Pith. sign in

REVIEW 2 major objections 4 minor 30 references

The error of a scenario-based optimal value follows a Gaussian-process infimum, and reduces to a normal distribution when the true optimum is unique.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

The scenario-based optimal value in uncertain optimal control converges at N^{-1/2} to a Gaussian limit, or to the lower envelope of a Gaussian process when multiple optima exist, enabling asymptotic confidence intervals.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely new stability estimate in the cumulative-input metric yields the SAA optimal-value CLT and confidence intervals for infinite-dimensional control-affine Mayer problems; the theory is sound, but the empirical claims rest on an unverified global reachable-set assumption. the 2 major comments →

arxiv 2607.14965 v1 pith:JUTHUZLE submitted 2026-07-16 math.OC math.STstat.TH

Statistical Inference for Scenario-Based Dynamic Optimization under Uncertainty

Aurya Javeed , Johannes Milz This is my paper
classification math.OC math.STstat.TH MSC 49K4560F1762F1290C15
keywords sample average approximationoptimal control under uncertaintyfunctional central limit theoremconfidence intervalsbatch process optimizationparametric uncertaintyoptimal value estimationsubsampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper studies the sample average approximation (SAA) method for finite-horizon open-loop optimal control problems with uncertain parameters. Its central claim is that the scaled optimal-value error N^{1/2}(J_N^* - J^*) converges in distribution to the infimum of a centered Gaussian process over the set of population-optimal controls. When the population problem has exactly one optimal control, that limit is normal, with variance equal to the variance of the terminal loss at the optimal control. The paper uses this limit to construct two asymptotically valid confidence intervals for the true optimal value: a cheap plug-in interval under uniqueness, and a subsampling interval that works even when multiple optimizers exist. Two fed-batch case studies show that the intervals quantify the sampling error, with the plug-in interval undercovering at small sample sizes and approaching nominal coverage as the scenario count grows.

Core claim

The principal result is that, under compactness and smoothness assumptions, the SAA optimal value converges at the standard Monte Carlo rate N^{-1/2}, with the exact limiting distribution equal to inf_{u in U*} Z(u), where Z is a centered Gaussian process on the control space with covariance Cov(Z(u), Z(v)) = Cov(F(x_u(t_f, ξ)), F(x_v(t_f, ξ))). If the population problem has a unique optimal control u*, the limit simplifies to N(0, Var(F(x_{u*}(t_f, ξ)))). The paper shows that this distributional statement directly justifies two confidence intervals for the population optimal value: one plug-in interval relying on the Gaussian limit under uniqueness, and one subsampling interval that does no

What carries the argument

The carrying object is the time-integrated control profile, U(u)(t) = ∫_0^t u(s) ds, and the associated distance d(u,v) = max_t |U(u)(t) - U(v)(t)|. The key stability estimate states that terminal states are Lipschitz continuous in this distance, uniformly over the uncertainty set. This estimate turns the sample objective into a stochastic process indexed by a compact metric space, enabling a functional central limit theorem for empirical processes. The same distance controls the convergence of the estimated variance in the plug-in interval and the consistency of the subsampling distribution.

Load-bearing premise

Everything rests on the assumption that a single compact set contains every state trajectory for every admissible control and every uncertain parameter (Assumption 2); without a verifiable bound like that, the Lipschitz stability estimate and both confidence interval guarantees have no foundation.

