Pith. sign in

REVIEW 2 major objections 7 minor 33 references

Tight Analyses of Ordered and Unordered Linear Probing

T0 review · 2 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Linear probing's amortized high-load insertion cost is $\Theta(x \log^{1.5} x)$, with matching upper and lower bounds.

desk verdict The path-surplus result is the real deal; the linear-probing theorems close the FOCS'21 gap but lean on black-box lemmas from [4]. read the letter →

arxiv 2501.11582 v1 pith:SMXV5URZ submitted 2025-01-20 cs.DS

Mark Braverman , William Kuszmaul This is my paper · ORCID
classification cs.DS MSC 68W4060C0568P05
keywords linearprobinghashtablesamortizedanalysispathsurplustombstonesanti-clusteringrandomwalkshighloadfactor
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Linear-probing hash tables were long believed to have expected insertion time $\Theta(x^2)$ at load factor $1 - 1/x$. This paper proves, with matching upper and lower bounds, that the amortized expected insertion/deletion time of an ordered linear-probing table with tombstones is $\Theta(x \log^{1.5} x)$ when the rebuild window is chosen optimally, while queries cost $\Theta(x)$. The combinatorial engine is the path-surplus problem: the maximum over monotone paths of blue-minus-red dots beneath the path in a random $m \times m$ grid, which the paper determines to be $\Theta(m \log^{0.75} m)$. The same tight bounds are extended to classical unordered linear probing under the average-case workload of deleting random elements and inserting new ones, showing that the anti-clustering effect is not an artifact of ordering.

What carries the argument

The load-bearing object is the path-surplus problem: place $\mathrm{Pois}(m^2)$ blue and $\mathrm{Pois}(m^2)$ red dots uniformly in $[0,m]^2$, and define the surplus of a monotone path as the number of blue dots minus red dots beneath it. The lower bound builds a path by recursively splitting a segment and shifting its midpoint by $\pm w_L/\sqrt{\log m}$ according to the surplus of a thin rectangle $R_L$; the rectangles at distinct tree nodes are disjoint, so the induced random walk has independent steps and reaches depth $\Theta(\log m)$ with constant probability. The upper bound discretizes paths by the diagonal lines they cross at scales $d_\ell = m/2^\ell$, then bounds the surplus gain at each level minus two potential functions, the transition potential $\Phi$ and the frame potential $\Psi$, whose sums over all levels are $O(m \log^{0.75} m)$. This geometric bound is transferred to insertion surplus through Poissonization, smoothing, and rescaling, and then to linear probing through crossing-number identities of the form $c_s = \max_{i<s}(\mathrm{surplus}(i,s-1) - \mathrm{free}(i,s-1))$.

What would settle it

Run the path-surplus experiment for $m = 2^8, 2^{10}, \dots, 2^{18}$ and compare the observed maximum surplus against $c m \log^{0.75} m$; if the growth rate is not within a constant factor of that curve, the central geometric claim is wrong. A second check is to implement ordered linear probing with tombstones at load factor $1 - 1/x$ with rebuild window $n/\log^{1.5} x$ and measure the amortized expected insertion time, which should grow as $\Theta(x \log^{1.5} x)$.

Watch

Extended reading notes

Core claim

The central discovery is that the expected maximum surplus of a monotone path through an $m \times m$ grid with independent Poisson blue and red dots of density one is $\Theta(m \log^{0.75} m)$, and that this quantity exactly controls the amortized cost of linear probing. For an ordered linear-probing table on $n$ slots with tombstones, rebuild window $R = n/\beta$, and load factor never exceeding $1 - 1/x$, the paper proves that the worst-case amortized expected insertion/deletion time is $\Theta(x \log^{1.5} x + \beta x)$ and the worst-case expected query time is $\Theta(x + x \log^{1.5} x / \beta)$. The two curves cross at $R = \Theta(n/\log^{1.5} x)$, the unique optimal rebuild window, where insertions cost $\Theta(x \log^{1.5} x)$ and queries cost $\Theta(x)$. The same amortized bounds are obtained for classical unordered linear probing under an average-case workload, with the paper noting that the workload restrictions are necessary to avoid forced $\Theta(x^2)$ operations.

