\)\\tilde\{\\tau\}\_\{m\}=f\(\\tau\_\{m\}\), each comparisonE~​\(x\)\>τ~m\\tilde\{E\}\(x\)\>\\tilde\{\\tau\}\_\{m\}is equivalent toE​\(x\)\>τmE\(x\)\>\\tau\_\{m\}\. Therefore,hE~​\(x\)=hE​\(x\)h\_\{\\tilde\{E\}\}\(x\)=h\_\{E\}\(x\)for allxx, and the induced ordinal prediction is unchanged\. ∎

## Appendix CArchitecture Details

### C\.1Temporal Convolutional Encoder

The encoder is a temporal convolutional network \(TCN\) composed ofB=6B=6residual blocks with exponentially increasing dilation factors\[1,2,4,8,16,32\]\[1,2,4,8,16,32\]\. Each block consists of two 1D dilated convolutions with kernel sizek=3k=3, followed by Group Normalization and GELU activation\. Residual connections are applied at the block level\.

The hidden dimension is fixed atH=128H=128across all layers\.

### C\.2Receptive Field

The effective receptive field of the network is

ℛ=1\+2​\(k−1\)​∑i=0B−12i=253,\\mathcal\{R\}=1\+2\(k\-1\)\\sum\_\{i=0\}^\{B\-1\}2^\{i\}=253,\(6\)which covers nearly the full temporal extent of the input after 2\-jet boundary trimming \(T′=T−2T^\{\\prime\}=T\-2\)\. This allows the model to capture long\-range temporal dependencies relevant for distinguishing predictability regimes\.

### C\.3Readout and Multi\-Scale Aggregation

The encoder produces a feature maph∈ℝH×T′h\\in\\mathbb\{R\}^\{H\\times T^\{\\prime\}\}\. RMS normalization is applied before aggregation\. Average pooling is performed at scaless∈\{4,16,64,256\}s\\in\\\{4,16,64,256\\\}to capture structure at multiple temporal resolutions; fors\>T′s\>T^\{\\prime\}the kernel is clamped tok=min⁡\(s,T′\)k=\\min\(s,T^\{\\prime\}\), so at scale256256this reduces to a global average over all254254timesteps\. Pooling uses no padding; non\-integer\-multiple tails are truncated byF\.avg\_pool1dwith defaultpadding=0\. The pooled representations are concatenated to form a feature vectorz∈ℝ4​Hz\\in\\mathbb\{R\}^\{4H\}\.

### C\.4MLP Head and Output Scaling

The pooled feature vectorz∈ℝ4​Hz\\in\\mathbb\{R\}^\{4H\}\(dimension512512\) is mapped to a scalar score via a two\-layer MLP with hidden dimensionH=128H=128:

Eθ​\(x\)=𝐖2​GELU​\(LN​\(𝐖1​z\+b1\)\)\+b2,E\_\{\\theta\}\(x\)=\\mathbf\{W\}\_\{2\}\\,\\mathrm\{GELU\}\\\!\\left\(\\mathrm\{LN\}\(\\mathbf\{W\}\_\{1\}z\+b\_\{1\}\)\\right\)\+b\_\{2\},\(7\)where𝐖1∈ℝ128×512\\mathbf\{W\}\_\{1\}\\in\\mathbb\{R\}^\{128\\times 512\}and𝐖2∈ℝ1×128\\mathbf\{W\}\_\{2\}\\in\\mathbb\{R\}^\{1\\times 128\}\. To prevent unbounded growth and maintain consistency with the fixed anchor targets, the output is smoothly clipped:

Eθ←c​tanh⁡\(Eθ/c\),c=5\.E\_\{\\theta\}\\leftarrow c\\tanh\(E\_\{\\theta\}/c\),\\quad c=5\.The full model contains616,233616\{,\}233trainable parameters\.

## Appendix DData Generation and Labeling Details

### D\.1Numerical Integration

Each system is simulated to produceN=4096N=4096uniformly sampled time steps\. The timestepΔ​t\\Delta tis selected to resolve the dominant dynamical timescale of each system\. We estimate a characteristic timescaleτ\\taufrom the mean inter\-peak interval of the trajectory, and setΔ​t\\Delta tsuch that approximately 64 dominant cycles are covered, withΔ​t\\Delta tclamped to\[10−4,10−1\]\[10^\{\-4\},10^\{\-1\}\]\.