What would settle it

Construct a control-affine system that satisfies the smoothness and box-control assumptions but has finite-time blow-up for some bounded control, so the compact reachable set in Assumption 2 does not exist; if the √N Gaussian-process limit still appears empirically, the theorem's stated hypotheses are not sharp. Conversely, on a problem engineered with two exactly optimal controls, check whether the plug-in interval's coverage drops below nominal while the subsampling interval's coverage stays near nominal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The statistical error in an SAA optimal value is of order N^{-1/2}; quadrupling the number of scenarios halves the error, independent of the dimension of the uncertain parameter vector.
  • When the true problem has a unique optimal policy, the plug-in normal interval gives an asymptotically calibrated 1−β confidence interval around the SAA value at negligible extra computational cost.
  • When several optimal policies exist, the limiting distribution is the lower envelope of a Gaussian process, so normal intervals fail; the subsampling interval maintains asymptotic coverage without requiring uniqueness.
  • In the two fed-batch studies, the plug-in interval's finite-sample coverage is below nominal for N between 8 and 64 and improves as N grows, so small-N plug-in intervals should be treated as diagnostics rather than calibrated statements.
  • The theory covers open-loop policies with risk-neutral terminal costs; the paper identifies risk-averse objectives, state-path constraints, and feedback policies as open extensions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the compact reachable-set assumption cannot be verified for a given model — for example, kinetics that admit finite-time blow-up — the interval constructions have no theoretical grounding; a practitioner would need a reachability certificate before trusting either interval.
  • Because the stability estimate is in terms of cumulative input rather than the control value itself, the sampling uncertainty should be insensitive to high-frequency control chatter as long as the cumulative delivery is close; this is consistent with the paper's singular-control examples.
  • A direct next experiment is to engineer a problem with a flat objective plateau (two or more exactly optimal controls) and compare the two intervals: the subsampling interval should hold its coverage, while the plug-in interval should fail even at large N.
  • The same Gaussian-process machinery likely extends to risk measures such as conditional value-at-risk if a Lipschitz terminal cost is retained, giving the error distribution for risk-averse versions of the problem.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. This paper develops asymptotic statistical inference for sample average approximation (SAA) in finite-horizon open-loop optimal control under parametric uncertainty. The main theoretical result is Theorem 3.1: under assumptions of weak compactness of the control set, a uniform compact reachable set, smoothness of the dynamics, and Lipschitz terminal cost, the scaled SAA optimal-value error converges in distribution to the infimum of a centered Gaussian process over the population argmin set; when the argmin is unique, the limit is normal. The paper derives two confidence-interval procedures: a plug-in normal interval under uniqueness and a subsampling interval that does not require uniqueness. The proofs are based on a Volterra-operator stability estimate (Proposition A.1) and an empirical-process FCLT (Theorem B.2). Numerical case studies on fed-batch reactor and ethanol fermentation problems illustrate the confidence intervals, and the authors report honest finite-sample undercoverage of the plug-in interval.

Significance. If the results hold, this is a useful and nontrivial contribution. The key analytic idea is Proposition A.1: under control-affine dynamics and a compact reachable set, terminal states depend Lipschitz-continuously on the cumulative input profile, allowing the control set to be equipped with a compact metric without imposing norm compactness on the controls. This cleanly yields the FCLT and the limiting distribution of the optimal-value error. The distinction between the Gaussian plug-in regime and the non-Gaussian nonunique-argmin regime is practically relevant for scenario-based dynamic optimization. The paper ships reproducible code and archived data, and the numerical study is transparent about the fact that the plug-in interval undercovers at the tested sample sizes. The main gap is formal verification of the reachable-set assumption for the numerical models; the argmin step in the proof of Theorem 3.1 is also deferred to a citation and should be spelled out.