Load-bearing premise

The linear-probing results assume that the cited lemmas translating geometric surplus into crossing numbers, and crossing numbers into real operation costs, are correct; if any one of those translation steps has a hidden flaw, the $\Theta(x \log^{1.5} x)$ time bound fails even though the geometric path-surplus theorem stands.

Editorial extensions

If this is right

  • At the optimal rebuild window $R = \Theta(n/\log^{1.5} x)$, ordered linear probing achieves worst-case amortized expected insertion/deletion time $\Theta(x \log^{1.5} x)$ and worst-case expected query time $\Theta(x)$ whenever the load factor stays at or below $1 - 1/x$.
  • Choosing the rebuild window too short adds a $\beta x$ penalty to insertions; choosing it too long adds an $x \log^{1.5} x / \beta$ penalty to queries, and the two penalties balance exactly at the optimal window.
  • The tight path-surplus bound $\Theta(m \log^{0.75} m)$ closes the gap between the prior $m \log \log m$ lower bound and $m \log^{O(1)} m$ upper bound for that problem.
  • Classical unordered linear probing, the version most widely implemented, inherits the same amortized bounds under the stated average-case workload, showing that anti-clustering is a general linear-probing phenomenon.
  • The workload restrictions for the unordered result are essential: if insertions repeat the same element or queries are adversarial, one can force $\Theta(x^2)$-time operations again.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the recursive midpoint-shift construction is effectively a martingale random walk with step $1/\sqrt{\log m}$, so the same technique may give tight logarithmic corrections for other signed path processes, such as last-passage percolation with positive and negative weights.
  • The exact location of the optimal rebuild window suggests a concrete engineering heuristic: tombstones should be left in place for roughly $n/\log^{1.5} x$ operations rather than a much shorter window, because tombstone-driven anti-clustering is what keeps insertion costs low.
  • The paper proves expectation bounds; a natural next step would be to test concentration of the maximum path surplus by simulation, which the paper's methods do not by themselves settle.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 7 minor

Summary. The paper resolves the path surplus problem, proving that the expected maximum surplus of a monotonic path through an m-by-m grid with independent blue and red Poisson dots is Θ(m log^{0.75} m), and then transfers this result to linear probing. The authors derive matching upper and lower bounds of Θ(x log^{1.5} x) for the amortized expected insertion/deletion time in ordered linear probing with tombstones and rebuild windows, and they obtain analogous bounds for classical unordered linear probing under an average-case workload. The main technical work is a new lower-bound construction based on a recursive random walk and an upper bound using frame decompositions, potential functions, and an entropy-compression argument.

Significance. If the technical gaps discussed below are repaired, this is a significant advance: the path surplus problem is settled with a clean Θ(m log^{0.75} m) bound, and the paper gives the first tight asymptotic characterization of the amortized insertion time for linear probing with tombstones. The path-surplus upper bound is a substantial and largely self-contained contribution, and the idea of isolating three 'core facts' to transfer the analysis from the Poisson path-surplus setting to insertion surpluses is elegant. The linear-probing results are conditional on several lemmas from Bender et al. [4]; this is a legitimate dependency rather than a circularity, but the authors should verify that all hypotheses of those lemmas are met in the exact parameter regime used here.

