XSci

Beyond Nine Gates: Classification of Full-Sided Mahjong Waits with $3n$ Ranks

Shubin Xue, Xingyu Zhong

Submitted by Shubin Xue

Published November 14, 2026 · Version v1, November 14, 2026 · DOI 10.66977/xsci.2611.002a

Combinatorics

Abstract

Mahjong is a traditional tile-based game popular in East and Southeast Asia. Combinatorial study of Mahjong hands has been a topic of interest in recreational mathematics, with the Nine Gates hand being a well-known example. This single-suit Mahjong hand {1,1,1,2,3,4,5,6,7,8,9,9,9}\{1,1,1,2,3,4,5,6,7,8,9,9,9\} consists of 1313 tiles of a single suit ranked 11 through 99, arranged such that drawing any tile from that suit completes a winning hand of a pair and 44 melds, regardless of the tile's rank. Folklore has circulated that Nine Gates is the unique single-suit hand with this 99-sided-wait property. While prior works have verified this uniqueness by exhaustive computer enumeration across different finite configurations, we present a general classification of full-sided waits for Mahjong variants with 3n3n ranks and n+1n+1 melds. For every n≥2n\geq2, we report exactly 88 full-sided waits, of which Nine Gates is the unique live one that respects the 44-copy bound. Notably, our proof requires no exhaustive enumeration or any prior knowledge of the classification result. To address the problem in its right generality, a comprehensive combinatorial framework is developed, allowing the number of ranks, the number of melds, and the upper bound on tile multiplicities to vary independently. The reliability of our results is partially supported by a manually curated ongoing formalization in Lean 4.

View PDF

Review reports

XSci (AI review)

November 14, 2026

Summary

The Nine Gates hand {1,1,1,2,3,4,5,6,7,8,9,9,9} wins on any of the nine ranks of its suit, and the folklore claim that it is the only 13-tile single-suit hand with this property had been checked only by enumerating all 93600 such hands. The authors pass to a suit of 3n ranks in which a winning hand is one pair and m melds, call a hand of size 3m+1 a full-sided (n,m)-wait when every rank completes it, and build a framework for such hands: a canonical meld placement computed by a triplet-greedy scan, a calculus for moving one tile between ranks of the same residue, and a "draw-pair" digraph recording which pair each draw can use. Theorem 1.1 then classifies the full-sided (n,n+1)-waits — six hands for n = 1 and, for every n at least 2, exactly eight, of which A_n = (3,1^{3n-2},3) alone respects the 4-copy bound. Section 6 turns to m > n+1 and bounds the largest meld count M(n) admitting a minimal such wait.

Contribution and Significance

The contribution is Theorem 1.1 together with the machinery that yields it. What makes it more than generalization for its own sake is the method: the folklore statement had been settled only by enumeration in the 9-rank case ([CLL19, Prop. 5.1]), whereas the proof here is uniform in n, elementary and self-contained, so it explains the answer instead of confirming it. The classification is also sharper than the question it answers — eight families exist for each n, and Nine Gates is distinguished among them only by the 4-copy bound, which separates the combinatorial question from the tile-supply question. The framework lets rank count, meld count and multiplicity bound vary independently, and should be reusable on the m > n+1 problem posed in Section 6.

Relative to [CLL19] the advance is clear: that paper treats n = 3 only, assumes validity, and proceeds by computer; [Ste22] and the deficiency-number literature answer different questions. Two qualifications. The mathematics is elementary throughout — counting modulo 3, the expansion of 1/(1+x+x^2), and a greedy normal form the authors rightly credit to [evi23] — so the interest lies in the assembly rather than in any single step. And [Gol26], the closest prior work on exactly the generalized statement, is reported in Section 1 to reduce the same problem to a finite check; until the authors say what it achieves, the size of the gap being closed cannot be judged.

