{"id":"1e5019ee-17ae-414d-b1dd-b47496baca9c","arxiv_id":"2610.19656","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PICES selects both the action chunk and execution horizon from perturbation-based sensitivity scores, cutting VLA policy calls by 25-60% without retraining.","lead":"This robotics paper proposes PICES, a training-free method that makes vision-language-action robot policies faster by generating several candidate action plans in one batch and using small perturbations of the random inputs to decide which plan to follow and for how long. On Jetson edge GPUs, simulation benchmarks, and a real robot arm, it reports 1.2-2.3x inference-time speedups with unchanged or slightly better task success.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The central assumption that stepwise initial-noise sensitivity G_k^(n) is a reliable proxy for action quality and replanning time is not established beyond aggregate outcomes; a direct per-candidate correlation test is needed.","rationale":"The reader's weakest_assumption is exactly the assumption I find most load-bearing: that G_k^(n), and the derived S_n and h_n, track action quality and the right replanning time. The paper's own limitation statement in Appendix D (sentence b349) confirms that no analytic guarantee is offered, so the burden falls on empirical validation. Figure 5 and the closed-loop tables are suggestive, but they do not establish the correlation: the open-loop evaluation averages over contexts instead of reporting per-candidate agreement, and the closed-loop results could in principle be explained by multi-candidate sampling and longer average horizons alone. I do not see an internal inconsistency in Eqs. 3-6, and the roofline argument and batch-latency measurements are reasonable independent support for the cost side of the claim. The W sensitivity and the small success-rate differences without confidence intervals are secondary concerns; they affect tuning and reporting but not the validity of the mechanism. Because the concern is a specific missing empirical link rather than a demonstrated contradiction, the appropriate disposition remains CONDITIONAL: the reader's verdict already requests justification, and my check specifies what would settle it. No verdict change is needed.","tokens_in":23657,"tokens_out":5876,"duration_ms":57862,"concrete_test":"Re-run the Appendix C open-loop protocol (1,000 Fractal contexts) while storing per-candidate records: S_n, the full profile {G_k^(n)}, and per-step reference RMSE for every candidate. Then compute (i) the Spearman correlation across candidates between S_n and prefix RMSE at h=W and at h=H; (ii) the rank correlation between the Eq. 6 horizon h_n and the first offset at which the candidate's prefix RMSE exceeds its own h=W error; and (iii) compare PICES's closed-loop success and policy calls on SimplerEnv against a control that samples N candidates and executes the longest horizon among a randomly chosen half of them. If (i) or (ii) is near zero or negative, or if the control matches Table 1, the sensitivity proxy is not the load-bearing component of PICES.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on a specific empirical premise: the normalized difference G_k^(n) in Eq. 3 is treated both as a quality signal (low initial-window average S_n marks trustworthy candidates via Eq. 5) and as a timing signal (the first offset k at which G_k^(n) exceeds S_n marks where execution should stop via Eq. 6). The paper explicitly acknowledges in Appendix D that this local-interpretation argument 'does not provide a bound on action prediction error' (sentence b349), so the proxy must be validated empirically. The open-loop study in Section 3.5 reports only aggregate curves (mean RMSE against mean selected horizon), not a per-candidate or per-step correlation between G_k^(n) and reference-action error. The closed-loop tables show aggregate success and policy-call improvements, but they do not separate the contribution of the sensitivity signal from the trivial benefits of sampling N candidates and executing longer horizons. The unresolved possibility is that PICES succeeds because multi-candidate sampling plus horizon regularization helps, while the specific sensitivity ranking is not the operative mechanism. That would make the method's central novelty rest on an untested correlation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes PICES, a training-free inference-time method to accelerate flow-matching vision-language-action (VLA) policies. At each policy call, PICES draws