{"id":"6a0313ac-54c5-4555-b7a8-b76c63e0d5d3","arxiv_id":"1908.02547","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"For a three-unit cold-standby system with a regular and an expert repairer, the authors derive exact limiting availability and profit for four policies and recommend a deterministic patience time with the expert repairing all failed units.","lead":"This paper derives formulas for long-run availability and profit of a machine with two spare units, repaired by a regular worker or a faster expert. It compares four maintenance policies and recommends when the expert should repair all failed units and how long to wait before calling her.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DPT models are not semi-Markov: residual patience T-X is history-dependent and omitted from the six-state chain, so Theorem 4.1 and the Model 3/4 formulas (4.18)-(4.23) are unsupported.","rationale":"The preprint's goal is to extend the one-spare analysis of [2] by deriving exact limiting availability and profit for four repair policies and using those formulas to rank the policies. That central claim is sound only if each policy's embedded chain is a Markov chain on the six states listed in Section 3. For RPT models, the exponential memorylessness of the patience time justifies the six-state SMP, and those portions appear internally consistent. For DPT models, however, the residual patience T-X is a function of the age X at which State 4 was entered, so the next transition from State 4 depends on history. The paper's own Section 6(ii) acknowledges that the Markovian property fails under the DPT policy. Section 4.3's P45 computation also mishandles the conditioning: it averages over the unconditional X density and conditions only on X<T, not on X<min(Y,T). Conditioning correctly gives a different P45; for the Section 5 parameters the difference is material. Because the paper's headline comparisons (MRE-DPT best for availability, DPT better than RPT for profit at a tuned T) rest on Models 3 and 4, the central claim is not supported as written. This is an internal-consistency problem rather than a disagreement with the literature, and it is repairable by augmenting the state space or using a different regenerative argument. I agree with the reader's assessment and would keep the reject verdict.","tokens_in":12738,"tokens_out":13390,"duration_ms":135791,"concrete_test":"Simulate the exact MRE-DPT process as a continuous-time event simulation with deterministic patience T at the Section 5 parameter values: lambda=0.5, beta=0.35, gamma=0.75, T=1.62. Record, for each entry into State 4, the age X and the subsequent transition. Check whether the empirical transition probability from State 4 to State 5 equals the paper's P45 = lambda(e^{-lambda T} - e^{-(lambda+beta)T})/beta and whether that transition probability depends on the recorded age X. If either check fails, or if the simulated limiting availability differs from Eq. (4.20) beyond sampling error, the six-state SMP derivation is invalid and the DPT comparison in Section 5 is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim relies on Theorem 4.1 for all four models. For DPT models, the embedded six-state chain is not a Markov chain. When the chain enters State 4 from State 2, the remaining patience time is T'=T-X for the realized lifetime X in State 2. Under deterministic T, T' is not exponential and is not encoded in the state; the transition probabilities out of State 4 (P42, P45, P46 in (4.18)/(4.21)) therefore depend on the history X. Consequently, the Markov renewal property fails and Theorem 4.1 does not apply to Models 3 and 4. The error is visible in the Section 4.3 derivation: P45 is written as P{T' < min{X',Y'} | X<T} and evaluated as the integral over [0,T] of e^{-(lambda+beta)(T-x)} lambda e^{-lambda x} dx, which is not a conditional probability (no denominator) and conditions on X<T rather than on X<min(Y,T), the event that actually sends the chain to State 4. Conditioning properly on that entry event yields P45 = (lambda+beta) T e^{-(lambda+beta)T} / (1 - e^{-(lambda+beta)T}), not the expression used in the paper. Since the Section 5 comparisons and the claimed optimal DPT range [1.45, 1.62] rely on these models, the DPT-based conclusions are unsupported. The paper itself concedes in Section 6(ii) that the Markovian property fails under the DPT policy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript extends the one-spare-unit system of Bieth, Hong, and Sarkar [2] to a one-unit system supported by two cold-standby spares, serviced