Integration is performed using either a fixed\-step fourth\-order Runge–Kutta method or the adaptive Dormand–Prince \(DOPRI5\) solver, depending on system stiffness\.

### D\.2Predictability Ladder Construction

Deterministic regimes follow Definition[1](https://arxiv.org/html/2605.30592#Thmdefinition1), withλ1\\lambda\_\{1\}estimated viaRosensteinet al\.\[[1993](https://arxiv.org/html/2605.30592#bib.bib9)\]\.L0L\_\{0\}corresponds toλ1<0\\lambda\_\{1\}<0;L1L\_\{1\}andL2L\_\{2\}are distinguished by the relative magnitude ofλ1\\lambda\_\{1\}, cross\-checked by visual inspection of attractor phase portraits\. Because attractor geometries differ substantially across systems, no single numerical threshold is applied uniformly; regime boundaries are confirmed by qualitative separation in both the exponent estimate and the phase portrait\. These boundaries are fixed at data generation time and held constant throughout all experiments; the qualitative inspection step affects only this initial label assignment and does not enter any learned component of the method\.L3L\_\{3\}is generated fromL2L\_\{2\}trajectories using AAFT, IAAFT, WLS, and phase randomization\[Theileret al\.,[1992](https://arxiv.org/html/2605.30592#bib.bib10), Schreiber and Schmitz,[1996](https://arxiv.org/html/2605.30592#bib.bib11), Keylock,[2006](https://arxiv.org/html/2605.30592#bib.bib12), Schreiber and Schmitz,[2000](https://arxiv.org/html/2605.30592#bib.bib100)\]\.L4L\_\{4\}is i\.i\.d\. Gaussian noise matched in length and dimensionality\.

### D\.3Data Splits

Train/validation/test splits are performed at the trajectory level using an 80/10/10 partition\. All normalization statistics are computed on the training split and held fixed for validation and testing\.

## Appendix EGauge Fixing and Ordinal Identifiability

Ordinal regression determines only the ordering of latent scores\. For any strictly increasingff, the transformed scoreE~​\(x\)=f​\(E​\(x\)\)\\tilde\{E\}\(x\)=f\(E\(x\)\)induces identical ordinal decisions under the corresponding threshold transformation, soEEis identifiable only up to a monotone reparameterization\[McCullagh,[1980](https://arxiv.org/html/2605.30592#bib.bib98)\]\. Within a single system, this ambiguity is harmless: thresholds are learned jointly with the score, and ranking suffices\. Under cross\-system transfer, it becomes consequential because identical ordinal predictions on two systems need not assign the same numeric value to the same predictability regime\.

The anchor objective removes this degree of freedom by fixing the level\-wise score means,

𝔼​\[E​\(x\)∣y=k\]=tk,\\mathbb\{E\}\[E\(x\)\\mid y=k\]=t\_\{k\},standardizing the macroscopic location and scale of the score distribution, breaking the unbounded gauge freedom of ordinal regression\. The variance term further restricts within\-level dispersion\. Together, they establish a shared coordinate convention across systems without claiming strict uniqueness: infinitely many nonlinear strictly increasing functions could, in principle, satisfy these moment constraints, but the anchor fixes the operationally relevant degrees of freedom for cross\-system comparison\.

This differs from post\-hoc calibration, which rescales scores after training on a per\-system basis and, therefore, cannot produce a universal numeric coordinate\. Gauge fixing constrains the representation during learning, so the coordinate generalizes by construction\.

## Appendix FBenchmark System Equations and Parameters

For each system, we report the governing equations, fixed parameter values, the single parameter varied to traverse the predictability ladder, and the parameter range assigned to each regime\. Regime boundaries were determined by the protocol described in Appendix[D](https://arxiv.org/html/2605.30592#A4): the sign and relative magnitude of the leading Lyapunov exponentλ1\\lambda\_\{1\}cross\-checked by visual inspection of attractor phase portraits\. Initial conditions were sampled from multiple distributions per system \(Gaussian perturbations about a base point, uniform on a sphere, and system\-specific defaults\); all distributions produced qualitatively consistent regime behavior\.

The*stable*range corresponds to L0, the*transition*range to L1 \(weakly chaotic\), and the*chaotic*range to L2 \(strongly chaotic\)\.

