{"id":"ddca91ae-63df-43fe-ad76-f730fc14611e","arxiv_id":"2607.07650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":11,"one_line_summary":"Simulation of USSD authentication shows a non-linear 'Success Cliff' in session completion when stochastic SMS OTP blocking delay interacts with fixed session time budgets.","lead":"This paper simulates USSD mobile-money authentication workflows and finds that session success rates collapse sharply when an out-of-band SMS OTP delay is added to an already complex flow. The result matters for financial inclusion: stronger authentication may make mobile money unusable for rural feature-phone users.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Cliff detection is structurally guaranteed: blocking delay appears only at C4, so the largest success-rate drop will always occur at the C3→C4 transition regardless of delay distribution or detection threshold.","rationale":"The paper's architectural insight — that out-of-band blocking delay is more dangerous to time-constrained sessions than in-band complexity — is sound and practically relevant. The formalization of the Success Cliff (Eqs. 10–12) is a reasonable attempt to operationalize non-linear failure detection. However, the experimental design cannot substantiate the central claim of non-linear collapse because blocking delay is varied as a binary factor (present at C4, absent at C1–C3) rather than as a continuous variable. This means the cliff location is determined by experimental structure, not discovered empirically. The sensitivity analysis, while methodologically careful, tests parameters that cannot change the structural outcome. The reader's verdict of CONDITIONAL is appropriate, but the specific concern should be reframed: the issue is not merely that U(5,30) is unvalidated (though it is), but that no delay distribution or parameter variation can distinguish a genuine cliff from a proportional delay effect under the current binary design. The paper is honest about its simulation-only nature (Section IX-D) and does not overclaim field validation. The contribution stands as a simulation-based architectural argument, but the 'cliff' framing overstates what the design can demonstrate. A continuous sweep of blocking delay magnitude, as described in the concrete test, would settle whether the non-linearity is real or imposed by design. If such a sweep shows a smooth degradation curve, the paper's contribution would still hold as a warning about blocking delay in USSD, but the 'Success Cliff' as a distinct non-linear phenomenon would not be established.","tokens_in":10695,"tokens_out":2972,"duration_ms":213734,"concrete_test":"Fix the workflow at C4 and sweep the SMS delay distribution mean: run simulations with W_SMS ~ U(a, a+25) for a ∈ {0, 1, 2, 3, 5, 8, 10, 15, 20, 25}, holding all other parameters at baseline. Plot session success rate as a function of mean blocking delay. If the curve shows a sharp, non-linear collapse at a critical delay threshold (consistent with Eq. 12's δ and γ criteria applied to the delay dimension), the Success Cliff phenomenon is genuine. If the curve is smooth and approximately linear/monotonic, the 'cliff' is an artifact of introducing a large delay at a single discrete complexity level, and the finding reduces to 'adding 5–30 seconds of delay to a 120-second session causes proportional failure.'","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The reader correctly identifies the SMS blocking delay as the load-bearing area but focuses on the distribution parameters U(5,30). The deeper issue is structural: blocking delay (H_k > 0) is introduced exclusively at complexity level C4 (Table V). Since the Success Cliff definition (Eq. 12) identifies the complexity level with the largest consecutive drop Δ(C) and positive acceleration Γ(C), and since blocking delay is the single largest time perturbation in the system (5–30 seconds vs. sub-second to few-second components), the cliff will always be detected at C3→C4 by construction. The sensitivity analysis (Section VIII-E) confirms robustness across error rates and γ thresholds, but this robustness is trivially expected: no variation in error probabilities or detection thresholds can move the cliff away from the only level where blocking delay is introduced. The Q range of 0.000 across all 21 combinations is not evidence of a genuine threshold phenomenon — it is a direct consequence of binary experimental design. To distinguish a true non-linear collapse from a proportional effect of adding delay, one must vary the blocking delay magnitude continuously and show that success rate collapses sharply at a critical delay threshold rather than degrading smoothly. The paper does not do this. The multi-mechanism argument (Section VIII-D: concurrent session timeout and user response timeout) is a reasonable qualitative point, but it does not rescue the cliff detection from being tautological — any sufficiently large added delay in a time-constrained system will eventually trigger multiple timeout mechanisms. The question is whether this happens at a sharp threshold (cliff) or gradually, and the binary design cannot answer that.