by a regular in-house repairer and a visiting expert repairer. Four policies are considered, formed by crossing single versus multiple repair by the expert (SRE versus MRE) with random versus deterministic patience time (RPT versus DPT). The authors model the system as a six-state semi-Markov process (SMP), derive the stationary state proportions in closed form, and use them to compute the limiting availability A_inf and limiting profit per unit time omega for each model. The main advertised conclusions are that MRE dominates SRE in availability, that for suitable deterministic patience time the DPT policy dominates the RPT policy in profit, and that for the numerical example the MRE-DPT model is optimal for patience times in [1.45,1.62].","tokens_in":12997,"tokens_out":8555,"duration_ms":89587,"significance":"If the derivations were correct, the paper would provide a useful extension of [2] and a concrete maintenance-engineering decision rule. The RPT models (Sections 4.1 and 4.2) are standard SMP calculations and appear internally coherent; the extension from one to two spare units is natural, and the explicit algebraic formulas are a strength. However, the DPT models (Sections 4.3 and 4.4) are not valid SMPs: the residual patience time is not Markovian in the six-state chain, and the transition probabilities out of State 4 are derived using an incorrect conditioning. Since the central profit comparison, including the claimed optimal DPT range, depends on these models, the main new conclusions are unsupported. The paper also makes useful numerically illustrated comparisons, but no code is provided.","major_comments":[{"comment":"The justification that the process is an SMP applies only to the RPT models. Under a deterministic patience time T, the remaining patience when the system enters State 4 is T' = T - X, where X is the already-realized lifetime in State 2; this quantity is not exponentially distributed and is not part of the six-state description. Consequently the transition probabilities out of State 4 depend on the history X, the embedded six-state chain is not a Markov chain, and Theorem 4.1 cannot be applied to Models 3 and 4. The paper itself notes in Section 6(ii) that the Markovian property fails under the DPT policy in a related setting. Because the formulas (4.18)-(4.23) and the Section 5 conclusions, including the optimal range [1.45,1.62], rest on this application, the DPT-based central claims are not supported.","section":"Section 4, Theorem 4.1 and Sections 4.3-4.4"},{"comment":"The derivation of P45 conditions on the wrong event and uses the wrong density. State 4 is entered from State 2 when X < min(Y,T), not merely when X < T, and the conditional density of X given this entry event is (lambda+beta)e^{-(lambda+beta)x}/(1-e^{-(lambda+beta)T}) for x in [0,T], not the unconditional density lambda e^{-lambda x}. The displayed integral has no conditioning denominator and therefore yields the incorrect value lambda e^{-lambda T}(1-e^{-beta T})/beta. Conditioning correctly gives P45 = (lambda+beta)T e^{-(lambda+beta)T}/(1-e^{-(lambda+beta)T}). Since P42 and P46 are then defined from this P45, the transition probabilities and the stationary distribution (4.19) are incorrect.","section":"Section 4.3, Eq. (4.18), derivation of P45"},{"comment":"In the SRE model the transition matrix sends State 6 to State 4, so State 4 is reached not only from State 2 (with residual patience T-X) but also after an expert visit. The paper simply reuses the MRE-DPT values of P42, P45, and P46 in equation (4.21). These two entry paths have different patience-time clocks unless the model explicitly specifies that patience restarts and unless the state label encodes the entry type. Without such an augmentation, the transition probabilities out of State 4 are path-dependent, and the stationary distribution (4.22) together with the SRE-DPT values of A_inf and omega are unsupported.","section":"Section 4.4, Eq. (4.21), SRE-DPT model"}],"minor_comments":[{"comment":"The affiliation line contains a typo: \"Indiana Universiry-Purdue University Indianapolis\" should read \"Indiana University-Purdue University Indianapolis.\"","section":"Title page / affiliation"},{"comment":"The statement that Wald's First Identity gives the expected number of expert visits per unit time as the reciprocal of the cycle length is imprecise; the renewal reward theorem is the standard result for the long-run rate of renewals, and citing it would be more accurate.","section":"Section 3, Eq. (3.2)"},{"comment":"There is a typo: \"th opposite\" should be \"the opposite.