Unless noted otherwise, initial conditions were sampled from three distributions: Gaussian perturbations about a base point, uniform on a sphere, and system\-specific defaults\. All distributions produced qualitatively consistent regime behavior\. Two exceptions are noted explicitly: Hastings–Powell, which requires ecologically structured non\-negative initial conditions, and Chua, which additionally uses double\-scroll sampling centered on the two attractor scrolls\.

### Source Systems \(Training Set\)

#### Chen\[Chen and Ueta,[1999](https://arxiv.org/html/2605.30592#bib.bib30)\]

x˙\\displaystyle\\dot\{x\}=a​\(y−x\)\\displaystyle=a\(y\-x\)\(8\)y˙\\displaystyle\\dot\{y\}=\(c−a\)​x−x​z\+c​y\\displaystyle=\(c\-a\)x\-xz\+cy\(9\)z˙\\displaystyle\\dot\{z\}=x​y−b​z\\displaystyle=xy\-bz\(10\)
Fixed parameters:a=35a=35,b=3b=3\. Varied parameter:cc\.

#### Chua\[Chuaet al\.,[1986](https://arxiv.org/html/2605.30592#bib.bib31)\]

x˙\\displaystyle\\dot\{x\}=α​\(y−x−f​\(x\)\)\\displaystyle=\\alpha\\bigl\(y\-x\-f\(x\)\\bigr\)\(11\)y˙\\displaystyle\\dot\{y\}=x−y\+z\\displaystyle=x\-y\+z\(12\)z˙\\displaystyle\\dot\{z\}=−β​y\\displaystyle=\-\\beta y\(13\)
where the piecewise\-linear Chua diode is

f​\(x\)=m1​x\+12​\(m0−m1\)​\(\|x\+1\|−\|x−1\|\)\.f\(x\)=m\_\{1\}x\+\\tfrac\{1\}\{2\}\(m\_\{0\}\-m\_\{1\}\)\\bigl\(\|x\+1\|\-\|x\-1\|\\bigr\)\.
Fixed parameters:β=28\.0\\beta=28\.0,m0=−1\.143m\_\{0\}=\-1\.143,m1=−0\.714m\_\{1\}=\-0\.714\. Varied parameter:α\\alpha\.

Chua additionally uses double\-scroll sampling \(σ=0\.3\\sigma=0\.3about\(±1\.5,0,0\)\(\\pm 1\.5,\\ 0,\\ 0\)\) and Gaussian perturbations \(σ=0\.01\\sigma=0\.01\) about the base point\(0\.7,0,0\)\(0\.7,\\ 0,\\ 0\)\.

#### Duffing\[Duffing,[1918](https://arxiv.org/html/2605.30592#bib.bib32)\]

The Duffing oscillator is non\-autonomous; the cosine forcing term renders the effective phase space three\-dimensional\.

x˙\\displaystyle\\dot\{x\}=y\\displaystyle=y\(14\)y˙\\displaystyle\\dot\{y\}=−δ​y−α​x−β​x3\+γ​cos⁡\(ω​t\)\\displaystyle=\-\\delta y\-\\alpha x\-\\beta x^\{3\}\+\\gamma\\cos\(\\omega t\)\(15\)
Fixed parameters:δ=0\.1\\delta=0\.1,α=−1\.0\\alpha=\-1\.0,β=1\.0\\beta=1\.0,ω=1\.0\\omega=1\.0\. Varied parameter:γ\\gamma\(forcing amplitude\)\.

#### Finance\[Cai and Huang,[2007](https://arxiv.org/html/2605.30592#bib.bib39)\]

x˙\\displaystyle\\dot\{x\}=\(1b−a\)​x\+z\+x​y\\displaystyle=\\left\(\\tfrac\{1\}\{b\}\-a\\right\)x\+z\+xy\(16\)y˙\\displaystyle\\dot\{y\}=−b​y−x2\\displaystyle=\-by\-x^\{2\}\(17\)z˙\\displaystyle\\dot\{z\}=−x−c​z\\displaystyle=\-x\-cz\(18\)
Fixed parameters:b=0\.2b=0\.2,c=1\.1c=1\.1\. Varied parameter:aa\.

#### Genesio–Tesi\[Genesio and Tesi,[1992](https://arxiv.org/html/2605.30592#bib.bib40)\]

x˙\\displaystyle\\dot\{x\}=y\\displaystyle=y\(19\)y˙\\displaystyle\\dot\{y\}=z\\displaystyle=z\(20\)z˙\\displaystyle\\dot\{z\}=−c​x−b​y−a​z\+x2\\displaystyle=\-cx\-by\-az\+x^\{2\}\(21\)
Fixed parameters:b=1\.1b=1\.1,c=1\.0c=1\.0\. Varied parameter:aa\.