major comments (2)
  1. [§5.1–5.2, Assumption 2] Assumption 2 is stated but never verified for the two fed-batch models in §5. The uniformity constant C_{U,t_f} in Prop. A.1 is obtained from a common compact reachable set X containing x_u(t,ξ) for all u∈U, ξ∈Ξ, and t∈[0,t_f]; this same boundedness drives the entropy estimate and Theorem B.2. The case studies only list control bounds and the parametric uncertainty model; no invariant-set or boundedness argument is given for the state components. Since a Lipschitz/continuity assumption alone is not enough to guarantee such a uniform bound (the growth-rate terms are polynomial/exponential in the states), the numerical illustrations of Theorems 4.1–4.2 are not formally covered unless Assumption 2 is checked. Please add a short verification (e.g., monotonicity/invariance boxes using the sign of ξ_8, the feed concentration, and the terminal-volume bound) or explicitly downgrade the examples
  2. [Appendix B, proof of Theorem 3.1] The transition from Theorem B.2 to (3.1) is the sentence that the arguments from Shapiro et al. (2021), Theorem 5.7, apply verbatim. This is the load-bearing argmin step. Please state the theorem and verify its hypotheses in the present infinite-dimensional setting: compactness of K=(U,d) has been established, but the existence/measurability of SAA minimizers and the treatment of the nonunique argmin set U* should be made explicit. A short direct proof using the continuity of the infimum map and an argmin expansion would make the paper self-contained. As written, the central limit result inherits all of its difficulty from this citation.
minor comments (4)
  1. [§2, Assumption 2] The word 'convex' in Assumption 2 is not used anywhere in the proofs (compactness suffices). If convexity is intended for later use, say so; otherwise remove it.
  2. [§5.2] The terminal-state constraint x_4(t_f) ≤ 200 should be incorporated explicitly into the definition of the feasible set U for Example 5.2 (e.g., as an integral bound on u), since the theory's U is described in §2 by control bounds alone.
  3. [Notation] The space C(0,t_f;R^s) should be written as C([0,t_f];R^s), since all estimates take suprema over the closed interval, including t_f.
  4. [§5.1, Table 1] The reference value J*_{N_ref} is itself a random proxy for J*, so the reported coverage is coverage of J*_{N_ref} rather than of J*; a sentence noting this additional noise source would be useful.

Circularity Check

0 steps flagged

No significant circularity: Theorem 3.1 is derived from explicit Assumptions 1-4 via classical CLT and compactness arguments, not from fitted parameters or self-citations; the few self-citations are background/software and not load-bearing.

full rationale

Walking the derivation chain: Proposition A.1 proves a genuine Lipschitz stability estimate for terminal states in the primitive metric d(u,v), with the constant CU,tf obtained from Assumptions 2 and 3. Lemma B.1 derives compactness of (U,d) from weak compactness of U and compactness of the Volterra operator. Theorem B.2 applies the classical Jain-Marcus FCLT, verifying the entropy condition via the Birman-Solomjak Sobolev entropy estimate; the only inputs are Assumptions 1-4. Theorem 3.1 then follows by the standard argmin mapping argument from Shapiro, Dentcheva and Ruszczynski (2021), an external textbook result. The confidence-interval theorems are consequences of Theorem 3.1 plus consistency of the plug-in variance estimator and the Politis-Romano subsampling principle. No fitted parameter is renamed as a prediction: the numerical section estimates sigma^2 from terminal-loss samples at the computed SAA optimizer, which is the standard plug-in estimator, and validates coverage against an independent reference SAA value. The self-citations (EnsembleControl software, consistency results, sample-complexity bounds) are used only for reproducibility and background; they do not supply the limit theorem or the confidence-interval validity. The unverified reachability Assumption 2 in the case studies is a potential correctness gap, but it is not circular: the theory is explicitly conditional on that assumption. No step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 8 axioms · 0 invented entities

The theory is parameter-free: no constants are fitted to data; the only estimated quantity is σ², which is a standard plug-in variance. The assumptions are explicit regularity conditions on the control set, the reachable set, the dynamics, and the terminal loss. The numerical sections introduce user-selected values (r, q, b_N, m_N) that affect the illustrations but not the central theorem. No invented entities.