\"","section":"Section 5, item after Figure 4"},{"comment":"The numerical comparisons are presented for a single parameter set without supporting code or tabulated values; since the paper's main claim is a comparison across policies, it would be helpful to state explicitly whether the figures are generated from the displayed closed forms or by simulation, and ideally to provide the data or code used.","section":"Sections 5, Figures 2 and 3"}],"recommendation":"reject","confidential_remarks":"The DPT part of the paper cannot be repaired by a local correction: a proper treatment would require either enlarging the state space to include the residual patience time, which is a continuous variable and changes the nature of the analysis, or restricting the paper to RPT policies, which would remove the headline DPT-over-RPT conclusion. I would therefore recommend rejection, though a substantially revised RPT-only version could be of interest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: the paper is a genuine but modest extension of Bieth-Hong-Sarkar's one-spare model to two spares, and the RPT half is mostly careful work. The DPT half is wrong in a load-bearing way, and the paper's own conclusion—that a tuned deterministic patience time beats exponential patience—rests on that half.\n\nWhat's new and good: the state space grows from the one-spare model, and the recursive cycle-time equations adapt cleanly to MRE and SRE under exponential patience. The derivations in Sections 4.1 and 4.2 look coherent; I spot-checked the stationary structure and it is consistent. The comparison figures also make sense for the RPT models. Credit where due: it is clearly written and the algebra is checked.\n\nThe soft spot is not minor. Models 3 and 4 assume the embedded six-state chain is Markovian when the patience time is deterministic, but the residual patience T' = T - X depends on the lifetime X that just occurred. That history is not in the state. So the transition probabilities out of State 4 depend on the age of the regular repair, and Theorem 4.1 does not apply. The P45 derivation in Section 4.3 confirms the problem: it averages the residual patience over the unconditional density of X, rather than conditioning on X < min(Y,T), the event that actually puts the system in State 4. Correcting that conditioning gives a different formula, and even then the state is still not sufficient. The paper tips its hand in Section 6(ii), where it says the Markovian property fails under DPT—that is true for the current model, not just the future two-facility version.\n\nSo the RPT models stand, the DPT models fall, and the central practical claim (\"MRE-DPT is best for T in [1.45,1.62]\") is unsupported. The paper could be repaired by dropping the DPT models or reworking them with an augmented state or a regenerative analysis, but as submitted it needs major revision.\n\nWho this is for: reliability engineers who want the two-spare RPT formulas, and readers who want a cautionary example of why residual times must be in the state. It deserves a serious referee—the RPT half is sound and worth having—but I would reject the current version. Send it to peer review only if the editor expects heavy revision; otherwise desk reject and invite a resubmission with DPT fixed.","headline":"The two-spare random-patience models are a solid but modest extension; the deterministic-patience models have a load-bearing Markovianity flaw that sinks the paper's headline recommendation.","tokens_in":13564,"tokens_out":3820,"would_cite":false,"duration_ms":38302,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K15","60K20","90B25"],"pacs":[],"model":"deepseek-v4-flash","headline":"For all four repair policies, the long-run availability and profit of a two-spare repairable system are derived exactly from a six-state semi-Markov process.","keywords":["cold standby","repairable system","semi-Markov process","limiting availability","limiting profit per unit time","patience time","multiple repair by expert","single repair by expert"],"falsifier":"Compare the paper's $P_{45}$ with the value obtained by conditioning on the actual event that the system enters State 4, namely $X < \\min(Y,T)$, rather than only on $X < T$; a discrepancy in the transition probability, or a mismatch between simulated $\\theta_6$ and equations (4.20) or (4.23), would show that the deterministic-patience models are approximations.","tokens_in":12453,"feed_emoji":"🔧","tokens_out":9270,"duration_ms":90293,"temperature":0.7,"pith_summary":"This paper