#### Halvorsen\[Sprott,[2010](https://arxiv.org/html/2605.30592#bib.bib75)\]

x˙\\displaystyle\\dot\{x\}=−a​x−b​y−b​z−y2\\displaystyle=\-ax\-by\-bz\-y^\{2\}\(22\)y˙\\displaystyle\\dot\{y\}=−a​y−b​z−b​x−z2\\displaystyle=\-ay\-bz\-bx\-z^\{2\}\(23\)z˙\\displaystyle\\dot\{z\}=−a​z−b​x−b​y−x2\\displaystyle=\-az\-bx\-by\-x^\{2\}\(24\)
Fixed parameter:b=4\.0b=4\.0\. Varied parameter:aa\.

#### Hastings–Powell\[Hastings and Powell,[1991](https://arxiv.org/html/2605.30592#bib.bib33)\]

x˙\\displaystyle\\dot\{x\}=x​\(1−x\)−a1​x​y1\+b1​x\\displaystyle=x\(1\-x\)\-\\frac\{a\_\{1\}xy\}\{1\+b\_\{1\}x\}\(25\)y˙\\displaystyle\\dot\{y\}=a1​x​y1\+b1​x−a2​y​z1\+b2​y−d1​y\\displaystyle=\\frac\{a\_\{1\}xy\}\{1\+b\_\{1\}x\}\-\\frac\{a\_\{2\}yz\}\{1\+b\_\{2\}y\}\-d\_\{1\}y\(26\)z˙\\displaystyle\\dot\{z\}=a2​y​z1\+b2​y−d2​z\\displaystyle=\\frac\{a\_\{2\}yz\}\{1\+b\_\{2\}y\}\-d\_\{2\}z\(27\)
Fixed parameters:a2=0\.1a\_\{2\}=0\.1,b1=3\.0b\_\{1\}=3\.0,b2=2\.0b\_\{2\}=2\.0,d1=0\.4d\_\{1\}=0\.4,d2=0\.01d\_\{2\}=0\.01\. Varied parameter:a1a\_\{1\}\(prey–predator interaction rate\)\.

Initial conditions were absolute\-valued to ensure non\-negative populations, with ecological structured sampling \(prey∈\[0\.5,1\.0\]\\in\[0\.5,1\.0\], predator∈\[0\.1,0\.4\]\\in\[0\.1,0\.4\], superpredator∈\[0\.2,0\.8\]\\in\[0\.2,0\.8\]\) used in place of sphere sampling\. Integration used an adaptive timescale with an8×8\\timesecological multiplier applied to the estimated dominant periodτ\\tau, reflecting the longer characteristic timescales of trophic oscillations\.

#### Lorenz\-63\[Lorenz,[1963](https://arxiv.org/html/2605.30592#bib.bib34)\]

x˙\\displaystyle\\dot\{x\}=σ​\(y−x\)\\displaystyle=\\sigma\(y\-x\)\(28\)y˙\\displaystyle\\dot\{y\}=x​\(ρ−z\)−y\\displaystyle=x\(\\rho\-z\)\-y\(29\)z˙\\displaystyle\\dot\{z\}=x​y−β​z\\displaystyle=xy\-\\beta z\(30\)
Fixed parameters:σ=10\.0\\sigma=10\.0,β=8/3\\beta=8/3\. Varied parameter:ρ\\rho\.

#### Lorenz\-84\[Lorenz,[1984](https://arxiv.org/html/2605.30592#bib.bib41)\]

x˙\\displaystyle\\dot\{x\}=−a​x−y2−z2\+a​F\\displaystyle=\-ax\-y^\{2\}\-z^\{2\}\+aF\(31\)y˙\\displaystyle\\dot\{y\}=−y\+x​y−b​x​z\+G\\displaystyle=\-y\+xy\-bxz\+G\(32\)z˙\\displaystyle\\dot\{z\}=−z\+b​x​y\+x​z\\displaystyle=\-z\+bxy\+xz\(33\)
Fixed parameters:a=0\.25a=0\.25,b=4\.0b=4\.0,G=1\.0G=1\.0\. Varied parameter:FF\(forcing\)\.