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper models failure dynamics in time-constrained USSD authentication systems, proposing a formal definition of a 'Success Cliff'—a sharp, non-linear collapse in session success rate. Through simulation, the authors vary authentication complexity, network round-trip time, and the presence of out-of-band blocking delay (SMS OTP). The central finding is that the cliff emerges specifically when stochastic blocking delay interacts with a fixed session time budget, rather than from authentication complexity or network latency alone. The experimental design is methodologically careful, with controlled one-factor-at-a-time variation and sensitivity analysis across error regimes and gamma thresholds.","tokens_in":11505,"tokens_out":1228,"duration_ms":178093,"significance":"The paper addresses a practically important problem for financial inclusion in USSD-based systems. Its strengths include a transparent simulation framework with clearly stated parameters anchored to documented operator thresholds and empirical KLM values, a formal and falsifiable definition of the Success Cliff (Eqs. 10–12), and a well-structured sensitivity analysis. The distinction between gradual degradation and non-linear collapse, and the identification of blocking delay as the necessary condition for the latter, is a useful architectural insight for authentication designers. However, the significance of the central claim is substantially undermined by a structural confound in the experimental design, detailed below.","major_comments":[{"comment":"The experimental design creates a structural near-tautology that undermines the central claim. Blocking delay (H_k > 0) is introduced exclusively at complexity level C4 (Table V). Since the Success Cliff definition (Eq. 12) identifies the complexity level with the largest consecutive drop Delta(C) and positive acceleration Gamma(C), and since blocking delay is the single largest time perturbation in the system (5–30 seconds vs. sub-second to few-second components), the cliff will always be detected at the C3-to-C4 transition by construction. The sensitivity analysis (Section VIII-E) confirms robustness across error rates and gamma thresholds, but this robustness is trivially expected: no variation in error probabilities or detection thresholds can move the cliff away from the only level where blocking delay is introduced. The Q range of 0.000 across all 21 combinations is not evidence of","section":null},{"comment":"Section III-F, Eq. (5): The SMS OTP delivery delay model W_SMS ~ U(5,30) is the load-bearing premise for the cliff phenomenon. The entire argument hinges on this blocking delay consuming enough of the per-step time budget to trigger concurrent timeout mechanisms. However, this distribution is stated as reflecting 'normal network conditions' but is not empirically validated. If real-world SMS delivery is faster (e.g., median 3–5 seconds rather than uniform 5–30), the cliff may not materialize at the severity reported. The authors acknowledge in Section IX-D that field measurement is needed, but the paper's framing presents the cliff as a robust structural finding without adequately foregrounding this dependency.","section":null},{"comment":"Section VIII-D: The paper claims that failure rates under blocking delay 'exceed the sum of their independent contributions,' implying super-additive interaction. However, because blocking delay is introduced only at C4, the comparison is between a configuration without blocking delay and one with it—this is a single design point, not a continuous manipulation. To distinguish a true non-linear collapse from a proportional effect of adding delay, the authors must vary the blocking delay magnitude continuously (e.g., W_SMS ~ U(a, b) with varying a, b) and show that success rate collapses sharply at a critical delay threshold rather than degrading smoothly. The current design cannot distinguish between these alternatives.","section":null}],"minor_comments":[{"comment":"Table I: The variable H_k is described as nonzero only for steps requiring delivery from a subsystem external to the active USSD session. It would help to explicitly note in the table caption or the table itself that H_k = 0 for C1–C3 and H_k > 0 only for C4, as this is central to interpreting the results.","section":null},{"comment":"Section VIII-D mentions 'complexity level 3.39 (C3)' and 'complexity 3.99 (C4).' These values are not clearly defined in the text or tables. Adding a column to Table V showing the computed complexity value C for each configuration would improve clarity.","section":null},{"comment":"Section VIII-D, paragraph on failure composition: the text notes that metrics sum to approximately 95% rather than 100%, leaving a residual attributable to user response timeout. It would be clearer to include user response timeout as an explicit metric in the evaluation metrics list (Section VII) and report it alongside the other timeout categories.","section":null},{"comment":"Section V-A: The KLM adaptation mentions standard operator values (M=1.35s, K=0.28s, H=0.40s) and states these are 'adapted to the feature phone numeric keypad context,' but the specific adaptation methodology is not described. Clarifying what was changed from standard KLM and why would strengthen reproducibility.","section":null},{"comment":"The paper uses 'Success Cliff' as a proper noun throughout. On