studies a single operating unit backed by two identical cold-standby spares. When the operating unit fails, the regular repairer works on it until either repair completes, a patience time expires, or the system goes down; if the patience time expires or the system fails, a faster but more expensive expert repairer is called. The central claim is that for all four policies—single versus multiple expert repairs, and random versus deterministic patience time—the long-run fraction of time the system is up and the long-run profit per unit time are given exactly by the stationary formulas of a six-state semi-Markov process under exponential lifetimes and repair times. If the claim is right, a maintenance engineer can compute the best patience time and expert workload from the model parameters, and the numerical comparisons show that repairing all failed units beats repairing one per visit on availability while a suitably tuned deterministic patience time can beat a random one on profit.","feed_headline":"Two spares: exact uptime and profit formulas for four policies","feed_subtitle":"Semi-Markov analysis gives long-run uptime and profit for each repair policy, so engineers can tune the patience time.","key_machinery":"The central object is the six-state semi-Markov process, a process that moves between states by a Markov chain after random sojourn times, with states recording how many units are operating, on standby, under regular repair, under expert repair, or awaiting repair. The load-bearing identity is $\\theta_k = \\pi_k \\mu_k / \\sum_{j=1}^6 \\pi_j \\mu_j$, which converts the embedded chain's stationary probabilities and mean sojourn times into long-run occupancy proportions; the paper then computes availability as $A_\\infty = 1 - \\theta_6$ and profit per unit time as $\\omega = A_\\infty(R_p-C_p) - [\\Theta_r C_r + \\Theta_e C_e + C_l/\\tau]$, with $\\tau$ obtained from recursive expected-cycle equations.","core_discovery":"Under exponential life and repair times (lifetime rate $\\lambda$, regular repair rate $\\beta$, expert repair rate $\\gamma$, patience rate $\\alpha$), the system is modeled as a semi-Markov process whose embedded chain has six states recording how many units are operating, on standby, under regular repair, under expert repair, or awaiting repair. The paper derives the transition matrix $P$, the stationary distribution $\\pi$ of the embedded chain, and the mean sojourn times $\\mu_k$ for each of the four models, and applies $\\theta_k = \\pi_k \\mu_k / \\sum_j \\pi_j \\mu_j$. This yields $A_\\infty = 1 - \\theta_6$ and the profit rate $\\omega = A_\\infty(R_p-C_p) - (\\Theta_r C_r + \\Theta_e C_e + C_l/\\tau)$, where $\\tau$ is the expected cycle time obtained by solving recursive relations and $C_l/\\tau$ follows from Wald's identity. The paper concludes that multiple-repair-by-expert dominates single-repair-by-expert in availability, that adding a second spare raises both availability and profit, and that for any cost parameters there is a threshold expert rate separating when MRE or SRE is more profitable, along with an optimal patience time under the deterministic-patience policy.","pith_inferences":["A natural extension not pursued in the paper is to re-derive the deterministic-patience models with an augmented state recording the remaining patience time; comparing those formulas with the paper's closed forms would show how much accuracy the six-state approximation sacrifices.","The same semi-Markov setup should extend to three or more spares by adding states for queues of failed units, with availability approaching one but profit eventually bounded by repair and trip costs.","A concrete numerical check would simulate the four models at the paper's example parameters to confirm that MRE-DPT maximizes both $A_\\infty$ and $\\omega$ in the patience interval $[1.45, 1.62]$ and that $\\omega$ peaks near $T=2.19$; this would also expose any bias introduced by the deterministic-patience Markov assumption."],"forward_implications":["The paper concludes that MRE yields higher limiting availability than SRE for both random and deterministic patience times, so if uptime is the priority and the expert can be afforded, the expert should repair all failed units during a visit.","A system with two cold-standby spares has higher limiting availability and limiting profit than the same system with one spare across all four policies, which quantifies the value of the extra spare when availability falls below an acceptable threshold.","For any parameter choice there is a patience-time interval