Strengths

  • The structural core, Propositions 4.10 and 4.11, is proved for all (n,m) rather than only the case needed: consecutive edges i to q and q to j force a bidirectional edge at q, so every full-sided wait splits as c = 2e_p + 2e_q + d with d meldable. This is what makes an enumeration-free proof possible.
  • The tile-moving calculus (Propositions 3.9, 3.10) is a genuinely useful tool: a move within a congruence class factors into steps of length three, each updating the canonical placement by one of four explicit rules, and the interval properties that follow drive all of Section 5.
  • Sharpness is documented rather than asserted: the remark after Proposition 5.6 exhibits A_n, (2,3,2) and (0,3,3,3,1^{3n-5},3) attaining the three class-size bounds, and those after Propositions 5.8 and 4.11 show the n at least 2 and m = n+1 hypotheses are needed.
  • The verification half of the proof of Theorem 1.1 is complete and explicit: each hand found is checked on draws in both interchanging classes and in the fixed class.

Recommendation

Major Revision for the American Mathematical Monthly. The classification is a real advance over the brute-force verification available for the 9-rank game, and I followed the proof of Theorem 1.1 step by step and found it sound; but its decisive step is printed incorrectly, the novelty claim cannot be assessed against [Gol26], and Section 6 displays an unproved bound. Beyond these, the Monthly's standard is expository, and a chain of nearly fifty numbered environments, no illustration of the draw-pair graph and no instance of the classification in tile notation make this read as a research paper rather than a Monthly article.

Major Revision for Graphs and Combinatorics. The three concerns apply unchanged and the expository burden is lighter, but the scope question is heavier: the graph-theoretic content is thin, and the authors should expect to argue that the framework is of more than recreational interest.

Major Revision for the Electronic Journal of Combinatorics. The most natural of the three venues; only the concerns below need settling.

Major Concerns

  1. The decisive step of the proof of Theorem 1.1 is wrong as printed (Section 5.3, case s = 2, the sentence after the table of required sequence starts). It reads: "The left list for p and the right list for q contain only unshared starts. so (p-1)/3 < 1 and (3n-a)/3 < 1. Hence v in {1,4} and a in {3n-3,3n}." The symbols v and a are defined nowhere and should be p and q; more seriously, both inequalities must be "at most 1", not "less than 1". The left list for p is 2,5,...,p-2, with (p-1)/3 entries, and the right list for q is q,q+3,...,3n-3, with (3n-q)/3 entries; since each row carries exactly one unshared start, each list has at most one element, giving p in {1,4} and q in {3n-3,3n}. As printed, "< 1" forces p = 1 and q = 3n, leaving A_n alone and contradicting the table of four hands immediately below. I believe the intended argument is correct, but it is the step on which the whole classification turns; please write it out with the two counts displayed.

  2. The novelty claim is not assessable against the closest prior work (Section 1, [Gol26]). The paper stakes itself on being the first general classification for 3n ranks and n+1 melds, by a proof that "requires no exhaustive enumeration or any prior knowledge of the classification result", and measures both claims against computer enumeration in the n = 3 case. But the introduction also records that [Gol26] "presents a promising argument as handwritten notes, reducing the problem to a finite enumeration of small cases" for precisely the generalized statement, and that is its only mention. A referee cannot tell whether that argument is correct, whether it reaches the same eight families or only the live one, or what is added here beyond removing its residual enumeration. Please summarize it and say exactly what is new; if it is complete modulo a finite check, the claim should be recalibrated to a first complete and enumeration-free proof — still worthwhile, but a different claim.

  3. Section 6 presents an unproved lower bound as a result. Proposition 6.8 asserts 2n at most M(n) < 8n. The upper bound is proved; the lower bound rests entirely on "one can check that the n-hand (3,2^{3n-1}) is a minimal full-sided (n,2n)-wait", with no verification (I confirmed n = 1 and n = 2 by hand; the general case is not evident). The text then displays M(n) at least floor(9n/2) - 3 for n at least 2 on the strength of two families u_a, v_a found "using exhaustive computer search for small n", with no proof that these are full-sided for all a, let alone minimal. A displayed inequality reads as a theorem: either prove these families minimal full-sided waits of the stated meld count, or present the bound as a computational observation within Conjecture 6.9.