major comments (2)
  1. [Section 4.1, Lemmas 4.4 and 3.10, Eq. (12)] The proof of Proposition 4.7 has an independence gap. In the modified insertion-surplus process, Core Fact 1 (Lemma 4.4) establishes independence only for vertical strips. Lemma 3.10, which is used verbatim in the transfer, decomposes the maximum transitional gain into a sum over j of random variables X_j that are argued to be independent because each depends on the dots in a strip between consecutive level-ℓ diagonals. These strips are not vertical, and in the insertion-surplus process a single Poissonized x-column can contribute dots to more than one such diagonal strip (for example, a key that is inserted at time i and deleted at time i' yields two dots in the same column at different times). The independence assertion in Lemma 3.10 therefore does not follow from Lemma 4.4. The authors need either to prove the needed independence for diagonal strips directly in the modified process, or to rework the frame decomposition so that it only relies on vertical-strip independence.
  2. [Section 6.1, Lemma 6.5, Eq. (36)] The tail bound on the sum ∑_ℓ A_{ℓ,s}K_ℓ is not justified as written. The proof states that this sum is bounded above by a binomial random variable with mean s/4, but A_{ℓ,s} is the event F_ℓ ≤ s + βs/8, and F_ℓ depends on the same sequence of failed deletions K_ℓ. The summands are therefore not independent and are not stochastically dominated by the claimed binomial merely because each conditional mean is at most (s+βs/8)/n. This step is load-bearing for Proposition 6.3 and hence for Theorem 1.4. A martingale or coupling argument is needed to justify the claimed 2^{-Ω(s)} decay.
minor comments (7)
  1. [Lemma 5.6] The display for t and the subsequent asymptotic bounds are garbled. With the intended t = (x^2/(qβ)) log^{1.5} x, the lower bound should read Ω((x/(β√q)) log^{1.5} x − (x/(qβ)) log^{1.5} x), which yields the claimed Ω(x log^{1.5} x / β) after summing over s. Please correct the typesetting and the intermediate inequalities.
  2. [Corollary 4.12] In the definition of surplus(O, [j−t,t]), the interval should be [j−t, j], consistent with the surrounding notation.
  3. [Lemma 3.13] The sentence 'This proves (21), which then implies (18) and (19)' should refer to (18) and (20), since (19) is proved separately by direct calculation.
  4. [Lemma 3.14] The bound |C_q| ≤ 2^{O(q)+ℓ} does not match the previously stated unary encoding length O(‖Δ(F)‖_1 + 2^ℓ). The correct bound is |C_q| ≤ 2^{O(q + 2^ℓ)}; the subsequent argument still works when α is sufficiently large, but the display should be fixed.
  5. [Lemma 4.1] The claim that the partial smoothing step 'does not change surplus(P)' is not immediate for arbitrary monotone paths and for dots lying on the path boundary. Please add a formal argument or state the convention used for points on the path.
  6. [Lemma 3.8] The notation P_0 is used in the telescoping sum near the end of the proof but is not defined; it should denote the set of rigid paths at level 0.
  7. [Corollary 5.4] There is a typo in 'wort-case'; it should be 'worst-case'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the path-surplus analysis is self-contained, and the linear-probing conversion lemmas from prior work are independent dependencies, not inputs recycled as outputs.

full rationale

The paper's central derivation is self-contained: the tight path-surplus bounds (Theorem 3.1 lower bound and Theorem 3.2 upper bound) are proved directly from Poisson/Chernoff bounds, the explicit random-walk structure of the lower-bound construction, and the potential-function framework of Section 3.2. No parameter is fitted to data, and no theorem is made true by definition. The Insertion Surplus reduction in Sections 4.1 and 4.2 is also explicit: Poissonization, smoothing, and rescaling are justified with quantitative O(t^2/n) or O(1/m^2) error terms, and the lower bound reduces formally to the original path-surplus problem, not to the desired linear-probing conclusion. The final linear-probing theorems (Theorems 1.3 and 1.4) do delegate several conversion steps to Bender et al. [4], including Lemmas 5.1, 5.2, 5.5, 5.8 and Lemma 6.2, and these citations include the current coauthor. However, these are legitimate external dependencies: [4] predates the exact Θ(log^1.5 x) bound, and its lemmas convert crossing numbers to operation costs without assuming the target result. The paper supplies the new quantitative path-surplus and insertion-surplus bounds and feeds them into that cited machinery. A hidden flaw in one of the cited lemmas would be a correctness risk, not circular reasoning, because the present derivation does not reduce to its own output by construction. There is also no renaming of a known result: the Θ(m log^{0.75} m) path-surplus bound and the Θ(x log^{1.5} x) amortized insertion bound are new statements derived through the chain described in the paper. Overall, I find no circular step warranting a nonzero score.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted to data; the constants in the proofs are universal asymptotic constants. The paper introduces no new physical or mathematical entities beyond the formal definitions of path surplus and insertion surplus, which are not independent postulates. The main external dependencies are fully random hashing and black-box lemmas from the authors' prior FOCS'21 paper.