where the deterministic-patience policy dominates the random-patience policy in profit, and within the deterministic policy there is an optimal patience time $T$ maximizing $\\omega$, so an engineer who prefers the logistically simpler deterministic rule can use that interval.","There is a threshold on the expert's cost rate: below it MRE is more profitable than SRE, above it SRE is more profitable, and knowing this threshold decides how many failed units the expert should repair per visit."],"supporting_citations":[{"why":"Supplies the definition of limiting availability and Wald's First Identity, used to turn the expert trip charge into the per-unit-time cost $C_l/\\tau$ in the profit formula.","marker":"[1]"},{"why":"Defines the four repair-model policies for a one-spare system and gives the matching $T^*$ comparison; the present paper extends its derivations to two spares.","marker":"[2]"},{"why":"Supplies the semi-Markov process stationary-proportion theorem $\\theta_k = \\pi_k \\mu_k / \\sum_j \\pi_j \\mu_j$ and the embedded-chain equations that yield all formulas.","marker":"[16]"},{"why":"Earlier two-server model where the expert takes over only after the patience time expires; its expert-call rule is contrasted with the rule used here.","marker":"[12]"},{"why":"Earlier patience-time model that calls the expert when patience expires or the system fails, the call-in convention on which the present models are built.","marker":"[13]"}],"fun_headline_variants":["Two spares, two repairers: four policies compared","Semi-Markov model for repairable system with two spares","Optimize uptime and profit: repair policies with two spares","Expert repairer vs regular: best policy for profit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact formulas for the deterministic-patience models assume the six-state embedded chain remains Markovian after part of the patience time has already elapsed; if the remaining patience depends on that elapsed time in a way the state does not record, Models 3 and 4 are approximations rather than exact.","fun_headline_variants_meta":{"raw":{"variants":["Two spares, two repairers: four policies compared","Semi-Markov model for repairable system with two spares","Optimize uptime and profit: repair policies with two spares","Expert repairer vs regular: best policy for profit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000284,"raw_usage":{"total_tokens":1720,"prompt_tokens":1037,"completion_tokens":683,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":611}},"tokens_in":653,"tokens_out":683,"duration_ms":6518,"temperature":1.0,"reasoning_tokens":611,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:15:04.028739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the paper's $P_{45}$ with the value obtained by conditioning on the actual event that the system enters State 4, namely $X < \\min(Y,T)$, rather than only on $X < T$; a discrepancy in the transition probability, or a mismatch between simulated $\\theta_6$ and equations (4.20) or (4.23), would show that the deterministic-patience models are approximations.","supporting_citations":[{"cited_title":"An introduction to probability theory and its applications, volume 1","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of limiting availability and Wald's First Identity, used to turn the expert trip charge into the per-unit-time cost $C_l/\\tau$ in the profit formula."},{"cited_title":"A standby system with two types of repair persons","cited_arxiv_id":null,"evidence_quote":"Defines the four repair-model policies for a one-spare system and gives the matching $T^*$ comparison; the present paper extends its derivations to two spares."},{"cited_title":"Stochastic Processes","cited_arxiv_id":null,"evidence_quote":"Supplies the semi-Markov process stationary-proportion theorem $\\theta_k = \\pi_k \\mu_k / \\sum_j \\pi_j \\mu_j$ and the embedded-chain equations that yield all formulas."},{"cited_title":"Comparative study of the profit of a two server system including patience time and instruction time","cited_arxiv_id":null,"evidence_quote":"Earlier two-server model where the expert takes over only after the patience time expires; its expert-call rule is contrasted with the rule used here."},{"cited_title":"Stochastic behaviour of a two-unit standby system with two types of repairmen and patience time","cited_arxiv_id":null,"evidence_quote":"Earlier patience-time model that calls the expert when patience expires or the system fails, the call-in convention on which the present models are built."}],"review_version":1}