N nominal initial-noise samples and N locally perturbed counterparts, passes all 2N inputs through the action expert in one batched call with shared conditioning, and computes a stepwise sensitivity score G_k^(n) (Eq. 3) for each candidate. Candidates with initial-window sensitivity S_n above the median are discarded (Eqs. 4-5); each retained candidate is assigned an execution horizon h_n at the first step where G_k^(n) exceeds S_n (Eq. 6); the retained candidate with the longest horizon is executed. Experiments with GR00T N1.7 on SimplerEnv-Fractal, pi 0.5 and GR00T N1.5 on LIBERO, and pi 0.5 on a UR10e real-world platform report 25.4-60.3% fewer policy calls and 1.2-2.3x wall-clock policy-inference speedups on Jetson Orin and Thor relative to fixed-horizon baselines, with comparable or modestly higher success rates; the speedup also composes with FastV and GridS token pruning.","tokens_in":24016,"tokens_out":17784,"duration_ms":141115,"significance":"The headline speedup claim is credible as reported: the numbers are internally consistent across tables (e.g., Table 1: 18.6 calls times 242.7 ms versus 46.8 calls times 209.9 ms yields approximately 2.2x), the roofline analysis in Section 3.2 and Appendix B provides a concrete cost model for batched generation of 2N chunks, results replicate across three VLA backbones, two simulators, two edge devices, and a real robot, and the paper ships code with a detailed reproducibility statement. If the mechanism claim holds, PICES is genuinely useful: it is training-free, needs no auxiliary predictor, reduces policy calls rather than per-call FLOPs, and is orthogonal to token-pruning accelerators. The central risk is that the perturbation-sensitivity signal, which is the paper's distinguishing novelty, is validated only through aggregate outcomes; the ablations needed to establish that the sensitivity ranking rather than multi-candidate sampling plus horizon extension is the operative mechanism are not present. Additionally, the success-rate improvements are statistically unquantified and often within sampling noise, so the supported claim is success parity with an inference-time speedup.","major_comments":[{"comment":"PICES's central novelty is the claim that the stepwise sensitivity G_k^(n) is simultaneously a quality signal (low S_n marks trustworthy candidates via Eq. 5) and a horizon signal (the first crossing G_k^(n) > S_n sets h_n via Eq. 6). The authors correctly acknowledge in Appendix D (sentence b349) that this local-interpretation argument 'does not provide a bound on action prediction error,' so validity must come from the empirical evaluation. The current evidence is only aggregate: Figure 5 reports mean RMSE against mean selected horizon, with no per-context correlation between S_n and the reference-action RMSE of the corresponding candidate, and no per-step comparison between G_k^(n) and per-step error. Because the fixed-step curve in Figure 5 uses single-candidate chunks, the favorable RMSE of PICES at longer average horizons could reflect the benefit of drawing N=3 candidates rather than the sensitivity rule. Table 4 sharpens this concern: N=1, where the Eq. 5 filter is vacuous, already achieves 1.9-2.0x speedup, and N=3 adds only 0.2-0.3x, indicating that most of the gain comes from the horizon rule alone. I request (i) per-context correlation statistics between S_n and candidate prefix RMSE, and between the crossing offset and the offset at which per-step action error begins to rise, in the open-loop study; and (ii) closed-loop ablations that hold the N=3 pool and Eq. 6 fixed while replacing the Eq. 5 filter with random selection or with longest-horizon selection, at the same sampling budget. These controls are needed to establish that the perturbation signal, rather than multi-candidate sampling plus horizon extension, is the operative mechanism.","section":"Sections 3.3-3.5, Eqs. (3)-(6), Appendix D","authors_acknowledge":"This local interpretation does not provide a bound on action prediction error.","authors_acknowledge_p":0.77},{"comment":"The claim that PICES improves success rate while speeding up inference is not statistically supported as stated. In Table 3, the real-world average is 93.5% over 108 rollouts versus 90.7% for Fixed (h=8); with 36 trials per task, the per-task gaps (e.g., Tape: 88.9% versus 86.1%) are within binomial sampling error. In Table 1, PICES (68.7%) is statistically indistinguishable from Fixed (h=1) (68.5%), and in Table 2 the reported LIBERO improvements are 0.8 and 1.6 percentage points over 2,000 