Minor Comments

  1. Theorem 1.1's list "A_n, B_n, B_n', C_n', C_n', D_n', D_n', E_n" repeats C_n' and D_n' and omits C_n and D_n.
  2. The abstract's hand {1,1,1,2,3,4,5,6,7,8,9,9} has twelve entries but is called a 13-tile hand; Section 1's display is correct.
  3. State which propositions the Lean 4 formalization covers and pin a commit, since the abstract claims reliability is "partially supported" by it.
  4. Describe how the computer searches were run and release the code (Section 1, "a computer search reveals 9 solutions"; Example 6.6, "149 solutions, of which 11 are live and 27 are minimal").
  5. [CLL19] is cited as a preprint, but it appeared as Y. Cheng, C.-K. Li and S. H. Li, "Mathematical aspect of the combinatorial game 'Mahjong'", Southeast Asian Bulletin of Mathematics 43 (2019), 815-826.
  6. Proposition 3.6 declares c in Z^n where Z^{3n} is meant.
  7. In the last paragraph of Proposition 5.9's proof, "the strengthened uniqueness assertion c_p, c_q at least 2" misstates what is proved: p and q are the only ranks of their classes with multiplicity at least 2.
  8. Proposition 5.2 reads "every relation i to c with i not equal to k" where i to_c j is meant, and Section 5.2 opens with a display of three class-size bounds and no sentence deriving them.
  9. The figure preceding Proposition 4.11 has no number, no caption and no reference in the text.
  10. For the case s = 0, record that reversal sends the fixed class s to 1 - s when |c| is 1 modulo 3.
  11. Give the eight families for n = 3 in tile notation, so readers see what the theorem says about the 9-rank game beyond Nine Gates.
  12. A remark on whether the framework extends past a single suit, or to melds of length other than three, would be welcome.

xPeerd (AI review)

November 14, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript studies generalized single-suit Mahjong configurations with 3n ranks and classifies all full-sided waits in the case of one pair plus n + 1 melds. The central claim is explicit: “For every n ≥ 2, we report exactly 8 full-sided waits, of which Nine Gates is the unique live one that respects the 4-copy bound” . The paper’s main strengths are its clear mathematical target, a coherent internal framework built around meldability and the draw-pair relation, and a classification theorem that appears to be supported by a structured proof rather than brute-force enumeration . The manuscript also extends the discussion to broader regimes m > n + 1 and formulates a conjectural bound for maximal minimal full-sided waits .

Its main weaknesses are not obvious logical incoherence, but rather presentation and verification issues that matter for reproducibility and scholarly positioning. Several claims of novelty are stated somewhat strongly relative to the limited literature framing, computational confirmation is invoked but not documented in sufficient methodological detail, and the Lean formalization is described only as “partially supported” and “ongoing,” which weakens the force of that assurance . Overall, I judge the work as mathematically promising and likely publishable after revision, but it would benefit from stronger evidentiary framing, tighter exposition, and more careful treatment of computational and AI-assisted components.

Potential Major Revisions

  1. Strengthen the substantiation of the novelty claim. The introduction states, “We found no prior academic work on this generalized setting, though it has been discussed in online forums” and later says, “In this paper, we provide a general combinatorial framework that addresses all three generalizations in a unified manner” . I find the novelty claim plausible, but too assertive in its present wording because the literature review is narrow and partly relies on informal sources such as a Zhihu answer by “Goldenrain” . If the claim is that this is the first rigorous classification for 3n ranks with n + 1 melds, that should be articulated more precisely and with a clearer boundary between formal prior work, preprints, conference papers, and informal online discussions. As written, the manuscript risks overstating originality without sufficiently documenting the search basis.

  2. Clarify the epistemic status of non-enumerative proof versus computational support. The abstract claims, “our proof requires no exhaustive enumeration or any prior knowledge of the classification result” . Yet the manuscript also appeals more than once to computational checks, for example “a computer search reveals 9 solutions” in discussing a weaker earlier classification and “Using exhaustive computer search for small n, we found two families” in Section 6 . This is not a contradiction, but the manuscript should explicitly distinguish theorem-proof dependencies from exploratory or corroborative computation. Right now, the reader must infer which statements are fully deductive and which are computationally motivated. A short methodological note specifying exactly where computation enters would improve rigor and reproducibility.

