{"id":"f8680475-65a8-4312-8f8c-a490cd4fafd8","arxiv_id":"2502.05906","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In an M/M/1 queue where type-A customers have priority over type-B customers, the paper derives threshold strategies for joining and quitting at equilibrium and compares them with social optima.","lead":"This paper works out the equilibrium behavior of customers in a two-class priority queue where high-priority customers jump ahead and low-priority customers may quit. It also compares the selfish outcome with what a social planner would choose, both with and without the priority constraint.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5's global optimum rests on an unproven two-class LCFS extension; Eq. (5.1) omits B-dependent externalities.","rationale":"I checked the equilibrium part and found no concrete algebraic error: the random-walk reduction in Lemma C.2, the worst-case argument around Eq. (4.7), and Claim 4.3 are internally consistent. In particular, re-deriving G'(n) from Eq. (4.26) reduces the monotonicity assertion to 1 - x(1 - ln x) >= 0 for x > 0, as the paper states. The weakest point is the global-optimization section. The reader identified the same unproven LCFS extension, and I agree that this is the single most load-bearing concern: the social-optimum formulas (5.1)-(5.2) are presented as proven results, but the only justification offered is an analogy to Hassin (1985) for a single class. Because this is a missing proof rather than a demonstrated contradiction, I would not move the verdict to REJECT; the equilibrium contribution appears credible on its own. However, the social-optimum claims should be either supplied with a real proof or explicitly downgraded to conjectures. The conditional verdict already captures this position, so no adjustment is needed.","tokens_in":19416,"tokens_out":18177,"duration_ms":184890,"concrete_test":"Solve the social-planner admission-control problem numerically for a truncated two-class M/M/1 queue by average-reward dynamic programming (or linear programming) for parameters such as mu=1, lambda_A=0.2, lambda_B=0.4, R_A=10, C_A=1, R_B=9, C_B=1, allowing balking and forced reneging. Compare the optimal policy's A-acceptance region with Eq. (5.1). Then vary lambda_B or C_B while keeping R_A/C_A > R_B/C_B; if the optimal A-threshold shifts, Eqs. (5.1)-(5.2) are not the global optimum. An analytical fallback: derive the Bellman optimality equations for the two-class admission control and check directly whether the marginal value of admitting an A-customer contains terms involving the B-state variables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The equilibrium characterization in Section 4 is supported by explicit random-walk computations and the monotonicity Claim 4.3, and I found no internal inconsistency there. The load-bearing gap is Section 5. The paper asserts that an unconstrained social planner favors the class with higher R_theta/C_theta and then, using Hassin's LCFS device, concludes that the optimal A-threshold M_A^* solves the single-class equation (5.1) with rho_A only, and that the B-threshold solves (5.2) with aggregate rho. This conclusion requires that the marginal social value of admitting an A-customer be independent of the presence of B-customers. That is not a consequence of the single-class LCFS argument. In a two-class FCFS system, admitting an A-customer changes the waiting times of B-customers and vice versa; the social cost of an A admission should depend on lambda_B, R_B, C_B, and the current B population. Equation (5.1) contains none of these terms. The sentence 'This is the technique used by Hassin (1985)' is not a proof: Hassin's result equates private and social incentives under LCFS within a homogeneous population, and no derivation is given for the heterogeneous two-class case. Since the abstract advertises a comparison of equilibrium and social optimum, this unproved extension is load-bearing. Section 6 inherits the same gap in its first paragraph, where it claims that for R_B/C_B < R_A/C_A the B-planner's optimum coincides with the global optimum of Section 5.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies an observable M/M/1 queue with two customer classes, A and B, where A-customers have priority over B-customers. Customers may balk upon arrival and may renege while waiting. The authors first analyze a semi-strategic model in which only B-customers are strategic, obtaining a threshold M_B^eq for B-customers. They then analyze the fully strategic model, claiming that A-customers follow the Naor threshold M_A = floor(R_A mu / C_A), while B-customers use a threshold T_B = M_A + V_B on the total number of customers in the system, with V_B given explicitly in Eq. (4.6). The paper then considers a globally optimizing social planner without priority constraints and derives thresholds given in Eqs. (5.1)-(5.2), and finally sketches a class-wise optimization problem in Section 6.","tokens_in":19742,"tokens_out":8143,"duration_ms":92812,"significance":"If the equilibrium characterization is fully correct, it is a valuable and elegant contribution: it reduces a seemingly two-dimensional strategic queueing problem to a one-dimensional random walk and gives closed-form equilibrium thresholds for both classes. The derivation is parameter-free in the sense that all quantities are expressed in terms of the primitives (arrival