episodes. The success-rate claims should be accompanied by confidence intervals or a per-task count breakdown, and the contribution should be phrased as maintaining success while reducing inference time unless a configuration shows a measurable improvement.","section":"Tables 1-3, success-rate columns","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The method A3 is attributed to Zhao et al. (2026) in Section 1 and to Chen et al. (2026) in Appendix G; both papers appear in the reference list, so the intended source is ambiguous and should be fixed (the same passages also mix 'A3' and 'A 3' spellings).","section":"Section 1 and Appendix G","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The columns of Table 1 are misaligned in the current rendering: per-task success counts, average success, policy calls, and latency/speedup columns run together (visible in the Fixed (h=2) and AAC rows), making the table hard to check; please reformat.","section":"Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation for the W variants is inconsistent (PICES* with W=2 in Table 1 versus PICES** with W=8 in Table 2), and the sigma-sweep in Figure 9 reports only success-rate variation: policy-call counts and speedups at each sigma should also be reported, since sigma influences the Eq. 6 horizon estimates.","section":"Tables 1-2 and Figure 9","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 1's caption restricts 'mean policy calls' to successful episodes, while Tables 2-3 average over all episodes; the denominator should be stated consistently or both numbers reported, because early-terminating failures could bias the policy-call reduction figures.","section":"Table 1 caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Formatting and typos: the sentence defining A in Section 3.1 is missing a terminating period and has spacing issues ('R d x H To generate'), Section 4.2 contains 'achiving' (should be 'achieving'), and Section 1 contains 'exceuction' (should be 'execution').","section":"Section 3.1 and Section 4.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The real-world speedups (1.39x/1.36x) are computed as the product of average policy calls and per-call GPU latency rather than measured wall-clock episode time; if episode wall-clock time was recorded, it should be reported directly, since the per-call latency metric excludes observation, control, and environment overhead in the loop.","section":"Section 4.2, real-world results","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"PICES is a clean, honest, practically useful acceleration method, and the empirical work backs it up. The claim that holds: for flow-matching VLA policies you can cut policy calls by 25-60% and wall-clock inference time by 1.2-2.3x on Jetson Orin and Thor while keeping success roughly flat, and the paper shows this across three models, two simulators, and a real UR10e. That is a real result for edge deployment.\n\nWhat's new is the specific mechanism: joint chunk and horizon selection driven by stepwise sensitivity of each candidate to small perturbations of its initial noise. That combination does not appear in the cited adaptive-horizon baselines (AAC uses entropy, AutoHorizon uses attention, A3 uses an extra flow-verification pass, MoH uses cross-horizon agreement). The batching-cost analysis is also good diligence: they measure arithmetic intensity against hardware ridge points and justify the candidate batch size per device, so the speedup arithmetic is internally consistent. The ablations (N sweep, sigma sweep, combination with FastV and GridS) are the right ones, and Appendix D is honest that the local sensitivity story does not bound action error.\n\nSoft spots, in proportion. First, the sensitivity proxy is a heuristic, and the evidence for it is aggregate. The open-loop study reports mean RMSE versus mean selected horizon; it does not show a per-candidate correlation between G_k and reference-action error. The alternative reading, that a good part of the gain is just multi-candidate sampling plus longer horizons, is not fully excluded, though the N=1 result and the lowest-Sn ablation push back on it. A per-candidate correlation plot would settle this. I do not think it is fatal, because the method's value is the closed-loop outcome and the authors never claim the proxy is more than a proxy.