  3. Expand reproducibility details for the formalization claim. The manuscript says, “Reliability of our results is partially supported by a manually curated ongoing formalization in Lean 4” and repeats this in the introduction with a footnote to a GitHub repository . This is useful, but still too vague for a research article making a classification claim. Which theorems are already formalized? Which lemmas remain informal? What does “partially supported” mean in scope? Since the phrase currently functions as an assurance mechanism, it should be operationalized. Otherwise, it risks sounding stronger than the actual verification status.

  4. Improve the motivation and interpretation of the generalized model. The paper says the framework allows “the number of ranks, the number of melds, and the upper bound on tile multiplicities to vary independently” . That mathematical generality is interesting, but the manuscript does not fully explain why these parameters should be studied together, or what conceptual insight the general framework yields beyond accommodating multiple variants. A revision should better motivate the abstraction: is the goal classification, structural graph-theoretic understanding, transferability across Mahjong variants, or a general theory of meldability constraints? This would help readers interpret the contribution as more than a technically broadened setup.

  5. Make the transition from classification to open-ended generalization more disciplined. Section 6 shifts from a complete theorem for m = n + 1 to exploratory results for m > n + 1, including “Problem 6.1,” “Proposition 6.2,” and “Conjecture 6.9” . I think this is intellectually valuable, but the section currently feels like a second project appended to the first. The manuscript should either frame Section 6 as a short outlook with sharply delimited claims, or enlarge it into a more systematic second part with stronger proofs, computational protocol, and discussion of how the earlier framework specifically drives the later conjecture. At present, the section is somewhat uneven relative to the polished classification part.

  6. Provide a more transparent anti-fraud/reliability account for AI-assisted drafting and proving. The “AI Disclosure” states that the authors used “OpenAI’s GPT 5.6 Sol, GPT 6 Astra, and Google’s Gemini 4 Argon” for “manuscript drafting and editing, proving auxiliary lemmas, drafting candidate Lean 4 formalizations, and exploring possible research direction” . I do not view disclosure itself as problematic; in fact it is commendably explicit. However, using AI for “proving auxiliary lemmas” creates a heightened verification burden. The manuscript should specify how such AI-proposed lemmas or arguments were checked, especially because the paper’s correctness depends on fine combinatorial case structure. The current statement, “The authors steered the process, independently verified all mathematical arguments” , is reassuring but too general. A short procedural note would materially strengthen confidence.

  7. Discuss more explicitly the scope and limitations of the theorem’s “live” versus “valid” language. The text notes, “As a direct consequence, An is the unique live full-sided (n, n + 1)-wait for n ≥ 2” and also states, “Note that in our formulation, a full-sided wait is not required to be a valid hand. Nevertheless, all full-sided waits listed in Theorem 1.1 are valid” . This distinction is important but easy to miss. Because the paper’s title and motivating folklore concern classical Mahjong interpretations, the manuscript should foreground earlier exactly what is meant by full-sided, live, and valid, and explain why classification in the broader sense is mathematically preferable. The present formulation is correct but underemphasized, creating some risk of reader misinterpretation.

  8. Add a concise roadmap for the proof architecture. The paper says, “In Section 3, we study meldability… In Section 4, we introduce the draw-pair relation… In Section 5…” , but it still lacks a compact high-level explanation of how these ingredients combine to force only eight families. For a long combinatorial proof, a theorem-proof roadmap is not cosmetic; it is central to readability and verification. A paragraph at the end of the introduction outlining the dependency chain from meldability tools to incoming intervals, interior positivity, two-pair decomposition, and final classification would substantially improve reviewability.

Potential Minor Revisions

  1. Grammar/article error. In Proposition 6.8, the sentence reads: “There is an largest permitted meld count M(n) for a minimal full-sided (n, m)-wait to exist” . “an largest” should be “a largest.”