axioms (8)
  • domain assumption Assumption 1: U ⊂ L2 is nonempty, bounded, weakly closed
    Defines the admissible-control class; enables weak compactness and compactness of the primitive control space (Lemma B.1). Standard for box-constrained controls.
  • domain assumption Assumption 2: there is a compact convex X containing all trajectories x_u(t, ξ)
    Global reachable-set bound; supplies uniform boundedness of f0, f1, Df0, Df1 on X×Ξ used throughout Prop A.1 and the FCLT. Not automatically satisfied by nonlinear kinetics; this is the fragile premise.
  • domain assumption Assumption 3: Ξ compact, x0 measurable, f0, f1 C^1 in state on a neighborhood of X, continuous on X×Ξ
    Smoothness+compactness needed for the Gronwall argument in Prop A.1 and measurability of the C(K)-valued process.
  • domain assumption Assumption 4: terminal loss F Lipschitz on X
    Converts endpoint stability into objective stability (eq. B.1).
  • domain assumption Standard measurability conditions so expectations and sample averages are well-defined
    Assumed in Section 2; technical but required for the probability machinery.
  • standard math Jain–Marcus (1975) CLT for C(S)-valued i.i.d. random elements under entropy-integrability; Sobolev entropy estimate of Birman–Solomjak
    Used in Theorem B.2 to obtain the FCLT on the compact metric space (U,d).
  • standard math Argmin-value functional continuity / Theorem 5.7 of Shapiro et al. (2021)
    Passes from convergence of √N(Ĵ_N−J) in C(K) to the limiting optimal-value error inf_{u∈U*}Z(u) (Theorem 3.1); invoked 'verbatim' without a self-contained adaptation.
  • standard math Politis–Romano (1994) subsampling principle, Corollary 2.1, as applied in Eichhorn–Römisch (2007)
    Validates the subsampling CI in Theorem 4.2 under b_N→∞, b_N/N→0, m_N→∞.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Statistical Inference for Scenario-Based Dynamic Optimization under Uncertainty." pith.science (2026). https://pith.science/paper/JUTHUZLE

@misc{pith2026260714965,
  author       = {Pith},
  title        = {Pith review of: Statistical Inference for Scenario-Based Dynamic Optimization under Uncertainty},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUTHUZLE}},
  note         = {Machine review of arXiv:2607.14965}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Motivated by batch and semi-batch process operation, we study finite-horizon open-loop dynamic optimization problems with uncertain parameters. A common computational approach replaces the expected performance criterion by an average over finitely many sampled parameter realizations. We develop a statistical theory for the resulting sample-based optimal value as an estimator of the population optimal value. The analysis is based on a stability estimate showing that terminal losses depend Lipschitz continuously on the time-integrated control, which records the cumulative input delivered up to each time. This estimate yields a functional central limit theorem for the sample-based objective and a statistical limit theorem for the corresponding optimal value error. As a consequence, we obtain confidence intervals for the population optimal value. When the population optimizer is unique, the limit is Gaussian and leads to a plug-in confidence interval. When multiple optimal policies may exist, we use a subsampling confidence interval that does not require uniqueness. The methodology is illustrated on two fed-batch case studies in which feed-rate profiles are optimized under parametric uncertainty.

Figures

Figures reproduced from arXiv: 2607.14965 by Aurya Javeed, Johannes Milz.

Figure 1
Figure 1. Figure 1: For the fed-batch reactor example, postprocessed nominal and SAA optimal solutions. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: For the fed-batch reactor example, finite-sample behavior of the SAA optimal value. bb [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: For the fed-batch reactor example, confidence intervals for the population optimal value b [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: For the ethanol fermentation example, postprocessed nominal and SAA optimal solutions. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: For the ethanol fermentation example, finite-sample behavior of the SAA optimal value. bb [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: For the ethanol fermentation example, confidence intervals for the population optimal b [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 13 canonical work pages

  1. [3]

    Output feedback stochastic nonlinear model predictive control for batch processes. Comput. Chem. Eng. 126, 434–450. doi:10.1016/j.compchemeng.2019.04.021. da Cunha, A.S., Peixoto, F.C., Prata, D.M.,

    work page doi:10.1016/j.compchemeng.2019.04.021 2019

  2. [14]