#### Rössler\[Rössler,[1976](https://arxiv.org/html/2605.30592#bib.bib36)\]

x˙\\displaystyle\\dot\{x\}=−y−z\\displaystyle=\-y\-z\(34\)y˙\\displaystyle\\dot\{y\}=x\+a​y\\displaystyle=x\+ay\(35\)z˙\\displaystyle\\dot\{z\}=b\+z​\(x−c\)\\displaystyle=b\+z\(x\-c\)\(36\)
Fixed parameters:a=0\.2a=0\.2,b=0\.2b=0\.2\. Varied parameter:cc\.

#### Rucklidge\[Rucklidge,[1992](https://arxiv.org/html/2605.30592#bib.bib42)\]

x˙\\displaystyle\\dot\{x\}=−a​x\+b​y−y​z\\displaystyle=\-ax\+by\-yz\(37\)y˙\\displaystyle\\dot\{y\}=x\\displaystyle=x\(38\)z˙\\displaystyle\\dot\{z\}=−z\+y2\\displaystyle=\-z\+y^\{2\}\(39\)
Fixed parameter:a=2\.0a=2\.0\. Varied parameter:bb\.

#### Shimizu–Morioka\[Shimizu and Morioka,[1980](https://arxiv.org/html/2605.30592#bib.bib37)\]

x˙\\displaystyle\\dot\{x\}=y\\displaystyle=y\(40\)y˙\\displaystyle\\dot\{y\}=x−a​y−x​z\\displaystyle=x\-ay\-xz\(41\)z˙\\displaystyle\\dot\{z\}=−b​z\+x2\\displaystyle=\-bz\+x^\{2\}\(42\)
Fixed parameter:b=0\.5b=0\.5\. Varied parameter:aa\.

### Target Systems \(Held\-Out Set\)

#### Forced Pendulum\[d’Humiereset al\.,[1982](https://arxiv.org/html/2605.30592#bib.bib46)\]

Non\-autonomous; cosine forcing renders the effective phase space three\-dimensional\.

θ˙\\displaystyle\\dot\{\\theta\}=v\\displaystyle=v\(43\)v˙\\displaystyle\\dot\{v\}=−γ​v−sin⁡θ\+A​cos⁡\(ω​t\)\\displaystyle=\-\\gamma v\-\\sin\\theta\+A\\cos\(\\omega t\)\(44\)
Fixed parameters:γ=0\.2\\gamma=0\.2,ω=2/3\\omega=2/3\. Varied parameter:AA\(forcing amplitude\)\.

#### Newton–Leipnik\[Leipnik and Newton,[1981](https://arxiv.org/html/2605.30592#bib.bib45)\]

x˙\\displaystyle\\dot\{x\}=−a​x\+y\+10​y​z\\displaystyle=\-ax\+y\+10yz\(45\)y˙\\displaystyle\\dot\{y\}=−x−0\.4​y\+5​x​z\\displaystyle=\-x\-0\.4y\+5xz\(46\)z˙\\displaystyle\\dot\{z\}=b​z−5​x​y\\displaystyle=bz\-5xy\(47\)
Fixed parameter:a=0\.4a=0\.4\. Varied parameter:bb\(z\-axis damping\)\.

#### Rabinovich–Fabrikant\[Rabinovich and Fabrikant,[1979](https://arxiv.org/html/2605.30592#bib.bib74)\]

x˙\\displaystyle\\dot\{x\}=y​\(z−1\+x2\)\+γ​x\\displaystyle=y\(z\-1\+x^\{2\}\)\+\\gamma x\(48\)y˙\\displaystyle\\dot\{y\}=x​\(3​z\+1−x2\)\+γ​y\\displaystyle=x\(3z\+1\-x^\{2\}\)\+\\gamma y\(49\)z˙\\displaystyle\\dot\{z\}=−2​z​\(α\+x​y\)\\displaystyle=\-2z\(\\alpha\+xy\)\(50\)
Fixed parameter:α=0\.1\\alpha=0\.1\. Varied parameter:γ\\gamma\.

Trajectories exceeding‖s‖∞=25\\\|s\\\|\_\{\\infty\}=25were truncated to the bounded portion before divergence; trajectories with fewer than 300 retained points were discarded\.