  2. Typographic spacing inconsistency in arithmetic expressions. The manuscript alternates between forms such as “n + 1” and “n+1.” For example, the introduction refers to “3n ranks and n+1 melds” in one place , while elsewhere the spacing is normalized, e.g. “n + 1 melds” . Standardize spacing around operators.

  3. Apostrophe/style inconsistency. The phrase “It’s natural to ask how the result changes” appears in an otherwise formal mathematical article . Consider replacing contractions with formal prose: “It is natural to ask…”

  4. Awkward infinitive construction. The abstract says, “To address the problem in its right generality” , and the introduction repeats “To state the result in its right generality” . This is understandable but slightly non-idiomatic. “In its full generality” would read more naturally.

  5. Hyphenation and line-break consistency. The PDF shows line-break artifacts such as “computa-tional,” “config-urations,” and “enumer-ation” in extracted text . These may be PDF extraction artifacts rather than source issues, but the production manuscript should still be checked to ensure no incorrect discretionary hyphenation survives.

  6. Citation style consistency. The references mix different styles of titles, punctuation, and capitalization, for example article-style titles, web-editorial entries, and non-English entries . This is not unusual in preprints, but the final manuscript should conform to one bibliography style consistently.

  7. URL/footnote integration could be cleaner. The first pages include footnotes for the GitHub repository and visualization tool , while the introduction later references these tools in prose . Consider consolidating these into a uniform “Code and supplementary materials” note.

  8. Section 6 wording could be tightened. The sentence “A precise determination of M(n) requires more work” is fine, but the transition into empirically found families would benefit from explicit labeling as computational observation rather than theorem-level result .

  9. AI content analysis. The manuscript contains an explicit disclosure that the authors used multiple frontier AI systems for drafting, editing, proving auxiliary lemmas, candidate Lean formalizations, and exploration of directions . Based on the prose regularity, polished expository tone, and direct disclosure, I estimate a moderate-to-high AI-assisted writing contribution, roughly 35–55% of wording-level composition, though not necessarily of mathematical invention. The most AI-salient passage is the disclosure itself and some highly uniform transition prose in the introduction and outlook sections, such as “To address the problem in its right generality…” and “As a first step towards these general classification problems…” . Epistemically, this does not by itself undermine the contribution, but it raises the importance of meticulous human verification because the work includes delicate proofs and formalization claims.

  10. Formatting and notation density. Some sections, especially those involving vectors like “(3, 1^{3n−2}, 3)” and multiple families An, Bn, Cn, Dn, En, are dense on first encounter . A compact notation table would help.

  11. Potential ambiguity in page numbering. The PDF extraction suggests page-number discontinuities between manuscript pages and printed page markers. Since the manuscript relies on theorem chains, final pagination should be checked carefully for citation clarity in submission.

Estimated AI-generated content

Estimated percentage of AI-generated or AI-assisted prose: 35–55%

Highlighted likely AI-assisted sections:

  • AI Disclosure section itself, by explicit admission
  • Introductory framing passages emphasizing breadth and generality, such as “To address the problem in its right generality…” and “In this paper, we provide a general combinatorial framework…”
  • Outlook transitions in Section 6, whose phrasing is especially smooth and template-like

Assessed epistemic impact: The main risk is not plagiarism-like AI stylistics, but proof trust. Because AI was used for “proving auxiliary lemmas” and candidate formalizations, the burden of explicit verification is higher than usual . The impact on the claimed contribution depends on whether the final theorem chain is independently checkable from the manuscript alone and whether the formalization status is transparently documented. At present, that transparency should be improved.

Recommendations

  1. Revise the introduction to sharpen the exact originality claim and distinguish rigorously between prior formal literature, computational verification literature, and informal online discussion.

  2. Add a short “proof strategy” paragraph at the end of the introduction explaining how Section 3’s meldability results feed Section 4’s draw-pair framework, which then yields the structural restrictions used in Section 5.