    A review on robust M-estimators for regression analysis. Comput. Chem. Eng. 147, 107254. doi:10.1016/j.compchemeng.2021.107254. Milz, J.,

    arXiv 2021

  3. [16]

    Consistency of sample-based stationary points for infinite-dimensional stochastic optimization. SIAM J. Optim. 35, 42–61. doi:10.1137/23M1600608. Milz, J., Surowiec, T.M.,

    work page doi:10.1137/23m1600608

  4. [17]

    Asymptotic consistency for nonconvex risk-averse stochastic optimization with infinite dimensional decision spaces. Math. Oper. Res. 49, 1403–1418. doi:10.1287/moor.2022.0200. Nemirovski, A., Shapiro, A.,

    arXiv 2022

  5. [18]

    Convex approximations of chance constrained programs. SIAM J. Optim. 17, 969–996. doi:10.1137/050622328. Paulen, R., Fikar, M.,

    doi:10.1137/050622328

  6. [19]

    Dynamic real-time optimization of batch processes using Pontryagin’s minimum principle and set-membership adaptation. Comput. Chem. Eng. 128, 488–495. doi:10.1016/j.compchemeng.2019.06.027. Phelps, C., Royset, J., Gong, Q.,

    work page doi:10.1016/j.compchemeng.2019.06.027 2019

  7. [21]

    Asymptotic properties of Monte Carlo methods in elliptic PDE-constrained optimization under uncertainty. Numer. Math. 156, 1887–1914. doi:10.1007/s00211-024-01436-5. Royset, J.O.,

    work page doi:10.1007/s00211-024-01436-5 1914

  8. [24]

    MOS-SIAM Ser

    Lectures on Stochastic Program- ming: Modeling and Theory. MOS-SIAM Ser. Optim.. 3rd ed., SIAM, Philadel- phia, PA. doi:10.1137/1.9781611976595. Terwiesch, P., Agarwal, M.,

    doi:10.1137/1.9781611976595

  9. [25]

    Batch unit optimization with imperfect modelling: A survey. J. Process Contr. 4, 238–258. doi:10.1016/0959- 1524(94)80045-6. Terwiesch, P., Ravemark, D., Schenker, B., Rippin, D.W.,

    doi:10.1016/0959-

  10. [27]

    With Applications to Statistics

    Weak Convergence and Empirical Pro- cesses. With Applications to Statistics. Springer Ser. Stat.. 2nd ed., Cham: Springer. doi:10.1007/978-3-031-29040-4. Wächter, A., Biegler, L.T.,

    doi:10.1007/978-3-031-29040-4

  11. [30]

    Optimal control of uncertain batch processes via Koopman expectation-assisted gradient methods. Comput. Chem. Eng. 203, 109353. doi:10.1016/j.compchemeng.2025.109353. 23

    arXiv 2025

  12. [39]

    Shapiro, A.,

    doi:10.1051/cocv/2023011. Shapiro, A.,

    arXiv

  13. [240]

    Scagliotti, A.,

    doi:10.1016/0959-1524(95)00015-I. Scagliotti, A.,

    work page doi:10.1016/0959-1524(95)00015-i

  14. [1953]

    On some asymptotic properties of maximum likelihood esti- mates and related Bayes’ estimates. Univ. California Publ. Stat. 1 , 277–329URL: https://hdl.handle.net/2027/wu.89045844305. Ledoux, M., Talagrand, M.,

    2027

  15. [1975]

    Central limit theorems forC(S)-valued random variables. J. Funct. Anal. 19, 216–231. doi:10.1016/0022-1236(75)90056-7. Javeed, A., Milz, J.,

    work page doi:10.1016/0022-1236(75)90056-7

  16. [1991]

    Classics in Mathematics, Springer

    Probability in Banach Spaces: Isoperimetry and Processes. Classics in Mathematics, Springer. doi:10.1007/978-3-642-20212-4. 20 Llanos, C.E., Sánchez, M.C.,

    doi:10.1007/978-3-642-20212-4

  17. [1994]