#### Thomas\[Thomas,[1999](https://arxiv.org/html/2605.30592#bib.bib43)\]

x˙\\displaystyle\\dot\{x\}=sin⁡y−b​x\\displaystyle=\\sin y\-bx\(51\)y˙\\displaystyle\\dot\{y\}=sin⁡z−b​y\\displaystyle=\\sin z\-by\(52\)z˙\\displaystyle\\dot\{z\}=sin⁡x−b​z\\displaystyle=\\sin x\-bz\(53\)
No fixed parameters\. Varied parameter:bb\(dissipation\)\.

#### Tigan\[Tigan and Opriş,[2008](https://arxiv.org/html/2605.30592#bib.bib44)\]

x˙\\displaystyle\\dot\{x\}=a​\(y−x\)\\displaystyle=a\(y\-x\)\(54\)y˙\\displaystyle\\dot\{y\}=\(c−a\)​x−a​x​z\\displaystyle=\(c\-a\)x\-axz\(55\)z˙\\displaystyle\\dot\{z\}=−b​z\+x​y\\displaystyle=\-bz\+xy\(56\)
Fixed parameters:a=2\.1a=2\.1,b=0\.6b=0\.6\. Varied parameter:cc\.

## Appendix GNeural Baseline Comparison

We compare GON with three neural baselines trained on the same 12 source systems using identical data, preprocessing, encoders, and optimization protocols\. The only change is the training objective\.

Regressionminimizes mean squared error to integer labels\{0,1,2,3,4\}\\\{0,1,2,3,4\\\}\.Classificationminimizes cross\-entropy over five classes\.CORAL\[Caoet al\.,[2020](https://arxiv.org/html/2605.30592#bib.bib2)\]uses cumulative link constraints for ordinal supervision\.

Each method is converted to a scalar score for evaluation: the raw output for regression, the expected label∑kk​pk\\sum\_\{k\}k\\,p\_\{k\}for classification, and the cumulative\-logit score for CORAL\.

##### In\-distribution\.

Table[5](https://arxiv.org/html/2605.30592#A7.T5)reports test performance on the 12 source systems\. Regression, classification, CORAL, and GON all achieve near\-perfect adjacent\-pair discrimination in\-distribution\. To make cross\-method comparisons meaningful despite different score scales, we report the scale\-normalized monotonicity slopeβnorm\\beta\_\{\\text\{norm\}\}rather than rawβ\\beta\. Under this normalization, regression, classification, CORAL, and GON all show nearly perfect global ordering on the source distribution\. This pattern indicates that source\-domain performance alone is insufficient to reveal whether a learned score will transfer with a stable numerical meaning\.

Table 5:Neural baselines — in\-distribution performance on the 12 source systems \(test split\)\.
##### Zero\-shot transfer\.

Table[6](https://arxiv.org/html/2605.30592#A7.T6)reports zero\-shot performance on the 5 held\-out target systems\. All supervised baselines cluster at similar values ofβnorm≈0\.46\\beta\_\{\\text\{norm\}\}\\approx 0\.46–0\.480\.48despite their different objectives, suggesting a shared limitation: none learns a score with a stable cross\-system coordinate\.

GON achieves the highest scale\-normalized slope,βnorm=0\.559\\beta\_\{\\text\{norm\}\}=0\.559, while also remaining competitive on the strongest transferable boundaryL2→L3L\_\{2\}\\rightarrow L\_\{3\}\. This is the relevant comparison for cross\-system ordinal structure, since rawβ\\betais not directly comparable across methods with different score scales\.

Table 6:Neural baselines — zero\-shot transfer on the 5 held\-out systems \(macro\-averaged over per\-system values\)\.

## Appendix HPer\-System Adaptation Results

Tables[7](https://arxiv.org/html/2605.30592#A8.T7)–[16](https://arxiv.org/html/2605.30592#A8.T16)report monotonicity slopeβ\\betaand all four adjacent\-pair AUROC values for each of the five held\-out systems, acrossk∈\{5,10,…,100\}k\\in\\\{5,10,\\ldots,100\\\}labeled windows\. All entries are mean±\\pmstd over 5 random seeds\. Zero\-shot \(ZS\) performance is listed in the caption of each system block for reference\. Columns correspond to the three adaptation strategies:scratch\(random initialization\),pre\_head\(frozen encoder, readout fine\-tuned\), andpre\_all\(full model fine\-tuned from the pretrained checkpoint\)\.