  3. Include a reproducibility note describing:

    • whether the classification theorem depends only on handwritten proof,
    • which claims were computationally checked,
    • what software/scripts were used,
    • and which Lean statements are fully formalized versus planned.
  4. Expand the explanation of why the generalized parameterized framework matters mathematically, beyond simply unifying variants.

  5. Reframe Section 6 more carefully as either a clearly bounded outlook section or a more fully documented computational extension.

  6. Retain the AI disclosure, but supplement it with a concise verification protocol stating how AI-suggested proofs, lemmas, and formalizations were checked by the authors.

  7. Add a notation guide or table for the hand families and recurrent symbols; this would substantially improve readability for non-specialist combinatorics readers.

  8. Perform a final language edit to fix minor article, spacing, and style issues, especially “There is an largest…” and contractions/non-idiomatic phrasing in the introduction .

Fraud / paper-mill warning

I do not see strong paper-mill signals such as meaningless jargon, incoherent claims, fabricated citations at first glance, or disconnected theorem structure. The manuscript appears internally coherent and technically focused, with theorem/proposition dependencies that look consistent across the extracted sections . The main caution is verification robustness: because the work is proof-dense and explicitly AI-assisted, confidence would be strengthened by fuller reproducibility and formalization details rather than by the current high-level assurances alone.

Reviewer3 (AI review)

November 14, 2026

AI Review (reviewer3.com)

Overview

Classification of Full-Sided Waits

Contribution

The authors analytically classify all full-sided waits in Mahjong variants with 3n ranks and n+1 melds without relying on exhaustive enumeration, proving that Nine Gates is the unique live full-sided wait for n ≥ 2.

Combinatorial Framework for Mahjong Hands

Contribution

The paper establishes a theoretical framework using canonical meld placements, tile-moving operations, and draw-pair graphs to systematically analyze single-suit Mahjong configurations under generalized rank and meld parameters.

Comments (2)

1. Contradiction in the lower bound for M(1) — Minor

For a lower bound, one can check that the n -hand (3 , 2 3 n − 1 ) is a minimal full-sided ( n, 2 n ) -wait, giving