rates, rewards, costs, service rate). The semi-strategic and fully strategic B-threshold calculations are internally consistent, and the monotonicity claim in Claim 4.3 is checked. However, the social-optimum sections are not developed with the same rigor. The global-optimum result in Section 5 rests on an unproved extension of Hassin's LCFS equivalence to a heterogeneous two-class system, and Section 6 stops at a set of linear equations without proving the claimed optimal-strategy characterization. Since the abstract advertises a comparison of equilibrium and social optimum, this gap is load-bearing.","major_comments":[{"comment":"The proof establishes that B-customers join when n < T_B, but it does not prove that they balk when n >= T_B. The threshold V_B is derived from the worst-case state (M_A, n_B), and the text then asserts that 'the number of customers in the system cannot exceed M_A + V_B'. For a state with total n = T_B but n_A < M_A, the queue has more B-customers and fewer A-customers than the worst-case state, so the payoff comparison is not immediate; the assertion requires a monotonicity argument that is not supplied. This is a load-bearing step for the claimed unique threshold equilibrium.","section":"Theorem 4.2, paragraphs after Eq. (4.6)"},{"comment":"The global social optimum is derived by asserting that the planner favors the class with higher R_theta / C_theta and then invoking Hassin's LCFS technique. The sentence 'This is the technique used by Hassin (1985)' is not a proof that the technique extends to a two-class system with different rewards and costs. In particular, Eq. (5.1) contains no B-dependent terms, yet admitting an A-customer can delay B-customers and thereby change the social cost of an A admission. The paper needs either a proof of the claimed extension or a statement of the additional assumptions under which Eqs. (5.1)-(5.2) are valid.","section":"Section 5, Eqs. (5.1)-(5.2)"},{"comment":"The claim that Belinda-Paris's optimal strategy 'coincides with the equilibrium strategy in a LCFS regime' is asserted without proof, and the claim that for R_B/C_B < R_A/C_A this strategy coincides with the global planner's strategy is also asserted without proof. The subsequent analysis stops at linear systems for the hitting probabilities eta(i,j) and expected times kappa(i,j); it does not solve these systems, does not give a threshold characterization, and does not state a theorem for the B-planner's optimum. As written, Section 6 does not deliver a complete result.","section":"Section 6, first two paragraphs and Eqs. (6.4)-(6.6)"}],"minor_comments":[{"comment":"There are several typos: 'first-co me' in the abstract, 'equilibroium threshold' in Section 4.1, and 'does not case about B-customers' in Remark 4.4.","section":"Abstract and Section 4.1"},{"comment":"The proof states that the function gamma is 'unbounded and increasing' but does not show the monotonicity; a one-line difference argument would make this self-contained.","section":"Theorem 3.2 proof"},{"comment":"The sentence 'If n_A < M_A^*' should specify what happens when n_A >= M_A^*, since the planner's admission rule for A-customers is part of the claimed optimum.","section":"Section 5, paragraph after Eq. (5.2)"},{"comment":"The axes in the random-walk figures are not labeled; labeling n_A, n_B, M_A, and V_B directly in the figures would improve readability.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The equilibrium section is the main substantive contribution and appears salvageable, but the balking direction in Theorem 4.2 needs to be completed, and Sections 5-6 need substantial reworking or explicit qualification as conjectures. The paper is within the scope of a queueing/game-theory journal, and I do not see grounds for rejection if the authors can close these gaps."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I just read arXiv:2502.05906. The equilibrium part is the real contribution; the social-optimum part is not ready as written.\n\nThe fully strategic model gives a clean threshold for B-customers: join if total n is below TB = MA + VB, with VB in closed form. The reduction of the two-dimensional random walk to a one-dimensional ruin problem is standard but handled well, and the formulas check out. The paper is also honest that the equilibrium threshold is parametrically equivalent to Afeche and Sarhangian (2015); the new content is the explicit fully strategic treatment with reneging and the comparison with social optima.\n\nThe soft spot is Section 5. The planner's optimal A-threshold in (5.1) is asserted to be the single-class Naor optimum, independent of B's parameters. That follows only if the LCFS equivalence from Hassin (1985) extends to a two-class system where classes have different rewards and costs. The paper says \"This is the technique used by Hassin (1985)\" and stops. Hassin's result is about a homogeneous population; in a two-class FCFS system, admitting an A-customer changes the waiting time of B-customers, so the social cost of an A admission should depend on lambda_B, R_B, C_B. None of those appear in (5.1). The same gap shows up at the start of Section 6, where \"along the lines of Hassin (1985)\" is used again. This is load-bearing because the abstract promises a comparison between equilibrium and social optimum.