△ Mechanism confirmed, baseline not beaten 2026

Lipschitz-Inflated Conformal Trajectory Tube

Implementation & benchmark of arXiv:2608.29559 — Conformal Prediction Regions for Continuous-Time Trajectories under Random Sampling

Usefulness8/10

Difficulty4/10

Novelty7/10

Sampling Stability Accuracy Generalization Ssm Rnn World model Probability Statistics Dynamical systems

Source paper: Conformal Prediction Regions for Continuous-Time Trajectories under Random Sampling arXiv:2608.29559 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Wrap a neural ODE, recurrent state-space model, or learned world model with a split-conformal prediction tube that is valid between irregularly sampled observations. Calibrate a pointwise residual quantile at observed times and inflate it at an unobserved time according to its distance from the nearest observed time and an estimated bound on the true and predicted trajectory slopes.

Formulas

$$L_i^{\mathrm{cont}}=\sup_{s,t\in[0,T^*],\,s\neq t}\frac{\|X_i(t)-X_i(s)\|}{|t-s|},\qquad \mathbb P\left(L_{m+1}^{\mathrm{cont}}\le \widehat L\right)\ge 1-\alpha_L.$$

$$S_i=\max_{j=1,\ldots,N_i}\left\|X_i(t_{ij})-\widehat X_i(t_{ij})\right\|,\qquad q_{1-\alpha}=\operatorname{Quantile}_{\lceil(m+1)(1-\alpha)\rceil}(S_1,\ldots,S_m).$$

$$\|X(t)-\widehat X(t)\|\le q_{1-\alpha}+\left(\widehat L+\widehat L_{\mathrm{pred}}\right)\delta(t),\qquad \delta(t)=\min_j|t-t_j|.$$

$$\mathcal C(t)=\left\{y\in\mathbb R^d:\|y-\widehat X(t)\|\le q_{1-\alpha}+\Gamma\delta(t)\right\},\qquad \Gamma=\widehat L+\widehat L_{\mathrm{pred}}.$$

Mathematical statement

Let X:[0,T*] -> R^d be the true trajectory, X_hat(t) the neural predictor, and T_i={t_i1,...,t_iNi} sampled times. Define L_i^cont=sup_{s != t} ||X_i(t)-X_i(s)||/|t-s| and let L_hat upper-bound the test trajectory constant with probability at least 1-alpha_L. For calibration trajectory i, define S_i=max_j ||X_i(t_ij)-X_hat_i(t_ij)|| and let q_(1-alpha) be its split-conformal quantile. If L_hat_pred bounds the predictor slope, the triangle inequality gives ||X(t)-X_hat(t)|| <= q_(1-alpha)+(L_hat+L_hat_pred) delta(t), where delta(t)=min_j |t-t_j|. The resulting tube has approximate failure probability at most alpha+alpha_L, subject to the paper's calibration split and coverage assumptions.

Implementation notes

Integrate this as an inference-time wrapper around a neural ODE, latent SSM, GRU, or world model. The base model receives irregular observations and produces a continuous prediction function X_hat(t); do not change its weights initially. Split trajectories into training and calibration sets. For every calibration trajectory, evaluate the model at its observed times, compute S_i=max_j ||X_i(t_ij)-X_hat_i(t_ij)||, and store the finite-sample conformal quantile q. Estimate L_hat from a separate high-frequency calibration set by computing the maximum finite-difference slope of the true trajectories, optionally applying a safety multiplier. Estimate L_hat_pred from dense solver evaluations of the neural predictor. At inference, for any requested time t, compute delta(t), the distance to the nearest observed input time, and return center X_hat(t) with radius q+(L_hat+L_hat_pred)delta(t). Pseudocode is: q=conformal_quantile(S); Gamma=L_hat+L_pred_hat; delta=min(abs(t-T_obs)); radius=q+Gamma*delta. The first cheap experiment is a GRU or neural ODE on noisy pendulum trajectories with randomly deleted observation times, comparing pointwise intervals against this tube. The quantitative prediction is dense-grid empirical coverage near 1-alpha-alpha_L, up to finite-sample slack, and approximately linear median radius growth with delta, with fitted slope near Gamma. Artificially using a sublinear gap inflation should cause coverage failures at the largest gaps.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a readable Lipschitz-inflated split-conformal trajectory-tube wrapper and numerical verifier. The triangle-inequality bound had zero measured violation, radius slope matched the predicted Gamma (0.9210526) to numerical precision, and the conditional coverage transition occurred exactly at multiplier 1.0; a 0.75 multiplier failed immediately after the largest observation gap. Marginal coverage at the estimated slope quantile was 0.988 versus the predicted slope-event coverage of 0.9, showing the mechanism but also conservatism from the grid/test construction.

Agent confidence: 8/10

Baseline: Constant observed-time conformal radius q=0; dense-grid coverage on the boundary test trajectory was 0.75025.
Idea: At Gamma=Lhat, boundary trajectory coverage was 1.000; iid test-slope marginal coverage was 0.98826, versus 1.000 with a 1.1 safety multiplier. Radius slope predicted/observed was 0.9210526/0.9210526; maximum triangle-bound violation was 0.

Limitations: This was a deterministic synthetic piecewise-linear toy study with a zero predictor, scalar trajectories, and no neural ODE/GRU training or noisy pendulum experiment. It did not test finite-sample behavior across many independent calibration splits, predictor slope estimation error, multidimensional norms, or irregular observation processes beyond the fixed sparse grid.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.29559) Sep 2, 2026

No significant difference vs. the tuned baseline (−-0.0%, p=1, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0091
Idea mean: 0.0091
Effect (Δ): 0 (−-0.0%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

The fixed dynamics track predicts one endpoint from an eight-step window rather than a true continuous-time rollout, so dense tube coverage is a diagnostic interpolation signature rather than the primary task metric. No multi-step neural ODE or independently sampled high-frequency calibration split was tested.

How to run: python3 bench_lipschitz_dynamics.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 10/10

The Lipschitz-inflated conformal tube was implemented and evaluated on the matched dynamics/rnn_small benchmark. Its trained-model mechanism signature was confirmed: the observed radius slope matched predicted Gamma to numerical precision, and diagnostic dense coverage was slightly higher than the constant-radius baseline. However, the tube is inference-time uncertainty post-processing and did not change the independent endpoint MSE: paired delta_mean=0.0 with permutation p=1.0, so there was no benchmark win.

Baseline: Best constant-radius baseline at lr=0.01: eight-seed MSE mean 0.009070940839592367, std 0.0043108428117023265.
Idea: Inflated-tube system at lr=0.01: eight-seed MSE mean 0.009070940839592367, std 0.0043108428117023265; paired delta 0.0, permutation p=1.0.

Benchmark result