##### Fine\-tuning protocol\.

All three adaptation strategies use AdamW with learning rate2×10−32\\times 10^\{\-3\}, weight decay10−410^\{\-4\}, cosine annealing over 30 epochs, and gradient clipping at 5\.0; identical to pretraining\. Data augmentation and EMA are disabled during fine\-tuning\. Forpre\_head, all encoder parameters are frozen including GroupNorm weight/bias; only the readout, normalization layer, and scale/bias embeddings are updated\. The trajectory\-level train/val/test split is fixed across all runs \(seed 42\)\. For each\(k,seed\)\(k,\\text\{seed\}\)pair,kkwindows are drawn without replacement from the fixed training pool, with at least one window per ladder level before filling the remaining budget uniformly at random\.

### Newton–Leipnik

Zero\-shot:β=0\.737\\beta=0\.737,AUROC01=0\.539\\text\{AUROC\}\_\{01\}=0\.539,AUROC12=0\.514\\text\{AUROC\}\_\{12\}=0\.514,AUROC23=0\.856\\text\{AUROC\}\_\{23\}=0\.856,AUROC34=1\.000\\text\{AUROC\}\_\{34\}=1\.000\.

Table 7:Newton–Leipnik: monotonicity slopeβ\\beta\(mean±\\pmstd, 5 seeds\)\.Table 8:Newton–Leipnik:AUROC01\\text\{AUROC\}\_\{01\},AUROC12\\text\{AUROC\}\_\{12\},AUROC23\\text\{AUROC\}\_\{23\},AUROC34\\text\{AUROC\}\_\{34\}\(mean±\\pmstd, 5 seeds\)\.
### Forced Pendulum

Zero\-shot:β=0\.489\\beta=0\.489,AUROC01=0\.295\\text\{AUROC\}\_\{01\}=0\.295,AUROC12=0\.390\\text\{AUROC\}\_\{12\}=0\.390,AUROC23=0\.604\\text\{AUROC\}\_\{23\}=0\.604,AUROC34=1\.000\\text\{AUROC\}\_\{34\}=1\.000\.

Table 9:Forced Pendulum: monotonicity slopeβ\\beta\(mean±\\pmstd, 5 seeds\)\.Table 10:Forced Pendulum:AUROC01\\text\{AUROC\}\_\{01\},AUROC12\\text\{AUROC\}\_\{12\},AUROC23\\text\{AUROC\}\_\{23\},AUROC34\\text\{AUROC\}\_\{34\}\(mean±\\pmstd, 5 seeds\)\. Note:AUROC23\\text\{AUROC\}\_\{23\}does not improve beyond zero\-shot \(0\.604\) for any method or anykk, indicating that theL2→L3L\_\{2\}\\\!\\to\\\!L\_\{3\}boundary is not resolved by window\-level adaptation alone\.
### Rabinovich–Fabrikant

Zero\-shot:β=1\.270\\beta=1\.270,AUROC01=0\.483\\text\{AUROC\}\_\{01\}=0\.483,AUROC12=0\.576\\text\{AUROC\}\_\{12\}=0\.576,AUROC23=0\.725\\text\{AUROC\}\_\{23\}=0\.725,AUROC34=0\.872\\text\{AUROC\}\_\{34\}=0\.872\.

Table 11:Rabinovich–Fabrikant: monotonicity slopeβ\\beta\(mean±\\pmstd, 5 seeds\)\.scratchfails to recover ordinal structure across the entirekkrange, never exceedingβ=0\.84\\beta=0\.84even atk=100k=100\.Table 12:Rabinovich–Fabrikant:AUROC01\\text\{AUROC\}\_\{01\},AUROC12\\text\{AUROC\}\_\{12\},AUROC23\\text\{AUROC\}\_\{23\},AUROC34\\text\{AUROC\}\_\{34\}\(mean±\\pmstd, 5 seeds\)\.
### Thomas

Zero\-shot:β=0\.958\\beta=0\.958,AUROC01=0\.508\\text\{AUROC\}\_\{01\}=0\.508,AUROC12=0\.452\\text\{AUROC\}\_\{12\}=0\.452,AUROC23=0\.391\\text\{AUROC\}\_\{23\}=0\.391,AUROC34=0\.995\\text\{AUROC\}\_\{34\}=0\.995\.