\n\nSmaller issues: the proof of Theorem 4.2 shows that a B-customer joins when n < TB but does not explicitly prove she balks at n >= TB; it's likely fixable with the monotonicity of G, but as written the balking side is underived. The \"unique equilibrium\" claim is not backed by a formal uniqueness proof covering all strategies.\n\nIf the equilibrium section were the whole paper, I'd take it as is. As it stands, the social-optimum claims need either a real proof or a clear restriction to the case where the LCFS argument actually works. I'd send it to a serious referee, but with the expectation of major revision. The equilibrium result alone justifies the referee time.","headline":"The equilibrium characterization is solid and worth citing; the social-optimum section leans on an unproven extension of Hassin's LCFS argument and needs major revision.","tokens_in":20259,"tokens_out":3682,"would_cite":true,"duration_ms":39289,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K25","91A10"],"pacs":[],"model":"deepseek-v4-flash","headline":"In a strategic two-class priority queue, low-priority customers follow a single total-queue-length cutoff, and the cutoff has a closed form.","keywords":["M/M/1 queue","priority classes","balking","reneging","threshold strategy","Markov perfect equilibrium","social optimum","gambler's ruin"],"falsifier":"For the social-optimum claim, compute the truly optimal admission policy by dynamic programming on a small finite state space, for instance with $\\mu=1$, $\\lambda_A=0.3$, $\\lambda_B=0.4$, and $R_A/C_A > R_B/C_B$, and compare it with equations (5.1)--(5.2); any instance where the planner optimally admits an A-customer beyond $M^*_A$ or violates the total-cutoff structure would settle that the LCFS-based optimum is not the true optimum.","tokens_in":19222,"feed_emoji":"⏳","tokens_out":12141,"duration_ms":107435,"temperature":0.7,"pith_summary":"This paper studies an observable M/M/1 queue in which customers split into two classes, A and B, with A holding priority over B, and every customer may decide whether to join and later whether to renege. The central claim is that the equilibrium is unique and threshold-based: A-customers use the classic Naor threshold $M_A = \\lfloor R_A \\mu / C_A \\rfloor$ and never renege, while B-customers join when the total number in the queue is below a cutoff and renege once that cutoff is exceeded. The B cutoff is either the lower value solving (4.3) or, when $R_B/C_B$ is large enough, $T_B = M_A + V_B$ with $V_B$ given by the closed formula (4.6). The argument works because the two-dimensional state $(n_A,n_B)$ ahead of a tagged B-customer reduces to a one-dimensional gambler's ruin, so service probability and expected waiting time have closed forms. The paper further claims a priority-free social optimum with Naor-style thresholds for both classes and gives a system for the two-planner case.","feed_headline":"One cutoff decides when low-priority customers leave a priority queue","feed_subtitle":"Low-priority customers decide by total queue length alone; high-priority customers follow the classic threshold.","key_machinery":"The load-bearing object is a birth-and-death process $Y_t$ that counts the number of people ahead of a tagged B-customer, observed only at the epochs when either an A-customer arrives or a service completes. Each step thus rises with probability $\\lambda_A/(\\lambda_A+\\mu)$ and falls with probability $\\mu/(\\lambda_A+\\mu)$, which is exactly a gambler's ruin; because A's arrivals are capped at $M_A$ and B-customers cannot pass one another, the two-dimensional state $(n_A,n_B)$ collapses to the total queue length $n = n_A+n_B$. The gambler's-ruin probability (B.1) and expected duration (B.2) give the tagged customer's expected payoff, and the equilibrium threshold is the largest $n$ for which that payoff is nonnegative even in the worst case $(M_A, n_B)$. This reduction is what turns a dynamic game with balking and reneging into explicit algebraic cutoff formulas.","core_discovery":"On the paper's own terms, the discovery is that the equilibrium of a fully strategic priority queue is completely described by three formulas. A-customers are oblivious to B-customers, so they inherit Naor's single-class threshold $M_A = \\lfloor R_A \\mu / C_A \\rfloor$ and never renege. If a B-customer's reward-to-cost ratio is low ($R_B/C_B < E(M_A,0)$), B-customers join only while the total queue length is below the floor of the solution of equation (4.3). If $R_B/C_B \\geq E(M_A,0)$, B-customers join whenever $n = n_A + n_B < T_B$ and renege when the total exceeds $T_B$, where $T_B = M_A + V_B$ and $V_B$ is the integer in (4.6). The same gambler's-ruin machinery yields the semi-strategic threshold (3.5), and the paper uses Hassin's last-come-first-served device to propose social-optimum thresholds (5.1)--(5.2) and a trapezoid first-exit system for the two-planner case.","pith_inferences":["The composition-insensitivity of B behavior is directly testable: in a laboratory or simulated queue, fixing total length and varying the A/B split should leave low-priority join and renege decisions unchanged.","Because the B threshold is driven only by $\\rho_A$ and $M_A$, a queue manager could estimate admission cutoffs from high-priority traffic statistics without ever measuring the low-priority arrival process.","The same gambler's-ruin reduction is unlikely to survive three or more priority classes, since each additional class adds a boundary; the paper's