first use (Abstract and Section I), a brief parenthetical noting that this is a term coined by the authors would be appropriate.","section":null},{"comment":"Section IX-C: The claim that SMS OTP-based step-up authentication is 'not yet widely deployed for feature phone users' is important for contextualizing the findings. A citation or brief evidence for this claim would strengthen this point.","section":null}],"recommendation":"major_revision","confidential_remarks":"The structural confound identified in the major comments is the primary concern. The paper's framing implies that the cliff is a general phenomenon, but the experimental design makes its detection at C3→C4 structurally guaranteed. The authors should be given the opportunity to address this by adding a continuous variation of blocking delay magnitude, which is within the scope of the existing simulation framework. If they can show a sharp threshold in success rate as delay increases, the contribution would be substantially strengthened. The paper is otherwise well-written and addresses a relevant problem for the journal's scope."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and substantive review. The referee identifies three interconnected concerns about our experimental design: (1) a structural confound wherein blocking delay is introduced only at C4, making cliff detection at the C3-to-C4 transition near-tautological; (2) insufficient empirical grounding for the SMS OTP delay distribution; and (3) the absence of continuous variation in blocking delay magnitude, which prevents distinguishing non-linear collapse from proportional degradation. We find these concerns largely valid and outline revisions below.","responses":[{"response":"The referee is correct that our experimental design introduces blocking delay only at C4, and that this creates a structural confound: the sensitivity analysis over error rates and gamma thresholds cannot, by construction, move the cliff away from the C3-to-C4 transition. We concede this point. The Q range of 0.000 across all 21 combinations demonstrates that the cliff detection is stable under parameter perturbation, but as the referee notes, this stability is expected given that blocking delay is introduced at exactly one complexity level. We over-claimed in framing this robustness as evidence of a structural property of the authentication flow; it is more accurately described as evidence that the cliff detection is not sensitive to secondary parameters, conditional on blocking delay being present at C4. We will revise the manuscript to accurately characterize what the sensitivity analysis does and does not demonstrate, and remove language that implies the robustness analysis validates the cliff as a general structural finding rather than a consequence of the experimental design. The genuine contribution of our current experiments is the contrast between the with-blocking-delay and without-blocking-delay conditions at the same complexity levels, which shows that the same C4 configuration produces gradual degradation without blocking delay and abrupt collapse with it. However, we agree this contrast alone is insufficient to establish non-linear collapse without the continuous manipulation described in our response to the third comment.","revision_made":"yes","referee_comment":"The experimental design creates a structural near-tautology that undermines the central claim. Blocking delay (H_k > 0) is introduced exclusively at complexity level C4 (Table V). Since the Success Cliff definition (Eq. 12) identifies the complexity level with the largest consecutive drop Delta(C) and positive acceleration Gamma(C), and since blocking delay is the single largest time perturbation in the system (5–30 seconds vs. sub-second to few-second components), the cliff will always be detected at the C3-to-C4 transition by construction. The sensitivity analysis (Section VIII-E) confirms robustness across error rates and gamma thresholds, but this robustness is trivially expected: no variation in error probabilities or detection thresholds can move the cliff away from the only level where blocking delay is introduced. The Q range of 0.000 across all 21 combinations is not evidence of"},{"response":"The referee is correct that W_SMS ~ U(5,30) is a load-bearing parameter and that our paper does not adequately foreground the dependency of the cliff phenomenon on this distribution. We chose U(5,30) as a conservative representation of SMS delivery latency under normal conditions, but we lack direct empirical validation for this range in USSD deployment contexts. The referee's concern that faster real-world SMS delivery (e.g., median 3–5 seconds) could attenuate or eliminate the cliff is legitimate and cannot be dismissed. We will address this in two ways. First, we will add a sensitivity analysis that varies the W_SMS distribution parameters systematically (e.g., W_SMS ~ U(a, b) with lower bounds from 0 to 5 seconds and upper bounds from 10 to 30 seconds), which directly addresses both this comment and the third comment. This will show whether the cliff emerges at a critical delay threshold or degrades smoothly as the referee suggests. Second, we will revise the manuscript framing to explicitly state that the severity of the cliff is contingent on