Table 13:Thomas: monotonicity slopeβ\\beta\(mean±\\pmstd, 5 seeds\)\.pre\_allbeats zero\-shot at mostkkvalues but with high variance;scratchbecomes competitive onβ\\betaabovek=40k=40while remaining belowpre\_allonAUROC23\\text\{AUROC\}\_\{23\}throughout\.Table 14:Thomas:AUROC01\\text\{AUROC\}\_\{01\},AUROC12\\text\{AUROC\}\_\{12\},AUROC23\\text\{AUROC\}\_\{23\},AUROC34\\text\{AUROC\}\_\{34\}\(mean±\\pmstd, 5 seeds\)\.pre\_allis the only method that consistently improvesAUROC23\\text\{AUROC\}\_\{23\}over zero\-shot \(0\.391\)\.pre\_headsystematically degradesAUROC23\\text\{AUROC\}\_\{23\}below zero\-shot across allkk\.
### Tigan

Zero\-shot:β=2\.105\\beta=2\.105,AUROC01=0\.645\\text\{AUROC\}\_\{01\}=0\.645,AUROC12=0\.985\\text\{AUROC\}\_\{12\}=0\.985,AUROC23=0\.997\\text\{AUROC\}\_\{23\}=0\.997,AUROC34=1\.000\\text\{AUROC\}\_\{34\}=1\.000\.

Table 15:Tigan: monotonicity slopeβ\\beta\(mean±\\pmstd, 5 seeds\)\. All pretrained methods match zero\-shot fromk=5k=5;pre\_headalone is sufficient here, confirming that the pretrained representation already covers Tigan’s attractor geometry\.scratchconverges slowly and remains below zero\-shot throughout the range\.Table 16:Tigan:AUROC01\\text\{AUROC\}\_\{01\},AUROC12\\text\{AUROC\}\_\{12\},AUROC23\\text\{AUROC\}\_\{23\},AUROC34\\text\{AUROC\}\_\{34\}\(mean±\\pmstd, 5 seeds\)\.AUROC12\\text\{AUROC\}\_\{12\},AUROC23\\text\{AUROC\}\_\{23\}, andAUROC34\\text\{AUROC\}\_\{34\}are at ceiling for both pretrained methods fromk=5k=5\. The only boundary where scratch lags meaningfully isAUROC01\\text\{AUROC\}\_\{01\}\.

## Appendix IFull Noise Robustness Results

Table[17](https://arxiv.org/html/2605.30592#A9.T17)reports the complete inference\-time perturbation sweep on the 12 source systems \(in\-distribution\)\. No retraining is performed; the pretrained GON checkpoint is applied directly\.

##### Perturbation details\.

All perturbations are applied at inference only undertorch\.no\_grad\(\)\.Gaussian noise: SNR is defined on a mean\-square power basis,SNRdB=10​log10⁡\(Ps/Pn\)\\mathrm\{SNR\}\_\{\\mathrm\{dB\}\}=10\\log\_\{10\}\(P\_\{s\}/P\_\{n\}\)wherePs=mean​\(x2\)P\_\{s\}=\\mathrm\{mean\}\(x^\{2\}\)over all channels and timesteps jointly per sample; a single noise power scalar is shared across channels\.Temporal dropout: selected timesteps are zero\-masked in\-place \(sequence length preserved\), with the same mask applied across all channels simultaneously; each sample draws an independent mask\.Quantization: post\-training uniform asymmetric quantization applied per channel per sample \(min\-max over the time axis\) in pure PyTorch; zero is a representable level \(uniform mid\-tread\)\.

##### Results\.

\(1\) Ordinal coherence is preserved down to 20 dB SNR and falls below1\.01\.0between 30 and 20 dB; at 5 dB,β=0\.216\\beta=0\.216, but the score remains positively monotone on average\. \(2\) Degradation is boundary\-asymmetric throughout:AUROC34\\text\{AUROC\}\_\{34\}holds at1\.0001\.000through 30 dB Gaussian noise and all dropout fractions up to 30%, whileAUROC01\\text\{AUROC\}\_\{01\}falls below0\.50\.5at 20 dB and at 20% dropout;AUROC12\\text\{AUROC\}\_\{12\}is intermediate, trackingAUROC01\\text\{AUROC\}\_\{01\}in shape but degrading more slowly\. \(3\) Quantization is negligible at all tested precisions: no metric changes by more than0\.0020\.002between 16\-bit and 4\-bit, andβ\\betais unchanged to three decimal places down to 6\-bit\.

Table 17:Full noise robustness sweep on the source systems \(in\-distribution\)\. No retraining; pretrained GON checkpoint applied directly under each condition\.