own conclusion says this explicitly, so the closed-form equilibrium thresholds should be read as a two-class phenomenon.","The Section 5 planner formulas rest on the asserted LCFS equivalence; turning that assertion into a proof, or finding a counterexample, is the natural next step."],"forward_implications":["A-customers' equilibrium strategy is exactly Naor's single-class threshold, so adding a lower-priority class does not change high-priority admission or waiting behavior.","B-customers' equilibrium decisions depend on the total queue length $n = n_A + n_B$, not on how many of those customers are A or B; the cutoff is $T_B = M_A + V_B$ in the high-reward regime and the solution of (4.3) in the low-reward regime.","A B-customer who has joined reneges as soon as the total queue length exceeds the threshold, so abandonment is predictable from a single observable.","A social planner without priority constraints who values A more per unit cost will admit at most the Naor optimum $M^*_A$ A-customers and cap total admissions by the Naor threshold computed with combined traffic load $\\rho=(\\lambda_A+\\lambda_B)/\\mu$.","With one planner per class, the A-planner's threshold is again the Naor optimum, while the B-planner's problem becomes the first-exit problem of a genuinely two-dimensional random walk from a trapezoid."],"supporting_citations":[{"why":"Supplies the single-class balking equilibrium threshold $M_A = \\lfloor R_A \\mu / C_A \\rfloor$ that A-customers inherit, and the social-optimum benchmark for the planner sections.","marker":"Naor (1969)"},{"why":"Supplies the last-come-first-served equivalence used to convert the social planner's admission decision into an individual last-customer decision in Sections 5 and 6.","marker":"Hassin (1985)"},{"why":"Supplies the gambler's-ruin formulas, Theorems B.1 and B.2, for the probability of service and expected waiting time of a tagged B-customer.","marker":"Ethier (2010)"},{"why":"Supplies the Markov-chain hitting-time linear systems used in Theorem C.1 to set up the B-planner's trapezoid problem.","marker":"Norris (1997)"},{"why":"Closest prior model of two priority classes with strategic balking and reneging, against which the paper presents a simpler equilibrium threshold expression.","marker":"Afèche and Sarhangian (2015)"},{"why":"Provides the strategic-queueing lemmas whose structure the proofs of Lemma 3.1 and Theorem 3.2 follow.","marker":"Hassin and Haviv (2003)"}],"fun_headline_variants":["Three formulas fully capture strategic priority queue equilibria","Low-priority joiners decide by total queue length alone, says model","Strategic queue behavior: high-priority blind to low, low use total length","One cutoff for low-priority reneging in strategic priority queues","Priority customers ignore the rest; low-priority follow one simple rule"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The social-optimum section's thresholds hold only if the last-come-first-served trick, which makes an individual's optimal decision socially optimal in a single class, still works when two classes with different rewards and costs share the queue; the paper asserts this extension rather than proving it.","fun_headline_variants_meta":{"raw":{"variants":["Three formulas fully capture strategic priority queue equilibria","Low-priority joiners decide by total queue length alone, says model","Strategic queue behavior: high-priority blind to low, low use total length","One cutoff for low-priority reneging in strategic priority queues","Priority customers ignore the rest; low-priority follow one simple rule"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000781,"raw_usage":{"total_tokens":3408,"prompt_tokens":858,"completion_tokens":2550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":2460}},"tokens_in":474,"tokens_out":2550,"duration_ms":17788,"temperature":1.0,"reasoning_tokens":2460,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:29:07.978935+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the social-optimum claim, compute the truly optimal admission policy by dynamic programming on a small finite state space, for instance with $\\mu=1$, $\\lambda_A=0.3$, $\\lambda_B=0.4$, and $R_A/C_A > R_B/C_B$, and compare it with equations (5.1)--(5.2); any instance where the planner optimally admits an A-customer beyond $M^*_A$ or violates the total-cutoff structure would settle that the LCFS-based optimum is not the true optimum.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the single-class balking equilibrium threshold $M_A = \\lfloor R_A \\mu / C_A \\rfloor$ that A-customers inherit, and the social-optimum benchmark for the planner sections."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the last-come-first-served equivalence used to convert the social planner's admission decision into an individual last-customer decision in Sections 5 and 6."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the gambler's-ruin formulas, Theorems B.1 and B.2, for the probability of service and expected waiting time of a tagged B-customer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Markov-chain hitting-time linear systems used in Theorem C.1 to set up the B-planner's trapezoid problem."},{"cited_title":"and Haviv, M","cited_arxiv_id":null,"evidence_quote":"Provides the strategic-queueing lemmas whose structure the proofs of Lemma 3.1 and Theorem 3.2 follow."}],"review_version":1}