assumptions (3)
  • domain assumption The hash function h is fully random (uniform and independent for each key).
    Assumed in Section 2 and used throughout; the paper notes that standard techniques can simulate fully random hash functions, but it is open whether O(1)-independence suffices for the amortized bounds.
  • domain assumption Bender et al. [4] lemmas converting crossing numbers to operation costs (Lemmas 5.1, 5.2, 5.5, 5.8, and Proposition 3) are correct.
    Section 5 imports these results without reproof; the insertion and query time bounds in Theorems 1.3 and 1.4 depend on them.
  • standard math Standard probabilistic inequalities (Chernoff, Kolmogorov, Azuma, negative association) and Poissonization are valid in the regimes used.
    Used throughout Sections 3 and 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tight Analyses of Ordered and Unordered Linear Probing." pith.science (2026). https://pith.science/paper/SMXV5URZ

@misc{pith2026250111582,
  author       = {Pith},
  title        = {Pith review of: Tight Analyses of Ordered and Unordered Linear Probing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMXV5URZ}},
  note         = {Machine review of arXiv:2501.11582}
}
abstract

Linear-probing hash tables have been classically believed to support insertions in time $\Theta(x^2)$, where $1 - 1/x$ is the load factor of the hash table. Recent work by Bender, Kuszmaul, and Kuszmaul (FOCS'21), however, has added a new twist to this story: in some versions of linear probing, if the \emph{maximum} load factor is at most $1 - 1/x$, then the \emph{amortized} expected time per insertion will never exceed $x \log^{O(1)} x$ (even in workloads that operate continuously at a load factor of $1 - 1/x$). Determining the exact asymptotic value for the amortized insertion time remains open. In this paper, we settle the amortized complexity with matching upper and lower bounds of $\Theta(x \log^{1.5} x)$. Along the way, we also obtain tight bounds for the so-called path surplus problem, a problem in combinatorial geometry that has been shown to be closely related to linear probing. We also show how to extend Bender et al.'s bounds to say something not just about ordered linear probing (the version they study) but also about classical linear probing, in the form that is most widely implemented in practice.

Figures

Figures reproduced from arXiv: 2501.11582 by the authors.

Figure 1
Figure 1. An example with m2 = 100. In this case, there are 110 blue dots and 97 red dots. A surplus-maximizing monotonic path is given, and the surplus of the path is 32. this trivial path as our final solution. Otherwise, we remove the quadrant from the path, we break the path in half (one half traveling over the bottom left quadrant, and one half traveling over the top right quadrant) and we recursively apply the same cons… view at source ↗
Figure 2
Figure 2. The same path as in Figure 1, but in the rotated versi [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. An example of what the base-case subproblem (i.e., [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: If we look at the implied path Path(F) for F, we take the line segment between diagonals D (i−1) ℓ and D (i) ℓ , and we extend that segment to reach diagonal D (i+1) ℓ , then ∆i measures the distance (in multiples of √ 2qℓ) between where the extended segment hits D (i+…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [4]

    Linear Probing Revisited: Tombstones Mark the Death of Primary Clustering

    Michael A Bender, Bradley C Kuszmaul, and William Kuszma ul. Linear probing revisited: Tombstones mark the death of primary clustering. arXiv preprint arXiv:2107.01250, 2021

    work page Pith review arXiv 2021

  2. [1]

    Accessed: 2020-11-06

    Abseil, 2017. Accessed: 2020-11-06. URL: https://abseil.io/

    work page 2017

  3. [2]

    Ordered hash tables

    Ole Amble and Donald Ervin Knuth. Ordered hash tables. The Computer Journal , 17(2):135–142, January 1974. doi:10.1093/comjnl/17.2.135

    work page doi:10.1093/comjnl/17.2.135 1974

  4. [3]