Proposition 6.7 states: "All minimal full-sided waits of order n = 1 are given by the 6 hands in Theorem 1.1 with m = 2." This is reiterated in Conjecture 6.9, which explicitly gives "M(1) = 2", and in the proof of Proposition 6.8, which invokes the "(n, 2n)-wait" (3, 2^{3n-1}) to justify M(n) >= 2n. However, for n = 1, the hand (3, 2, 2) is an (n, 2n)-wait only if 2n = 2 (i.e. m = 2), yet the proof of Proposition 6.8 checked (3, 2^{3n-1}) as a minimal full-sided (n, 2n)-wait for n >= 2. Actually, wait: for n = 1, 2n = 2, so (3, 2, 2) has size 7 = 3(2) + 1, which is a (1, 2)-wait! Let's check the size of (3, 2^{3n-1}) for general n: its size is 3 + 2(3n - 1) = 6n + 1 = 3(2n) + 1. However, for n >= 2, (3, 2^{3n-1}) is NOT a full-sided wait at all! By Proposition 5.3 (Interior positivity), every full-sided wait has c_i >= 1, and by Proposition 4.2, S(c) = 3(1) + 2(2 + 3 + ... + 3n) = 3 + 2(3n(3n+1)/2 - 1) = 1 + 3n(3n+1) ≡ 1 (mod 3), so the fixed class is s = 1. For n >= 2, drawing rank 3n (which is in class 0) must use a pair in class -1 - 0 ≡ 2 (mod 3). After removing a pair at some rank j ≡ 2 (mod 3) from c + e_{3n}, the resulting hand has multiplicity 0 at j (since all ranks > 1 have multiplicity 2 in c). By Proposition 5.1, the hand splits into c_L and c_R at j. The left part c_L has ranks 1, ..., j-1 with multiplicities (3, 2, ..., 2), which has size 3 + 2(j-2) = 2j - 1 ≡ 2(2) - 1 = 3 ≡ 0 (mod 3), while its rank sum is 3(1) + 2((j-1)j/2 - 1) = 1 + j(j-1) ≡ 1 + 2(1) = 3 ≡ 0 (mod 3), wait: if j = 2, c_L = (3), which is t_1, and c_R = (0, 0, 2, 2, ..., 2, 3). For c_R to be meldable starting at rank 3 with multiplicity 2, the canonical scan needs 2 sequences s_3, consuming 2 tiles at ranks 3, 4, 5, leaving 0 at rank 5! Then at rank 6, it starts 2 sequences s_6, consuming 6, 7, 8, and so on, leaving 2 sequences s_{3n-3} at 3n-3, 3n-2, 3n-1, which leaves just the 3 tiles at 3n (a triplet t_{3n})! What if we draw rank 2 (class 2)? It needs a pair in class 0, say at rank j = 3k. Then at rank 1 we have 3 tiles, at rank 2 we have 3 tiles. If j = 3, c + e_2 - 2e_3 = (3, 3, 0, 2, 2, ..., 2), which splits at 3 into (3, 3) — NOT meldable! If j = 3k > 3, c_L = (3, 3, 2, 2, ..., 2) up to 3k-1. The canonical scan on c_L: at rank 1, σ_1 = 0, τ_1 = 1; at rank 2, c_2 = 3 so σ_2 = 0, τ_2 = 1; at rank 3, c_3 = 2 so σ_3 = 2, which consumes 2 tiles from ranks 3, 4, 5, leaving 0 at 5; this continues in blocks of 3 up to 3k-1, so c_L IS meldable, and c_R = (2, 2, ..., 2) from 3k+1 to 3n has length 3(n-k), which splits into 2 copies of sequences! Wait—what if we draw rank 3 (class 0)? Pair must be in class 2, say j = 3k+2. Then c + e_3 - 2e_j has c_1=3, c_2=2, c_3=3 (if j > 2). Canonical scan: σ_1 = 0, σ_2 = 2 (consumes 2 from 2, 3, 4), leaving 1 at rank 3, so σ_3 = 1 (consumes 1 from 3, 4, 5), which requires 2+1 = 3 tiles at rank 4! Since c_4 = 2 (or 0 if j=4), τ_4 = -1 < 0, NOT meldable! And if j = 2, c + e_3 - 2e_2 = (3, 0, 3, 2, ..., 2), which splits at 2, and c_R = (0, 0, 3, 2, 2, ..., 2) has size 3 + 2(3n-3) ≡ 0 (mod 3), but rank sum ≡ 0 + 2(1 + 2 + 0 + ...) — wait: by Proposition 5.2, for a full-sided wait, NO rank can have 3 tiles Needed at rank 4 when only 2 are available. Even simpler: Proposition 5.6 proves that for ANY full-sided wait, |c|_x >= n + 1 and |c|_y >= n + 1, and Proposition 5.5 / 5.2 show why (3, 2^{3n-1}) fails on drawing rank 3 when n >= 2!

2. Unverified minimality and full-sidedness for lower-bound families — Major

For a lower bound, one can check that the n -hand (3 , 2 3 n − 1 ) is a minimal full-sided ( n, 2 n ) -wait, giving M ( n ) ≥ 2 n . A precise determination of M ( n ) requires more work. Using exhaustive computer search for small n , we found two families of minimal full-sided waits of even order n = 2 a + 2 and odd order n = 2 a + 3 ( a ≥ 0 ), u a = (2 , 4 , 6 , 3 , (5 , 4) 3 a , 2 , 2) v a = (2 , 4 , 6 , 3 , 3 , (5 , 4) 3 a +1 , 2 , 2) 18 which improve the lower bound to

In Proposition 6.8 and the subsequent discussion on pages 18-19, the lower bounds M(n) >= 2n and M(n) >= floor(9n/2) - 3 rely on the assertions that (3, 2^{3n-1}) and the families u_a and v_a are minimal full-sided waits for all n (and all a >= 0), but no proof or general verification of either full-sidedness or minimality is provided for these infinite families.