\n\nSecond, W, the reference window, is a free parameter set to each model's default execution horizon, and the headline speedup is sensitive to it (W=2 gives 2.8x vs 2.2x in Table 1). The choice is defensible but not derived. Third, success-rate improvements over fixed baselines are small and reported without confidence intervals; the honest headline is \"success maintained while calls drop,\" and that is plenty. Fourth, minor but real: A3 is cited to Zhao et al. in the introduction and to Chen et al. in Appendix G; one of those is wrong.\n\nThis is for people doing VLA deployment on robots and anyone comparing adaptive-horizon methods. It deserves a serious referee. I would ask for the citation fix, a per-candidate sensitivity analysis if feasible, and confidence intervals or episode-level counts on success rates, but I would expect it to come out the other side intact.","headline":"A training-free, model-agnostic way to cut VLA policy calls by 25-60% and edge inference time up to 2.3x with success roughly maintained; the sensitivity mechanism is a heuristic validated only in aggregate, but the paper is honest and the evidence is solid.","tokens_in":25295,"tokens_out":841,"duration_ms":53595},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A frozen flow-matching robot policy can pick its action chunk and how long to run it from the chunk's sensitivity to small initial-noise perturbations, cutting policy calls by up to 60% and inference time by up to 2.3x.","keywords":["vision-language-action models","action chunking","execution horizon selection","flow matching","perturbation sensitivity","inference acceleration","robot manipulation","edge deployment"],"falsifier":"Using the paper's own open-loop protocol on GR00T N1.7 with the Fractal/RT-1 reference actions, compute for many contexts the sensitivity profiles $G_k^{(n)}$, the $S_n$ scores, and the true per-step prediction error against the recorded reference chunk. If the rank correlation between $S_n$ and prefix action error is not clearly negative, so that low-sensitivity chunks are not the accurate ones, then candidate filtering is selecting on noise rather than quality. Separately, test the horizon rule: check whether the first offset at which $G_k^{(n)}$ crosses above $S_n$ coincides, on average, with the offset at which the true prediction error begins to climb; if the crossing point carries no information about the error profile's change point, then the horizon estimate is not detecting when replanning is needed.","tokens_in":25149,"feed_emoji":"🤖","tokens_out":15946,"duration_ms":114701,"temperature":0.7,"pith_summary":"PICES attacks a specific inefficiency: a vision-language-action model that predicts a chunk of future actions is called anew every few executed actions even though the rest of the predicted chunk is discarded. The paper's proposal is to spend one batched policy call generating several candidate chunks, each from a different initial noise sample together with a locally perturbed twin, and to use the difference between each candidate and its twin as a free online signal about that candidate's quality. Candidates whose early steps react strongly to the perturbation are discarded, and for each surviving candidate the first step at which sensitivity rises above the early-window average sets how many actions to execute before replanning; the surviving candidate with the longest horizon is executed. The paper reports 25.4-60.3% fewer policy calls and 1.2-2.3x wall-clock inference speedups on Jetson Orin and Thor across GR00T and pi-0 models in simulation and on a real UR10e arm, with success rates maintained or slightly improved. If right, this means a training-free acceleration of frozen VLA policies that stacks with per-call optimizations such as token pruning, because it reduces the number of calls rather than the cost of each call.","feed_headline":"Perturbing robot-policy noise cuts policy calls by up to 60%","feed_subtitle":"PICES picks the action chunk and how long to execute it in one batched call, keeping success rates up on edge GPUs.","key_machinery":"The carrier of the argument is the stepwise action-sensitivity score $G_k^{(n)}$ of Eq. 3: the squared distance between the action at offset $k$ of a nominal chunk and the corresponding action of its perturbed twin, normalized by the squared Frobenius norm of the perturbation displacement. For small perturbations the score approximates the squared norm of the directional derivative of the action-output map along the sampled noise direction, an interpretation the paper derives in Appendix D from a first-order Taylor expansion; the paper also notes there that this local interpretation carries no bound on action prediction error. Two derived quantities do the selection work: the initial-window average $S_n$ over the first $W$ steps, which serves as both a candidate-quality score (below-median candidates survive) and a per-candidate threshold, and the first-crossing rule $h_n = \\min\\{k \\geq W : G_k^{(n)} > S_n\\}$, capped at $H$, which converts the sensitivity profile into an execution horizon. The whole mechanism is training-free: no Jacobians are constructed and no extra evaluator network is trained; the only added cost is one batched forward pass over the $2N$ noise samples, with the vision-language conditioning computed once per call.","core_discovery":"The central claim is that, for flow-matching VLA action experts (models that transform initial noise into an action chunk by iterative denoising), local stability of an action chunk under small perturbations of its initial noise is a usable online proxy both for chunk quality and for when replanning is needed. Concretely, PICES draws $N$ independent initial-noise samples $Z^{(n)}$, builds a perturbed twin $Z^{(n)}+\\sigma\\epsilon^{(n)}$ for each, generates all $2N$ chunks in one batched forward pass under a single conditioning encoding, and computes the stepwise score $G_k^{(n)} = \\|\\tilde{a}_k^{(n)} - a_k^{(n)}\\|^2/(\\sigma^2\\|\\epsilon^{(n)}\\|_F^2)$ for every action offset. The average of $G$ over the first $W$ steps, $S_n$, filters candidates to the below-median half, and for each kept candidate the first offset $k \\geq W$ at which $G_k^{(n)}$ exceeds $S_n$ defines its execution horizon $h_n$, capped at the chunk length $H$; the kept candidate with the longest horizon is the one executed. An open-loop study with recorded reference actions shows that oracle selection among several candidates yields lower action error over an eight-step horizon than a single candidate at one step, and a roofline analysis of the action experts shows that batch sizes around 6 on Orin and 10 on Thor stay close to memory-bound operation, so the extra candidates cost little. The paper's empirical case is that this joint selection reduces policy calls by 25.4-60.3% and speeds up wall-clock policy inference by 1.2-2.3x while matching or exceeding baseline success rates.","pith_inferences":["My extension: the sensitivity-to-error link can be tested head-on. The paper already has an open-loop setup with recorded reference actions; computing the rank correlation between $S_n$ and prefix RMSE across those contexts would tell whether low-sensitivity chunks are genuinely the accurate ones, independent of any closed-loop result.","My extension: the threshold is each candidate's own early-window average, which is self-referential; a calibrated or task-phase-aware threshold could make the reactivity-versus-speed trade-off explicit and could be tuned against measured action error rather than only against rollout success.","My extension: because the probe touches only the action expert's noise input, the same machinery applies to any stochastic sequence generator with chunked execution, including diffusion and flow policies without a vision-language backbone and non-robot predictors where replanning cost matters.","My extension: since the policy-call reduction is the lever, PICES should compound with asynchronous or overlapped inference, where the batched candidate call is scheduled during execution of the previous chunk so the horizon savings appear as wall-clock time rather than only as GPU idle time."],"forward_implications":["Fixed-horizon execution is a cost the model itself can waive: the same frozen checkpoint, given a bounded candidate batch, can decide per call how far to run, and the measured result is fewer policy calls without losing success.","Cumulative inference time on edge hardware is governed by the number of policy calls, not just per-call latency: cutting calls by 25.4-60.3% yields 1.2-2.3x wall-clock speedups on Jetson Orin and Thor even though each PICES call is slower than a single-chunk call.","PICES composes with per-call accelerations: with pi-0.5 on LIBERO, PICES combined with visual token pruning (FastV) or grid sampling (GridS16) reaches cumulative speedups up to 1.8-1.9x and 2.5-2.8x respectively, all measured relative to the unpruned fixed-horizon baseline.","The sensitivity signal transfers across backbones and benchmarks: GR00T N1.7 on SimplerEnv-Fractal, pi-0.5 and GR00T N1.5 on LIBERO, and pi-0.5 on three real UR10e manipulation tasks all show the same pattern of reduced calls with maintained or improved success.","Candidate diversity and batching cost set a practical operating point: $N=3$ nominal candidates (six chunks per call) is where speedup peaks on Orin at 2.2x, since larger batches add compute-bound latency faster than they save calls."],"supporting_citations":[{"why":"GR00T N1: the flow-matching VLA action expert that carries the main SimplerEnv-Fractal results and the open-loop accuracy study.","marker":"(Bjorck et al., 2025)"},{"why":"pi-0: the flow-matching VLA action expert used for the LIBERO evaluations and the real-world UR10e manipulation tasks.","marker":"(Black et al., 2024)"},{"why":"RT-1 Fractal dataset: supplies the recorded reference actions against which open-loop prediction error is measured.","marker":"(Brohan et al., 2022)"},{"why":"Provides the paired nominal-versus-perturbed-output finite-difference probing that PICES adapts for sensitivity estimation.","marker":"(Odena et al., 2018)"},{"why":"Perceptual path length: perturbing a latent code to quantify output stability, the direct inspiration for perturbation-based quality assessment.","marker":"(Karras et al., 2020)"},{"why":"Establishes local sensitivity as a reliability-relevant quantity, the conceptual root of using perturbation response to select candidates.","marker":"(Rifai, 2011)"},{"why":"AAC: the adaptive-execution-horizon baseline that PICES must beat on both policy calls and success rate.","marker":"(Liang et al., 2026)"},{"why":"A3: multi-candidate selection followed by flow-matching verification; its extra verification latency is the cost PICES avoids.","marker":"(Zhao et al., 2026)"},{"why":"Arithmetic-intensity profiling showing VLA action experts run memory-bound at small batch sizes, justifying the batched-candidate budget.","marker":"(Jiang et al., 2026)"}],"fun_headline_variants":["Perturbed noise picks robot action chunks and replan times","One batched pass selects chunk and execution horizon for VLA","Noise stability in VLA cuts policy calls by up to 60%","PICES uses perturbation probes to cut replanning on edge GPUs","Action-chunk noise checks trim VLA policy calls while aiding success"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The load-bearing premise is that a chunk's stepwise sensitivity to small perturbations of the initial noise tracks how accurate its actions are and when replanning is needed: low early sensitivity marks a trustworthy chunk, and the first step where the sensitivity profile $G_k^{(n)}$ rises above the early-window average $S_n$ marks the right point to stop executing; the paper itself notes in Appendix D that this interpretation gives no bound on action prediction error, leaving the open-loop and closed-loop results as the only support for the premise.","fun_headline_variants_meta":{"raw":{"variants":["Perturbed noise picks robot action chunks and replan times","One batched pass selects chunk and execution horizon for VLA","Noise stability in VLA cuts policy calls by up to 60%","PICES uses perturbation probes to cut replanning on edge GPUs","Action-chunk noise checks trim VLA policy calls while aiding success"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000109,"raw_usage":{"cost":0.0001092546,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0001092546,"upstream_inference_prompt_cost":4.02766e-05,"upstream_inference_completions_cost":6.8978e-05},"total_tokens":1608,"prompt_tokens":1229,"completion_tokens":379,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":983,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":299}},"tokens_in":1229,"tokens_out":379,"cost_source":"openrouter_usage","duration_ms":5459,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":299,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-16T01:49:56.681748+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Using the paper's own open-loop protocol on GR00T N1.7 with the Fractal/RT-1 reference actions, compute for many contexts the sensitivity profiles $G_k^{(n)}$, the $S_n$ scores, and the true per-step prediction error against the recorded reference chunk. If the rank correlation between $S_n$ and prefix action error is not clearly negative, so that low-sensitivity chunks are not the accurate ones, then candidate filtering is selecting on noise rather than quality. Separately, test the horizon rule: check whether the first offset at which $G_k^{(n)}$ crosses above $S_n$ coincides, on average, with the offset at which the true prediction error begins to climb; if the crossing point carries no information about the error profile's change point, then the horizon estimate is not detecting when replanning is needed.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":2}