    Bender, Martin Farach-Colton, Rob Johnson, R ussell Kaner, Bradley C

    Michael A. Bender, Martin Farach-Colton, Rob Johnson, R ussell Kaner, Bradley C. Kuszmaul, Dzejla Medjedovic, Pablo Monte s, Pradeep Shetty, Richard P. Spillane, and Erez Zadok. Don’t thrash: How to cache your has h on flash. Proceedings of the VLDB Endowment , 5(11):1627–1637, 2012. doi:10.14778/2350229.2350275

    arXiv 2012

  5. [5]

    Linear probing revisited: Tombstones mark the demise of primary clusterin g

    Michael A Bender, Bradley C Kuszmaul, and William Kuszma ul. Linear probing revisited: Tombstones mark the demise of primary clusterin g. In 2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS) , pages 1171–1182. IEEE, 2022

    work page 2021

  6. [6]

    Locally Uniform Hashing

    Ioana O Bercea, Lorenzo Beretta, Jonas Klausen, Jakob Bæ k Tejs Houen, and Mikkel Thorup. Locally uniform hashing. arXiv preprint arXiv:2308.14134, 2023

    work page Pith review arXiv 2023

  7. [7]

    Open-sourcing F14 for faster, more memory-efficient has h tables, 25 April 2019

    Nathan Bronson and Xiao Shi. Open-sourcing F14 for faster, more memory-efficient has h tables, 25 April 2019. Accessed: 2020-11-06. URL: https://engineering.fb.com/2019/04/25/developer-too ls/f14/

    work page 2019

  8. [8]

    Ian Munro

    Pedro Celis, Per- ˚ Ake Larson, and J. Ian Munro. Robin Hood hashing (preliminar y report). In 26th Annual Symposium on Foundations of Computer Science (FO CS’85), pages 281–288, Portland, Oregon, USA, 21–23 October 1985. doi:10.1109/SFCS.1985.48

    work page doi:10.1109/sfcs.1985.48 1985

Show all 33 references
  1. [9]

    Almost rand om graphs with simple hash functions

    Martin Dietzfelbinger and Philipp Woelfel. Almost rand om graphs with simple hash functions. In Proceedings of the Thirty-Fifth Annual ACM Symposium on Theory of Computing (STOC’03), pages 629–638, San Diego, California, USA, 9–11 June 2003. doi:10.1145/780542.780634

    2003

  2. [10]

    Adam Drozdek and Donald L. Simon. Data Structures in C . PWS, Boston, Massachusetts, USA, 1995

    1995

  3. [11]

    Quotient filters: Approximate membership queries on the GPU

    Afton Geil, Martin Farach-Colton, and John D Owens. Quotient filters: Approximate membership queries on the GPU. In Proceedings of the IEEE International Parallel and Distributed Processing Symp osium (IPDPS) , pages 451–462, Vancouver, British Columbia, Canada, 21–25 May 2018....

    2018

  4. [12]

    Notes on “open” addressing, 1963

    Don Knuth. Notes on “open” addressing, 1963

    1963

  5. [13]

    The Art of Computer Programming, Volume III: Sorting and Searchin g

    Donald Ervin Knuth. The Art of Computer Programming, Volume III: Sorting and Searchin g. Addison-Wesley, 2nd edition, 1998. URL: https://www.worldcat.org/oclc/312994415

    1998

  6. [14]

    Konheim and Benjamin Weiss

    Alan G. Konheim and Benjamin Weiss. An occupancy discip line and applications. SIAM Journal on Applied Mathematics , 14(6):1266–1274, November 1966. doi:10.1137/0114101

    1966 doi

  7. [15]

    Robert L. Kruse. Data Structures and Program Design. Prentice-Hall Inc, Englewood Cliffs, New Jersey, USA, 1984. 53

    1984

  8. [16]

    The multiplicative version of azuma’s inequality, with an application to contention analysis

