△ Mechanism confirmed, baseline not beaten 2026

Certified Temporal Budget for Neural Control

Implementation & benchmark of arXiv:2608.25228 — Compiling Spatial Certificates into Temporal Contracts for Latency-Aware Control

Usefulness7/10

Difficulty5/10

Novelty8/10

Dynamics Stability Inference speedup Accuracy Rl Rnn Ssm Inference speedup Control theory Dynamical systems Optimization Numerical analysis

Source paper: Compiling Spatial Certificates into Temporal Contracts for Latency-Aware Control arXiv:2608.25228 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Attach a learned controller to a physical or simulated plant and use a continuous safety certificate to compute a conservative remaining-time budget before the current action or latent prediction can become unsafe. Compile this spatial margin into a unit-rate temporal contract, allowing asynchronous inference, batching, or early execution without online rollout integration; trigger a new network evaluation only when the countdown reaches a guard threshold.

Formulas

$$\tau^*(x;I)=\inf\{\tau\geq0\mid x_I(t_0+\tau;x)\notin\mathcal A_I^*\}.$$

$$\mathcal D_I=\{x\in\mathcal X:\Phi_I(x)\geq0\},\qquad \Phi_I(x)\geq0\Rightarrow x\in\mathcal A_I^*.$$

$$\frac{d}{dt}\Phi_I(x(t))\geq-\gamma_I\quad\Longrightarrow\quad c_I(t)=\Phi_I(x(t_0))-\gamma_I(t-t_0)\geq0,$$

$$r_I(t)=\frac{\Phi_I(x(t_0))}{\gamma_I},\qquad \text{permit latency }L\text{ only if }L\leq r_I(t)-\delta.$$

Mathematical statement

For object or constraint $I$, let $x(t)\in\mathcal X$ be the plant state and $\mathcal A_I^*$ its ground-truth admissible region. The exact physical safety horizon is $\tau^*(x;I)=\inf\{\tau\ge 0:x_I(t_0+\tau;x)\notin\mathcal A_I^*\}$. Replace this unavailable quantity by a continuous persistence functional $\Phi_I:\mathcal X\to\mathbb R$, with certified domain $\mathcal D_I=\{x:\Phi_I(x)\ge0\}$ and soundness implication $\Phi_I(x)\ge0\Rightarrow x\in\mathcal A_I^*$. For an implementation, estimate a worst-case certificate decrease rate $\gamma_I\ge0$ satisfying $d\Phi_I(x(t))/dt\ge-\gamma_I$ over the relevant state/action set. The compiled temporal contract is $c_I(t)=\Phi_I(x(t_0))-\gamma_I(t-t_0)$; if $c_I(t)\ge0$, the certificate remains nonnegative. More generally, normalize $r_I=\Phi_I/\gamma_I$ when $\gamma_I>0$, so $\dot r_I\ge-1$ and the scheduler has a unit-rate remaining-time budget. With bounded inference and handoff latency $L$, safety requires $L\le r_I(t)-\delta$, where $\delta>0$ is a measurement and modeling margin.

Implementation notes

1. Integration point: place the mechanism between state sensing and a learned policy or model-predictive neural controller. The network still outputs the action $u=\pi_\theta(x)$, but a lightweight certificate module evaluates $\Phi_I(x)$ for every active constraint and maintains a temporal contract for the current action. If the smallest remaining budget is below the worst-case inference plus communication latency, switch to a preverified fallback controller, reuse the previous action only for its certified interval, or force an urgent high-priority evaluation. 2. Pseudocode: at a refresh time $t_0$, evaluate $p_I=\Phi_I(x_{t_0})$, set $r_I=p_I/\gamma_I$, and set deadline $d=t_0+\min_I r_I-\delta$. Launch the neural inference asynchronously. At every scheduler tick, decrement the contract by elapsed time; accept the network action only if current time is below $d$, otherwise use the fallback and refresh the certificate. Reject immediately when any $p_I<0$. 3. Computed versus estimated: $\Phi_I$, state measurements, and countdown arithmetic are computed online; $\gamma_I$ is estimated offline by maximizing $-\nabla_x\Phi_I(x)\cdot f(x,\pi_\theta(x))$ over a sampled state/action envelope, then multiplied by a safety factor. Latency $L$ is measured from hardware traces, including queue and handoff delay. 4. First cheap experiment: use a 2-D point-mass obstacle-avoidance environment with a small MLP policy, compare ordinary periodic inference, naive event-triggered inference, and this contract scheduler under injected random latency. Sweep latency and safety-factor values. The quantitative prediction is a sharp failure boundary near $L=\min_I\Phi_I/\gamma_I$: below the bound, certificate violations should remain zero in simulation; above it, violations should increase, while average network evaluations should decrease approximately in proportion to the certified horizon. Validate the estimated boundary against dense numerical integration within 20 percent.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Inconclusive

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 certified countdown scheduler with fallback gating and a reproducible toy verification. The derivative and contract residual errors were below 1.5e-13; conservative gamma produced zero unsafe permissions, the latency/gamma transition matched prediction within 0.01, and evaluation scaling matched the inverse-budget prediction within 0.8%. Certified gating reduced unsafe handoffs from 28.8% periodic and 5.77% naive-event to 0%, while reducing policy launches to 79.2% of episodes.

Agent confidence: 8/10

Baseline: Periodic: 28.8% unsafe handoffs and 1.00 evaluations per episode; naive event trigger: 5.77% unsafe handoffs.
Idea: Certified scheduler: 0% unsafe handoffs and 0.792 policy evaluations per episode; predicted gamma transition 0.909, observed transition 0.90-0.91; predicted versus observed evaluation ratio for gamma 2 versus 1: 2.25 versus 2.232.

Limitations: This is a deterministic 1-D constant-velocity reduction rather than the planned 2-D learned MLP controller and physical plant. The certificate and gamma bound are assumed rather than learned or formally validated; latency is sampled rather than measured asynchronously; fallback performance, perception noise, model mismatch, and end-to-end control quality were not tested.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.25228) Sep 2, 2026

Ahead of the baseline (−7.5%) but not statistically significant (p=0.5454, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0022
Idea mean: 0.002
Effect (Δ): -0.0002 (−7.5%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.5454 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±27.8%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 6 configs

Limitations:

The registered dynamics benchmark is supervised future-angle regression rather than closed-loop control, so asynchronous inference, latency-triggered refresh, fallback controller quality, and end-to-end action safety were not measured. The implementation uses a hand-designed certificate penalty during training rather than a true online countdown scheduler, and the certificate rate envelope was not learned or formally certified. The certificate penalty was apparently inactive in this setup, as its three sweep settings produced identical four-seed means.

How to run: python3 bench_experiment.py

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

Stage 2 — Benchmark Sep 2, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented the certified temporal-budget intervention on the structurally matched registered dynamics track using identical rnn_small systems and the benchmark's paired-seed protocol. The idea had lower mean MSE than the tuned baseline, but the paired permutation p-value was 0.54535, so it did not meet the required significant-win criterion. The behavior-based signature showed zero permitted certificate violations, but the predicted certificate decrease was strongly conservative and did not translate into a significant task-metric improvement.

Baseline: Best baseline lr=0.006, weight_decay=0.0001; full 8-seed MSE 0.0022117759071988985 ± 0.0004126881098400608.
Idea: Best idea lr=0.006, weight_decay=0.0001, cert_weight=0.01; full 8-seed MSE 0.0020463809778448194 ± 0.0006329416993684799; paired delta=-0.00016539492935407907; permutation p=0.54535; 4/8 paired wins.

Benchmark result