    Large sample confidence regions based on sub- samples under minimal assumptions. Ann. Stat. 22, 2031–2050. doi:10.1214/aos/ 1176325770. 21 Römisch, W., Surowiec, T.M.,

    doi:10.1214/aos/ 2031

  18. [1998]

    Semi-batch process optimization under uncertainty: Theory and experiments. Comput. Chem. Eng. 22, 201–213. doi:10.1016/S0098-1354(96)00359-6. van der Vaart, A.W., Wellner, J.A.,

    work page doi:10.1016/s0098-1354(96)00359-6

  19. [1999]

    Monte Carlo bounding techniques for determining solution quality in stochastic programs. Oper. Res. Lett. 24, 47–56. doi:10.1016/S0167-6377(98)00054-6. Melnikov, O., Milz, J.,

    work page doi:10.1016/s0167-6377(98)00054-6

  20. [2002]

    The sample average approx- imation method for stochastic discrete optimization. SIAM J. Optim. 12, 479–502. doi:10.1137/S1052623499363220. Le Cam, L.M.,

    doi:10.1137/s1052623499363220

  21. [2004]

    Comput.Chem.Eng.28, 333–346

    Sample average approximation methods for stochastic MINLPs. Comput.Chem.Eng.28, 333–346. doi:10.1016/s0098-1354(03)00194-

    work page doi:10.1016/s0098-1354(03)00194-

  22. [2005]

    Dynamic optimiza- tion of bioprocesses: Efficient and robust numerical strategies. J. Biotechnol. 117, 407–419. doi:10.1016/j.jbiotec.2005.02.013. Ben-Tal, A., El Ghaoui, L., Nemirovski, A.,

    work page doi:10.1016/j.jbiotec.2005.02.013 2005

  23. [2006]

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming. Math. Program. 106, 25–57. doi:10.1007/s10107-004-0559-y. Wei, J., Realff, M.J.,

    doi:10.1007/s10107-004-0559-y

  24. [2007]

    Stochastic integer programming: Limit theorems and confidence intervals. Math. Oper. Res. 32, 118–135. doi:10.1287/moor.1060

    doi:10.1287/moor.1060

  25. [2019]

    CasADi: a software framework for nonlinear optimization and optimal control. Math. Pro- gram. Comput. 11, 1–36. doi:10.1007/s12532-018-0139-4. 19 Banga, J.R., Balsa-Canto, E., Moles, C.G., Alonso, A.A.,

    doi:10.1007/s12532-018-0139-4

  26. [2021]

    Robust data reconcilia- tion in chemical reactors. Comput. Chem. Eng. 145, 107170. doi:10.1016/j. compchemeng.2020.107170. Eichhorn, A., Römisch, W.,

    arXiv 2020

  27. [2023]

    An efficient methodology to select high- performance M-estimators for robust data reconciliation. Comput. Chem. Eng. 176, 108297. doi:10.1016/j.compchemeng.2023.108297. Luus, R.,

    arXiv 2023

  28. [2024]

    doi:10.5281/zenodo.11669862

    EnsembleControl: Optimal control with an ensemble of dynamical systems. doi:10.5281/zenodo.11669862. Milz, J.,

    work page doi:10.5281/zenodo.11669862

  29. [2025]

    doi:10.48550/arxiv.2507.06982, arXiv:2507.06982

    Sample-based consistency in infinite-dimensional conic-constrained stochastic optimization. doi:10.48550/arxiv.2507.06982, arXiv:2507.06982. Jain, N.C., Marcus, M.B.,

    work page doi:10.48550/arxiv.2507.06982

  30. [2026]

    Convergence rates for ensemble-based solutions to optimal control of uncertain dynamical systems. SIAM J. Control Optim. 64, 1631–1650. doi:10.1137/24M1693076. de Menezes, D.Q.F., Prata, D.M., Secchi, A.R., Pinto, J.C.,

    work page doi:10.1137/24m1693076

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.