    William Kuszmaul and Qi Qi. The multiplicative version of azuma’s inequality, with an application to contention analysis. arXiv preprint arXiv:2102.05077, 2021

    2021 arXiv

  9. [17]

    Lewis and Larry Denenberg

    Harry R. Lewis and Larry Denenberg. Data Structures and Their Algorithms . HarperCollins Publishers, New York, New York, USA, 1991

    1991

  10. [18]

    Data Structures and Other Objects Using C++

    Michael Main and Walter Savitch. Data Structures and Other Objects Using C++ . Addison-Wesley, Boston, Massachusetts, USA, 2001

    2001

  11. [19]

    Probability and computing: Randomization and probabilistic techniques in algorithms and data analys is

    Michael Mitzenmacher and Eli Upfal. Probability and computing: Randomization and probabilistic techniques in algorithms and data analys is. Cambridge university press, 2017

    2017

  12. [20]

    Uniform hashing in constant t ime and optimal space

    Anna Pagh and Rasmus Pagh. Uniform hashing in constant t ime and optimal space. SIAM Journal on Computing , 38(1):85–96, 2008. doi:10.1137/060658400

    2008 doi

  13. [21]

    Linear probing with constant independence

    Anna Pagh, Rasmus Pagh, and Milan Ruzic. Linear probing with constant independence. In Proceedings 39th Annual ACM Symposium on Theory of Computing (S TOC’07), pages 318–327, San Diego, California, USA, 11–13 June 2007. doi:10.1145/1250790.1250839

    2007

  14. [22]

    Bender, Martin Farach-Colton, and Rob Johnson

    Prashant Pandey, Alex Conway, Joe Durie, Michael A. Bender, Martin Farach-Colton, and Rob Johnson. Vector quotient filters: Overcoming the time/space trade-off in filter design . In Proceedings of the 2021 ACM SIGMOD International Conference on Management of Data (SIGMO D’21), 20–25 2021

    2021

  15. [23]

    W. W. Peterson. Addressing for random-access storage. IBM Journal of Research and Development , 1(2):130–146, April 1957. doi:10.1147/rd.12.0130

    1957 doi

  16. [24]

    A se ven-dimensional analysis of hashing methods and its implications on query processing

    Stefan Richter, Victor Alvarez, and Jens Dittrich. A se ven-dimensional analysis of hashing methods and its implications on query processing. PVLDB – Porceedings of the VLDB Endowment, 9(3):96–107, November 2015. The 42nd International Confe rence on Very Large Data Bases, New...

    2015

  17. [25]

    Algorithms

    Robert Sedgewick. Algorithms. Addison-Wesley, Reading, Massachusetts, USA, 1983

    1983

  18. [26]

    Algorithms in C

    Robert Sedgewick. Algorithms in C . Addison-Wesley, Reading, Massachusetts, USA, 1990

    1990

  19. [27]

    On universal classes of extremely random constant-time hash functions

    Alan Siegel. On universal classes of extremely random constant-time hash functions. SIAM Journal on Computing , 33(3):505–543, 2004

    2004

  20. [28]

    Standish

    Thomas A. Standish. Data Structures, Algorithms, and Software Principles in C . Addision-Wesley, Reading, Massachusetts, USA, 1995

    1995

  21. [29]

    Sorenson

    Jean-Paul Tremblay and Paul G. Sorenson. An Introduction to Data Structures with Applications . McGraw-Hill, 1984

    1984

  22. [30]

    Negative association: definition, properties, and applications

    David Wajc. Negative association: definition, properties, and applications. Manuscript, available from https://goo. gl/j2ekqM , 2017

    2017

  23. [31]

    Data Structures and Problem Solving using C++

    Mark Allen Weiss. Data Structures and Problem Solving using C++ . Addison-Wesley, Reading, Massachusetts, USA, 2000

    2000

  24. [32]

    Linear probing,

    Wikipedia contributors. Linear probing,

  25. [2021]

    URL: https://en.wikipedia.org/wiki/Lienar_probing

    Accessed 31-May-2021. URL: https://en.wikipedia.org/wiki/Lienar_probing. 54

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.