the SMS delay distribution and that empirical validation of this distribution is a prerequisite for applying the finding to specific deployments. We will move this caveat from the Limitations section to the abstract and introduction so it is foregrounded rather than deferred.","revision_made":"yes","referee_comment":"Section III-F, Eq. (5): The SMS OTP delivery delay model W_SMS ~ U(5,30) is the load-bearing premise for the cliff phenomenon. The entire argument hinges on this blocking delay consuming enough of the per-step time budget to trigger concurrent timeout mechanisms. However, this distribution is stated as reflecting 'normal network conditions' but is not empirically validated. If real-world SMS delivery is faster (e.g., median 3–5 seconds rather than uniform 5–30), the cliff may not materialize at the severity reported. The authors acknowledge in Section IX-D that field measurement is needed, but the paper's framing presents the cliff as a robust structural finding without adequately foregrounding this dependency."},{"response":"We agree. The current binary manipulation (blocking delay present vs. absent at C4) cannot distinguish between a true non-linear collapse at a critical delay threshold and a smooth proportional degradation. Our claim of super-additive interaction is not supported by the experimental design as described. We will add a new experiment that varies W_SMS continuously—for example, W_SMS ~ U(0, b) with b swept from 0 to 30 seconds in increments, holding complexity at C4 and network conditions fixed. This will allow us to plot session success rate as a function of blocking delay magnitude and determine whether there is a sharp transition (supporting the cliff characterization) or a smooth monotonic decline (supporting a proportional-effect interpretation). If the result is a smooth decline, we will revise the paper's central claim accordingly and reframe the contribution as identifying blocking delay as the dominant factor in session degradation rather than claiming a non-linear collapse. We will also remove or qualify the super-additive interaction claim until the continuous manipulation provides evidence for or against it.","revision_made":"yes","referee_comment":"Section VIII-D: The paper claims that failure rates under blocking delay 'exceed the sum of their independent contributions,' implying super-additive interaction. However, because blocking delay is introduced only at C4, the comparison is between a configuration without blocking delay and one with it—this is a single design point, not a continuous manipulation. To distinguish a true non-linear collapse from a proportional effect of adding delay, the authors must vary the blocking delay magnitude continuously (e.g., W_SMS ~ U(a, b) with varying a, b) and show that success rate collapses sharply at a critical delay threshold rather than degrading smoothly. The current design cannot distinguish between these alternatives."}],"tokens_in":10399,"tokens_out":2134,"duration_ms":155288,"standing_objections":["The referee's observation that the sensitivity analysis robustness is trivially expected (Comment 1) is correct and cannot be refuted. We can reframe what the analysis demonstrates, but we cannot claim it provides independent evidence for the cliff as a structural property. The revised manuscript will acknowledge this limitation directly.","The empirical validity of W_SMS ~ U(5,30) cannot be established without field measurement, which is beyond the scope of this revision. We can add sensitivity analysis on the distribution parameters and foreground the dependency, but we cannot resolve whether the cliff materializes at the severity reported under real-world SMS delivery conditions."]},"desk_editor":{"model":"glm-5.2","letter":"The paper formalizes a failure mode in USSD authentication and identifies out-of-band blocking delay as the mechanism that produces non-linear session collapse. That framing is genuinely new — the decomposition of failure into in-band complexity (gradual degradation) versus out-of-band blocking delay (cliff) is a useful analytical distinction, and the formalization in Eqs. 10–12 is clean. The simulation is methodologically careful in its parameterization: KLM-derived interaction times anchored to empirical PIN-entry data, three network regimes, three abandonment models, and sensitivity sweeps. The authors are also honest about limitations in Section IX-D, which I appreciate. Credit earned on formalization and experimental discipline. The architectural insight — that blocking delay triggers concurrent timeout mechanisms (session timeout plus user response timeout) rather than a single failure path — is the paper's strongest point and is supported by the failure taxonomy in Figures 4–5. That is a real observation about how time-constrained systems fail. Now the soft spot, and it is central. The stress-test note is correct: blocking delay (H_k > 0) appears exclusively at complexity level C4 (Table V). The cliff detection mechanism (Eq. 12) finds the largest consecutive success-rate drop with positive acceleration. Since C4 is the only configuration with a 5–30 second stochastic perturbation — dwarfing all other time components — the cliff will always be detected at C3→C4. The sensitivity analysis (Section VIII-E) varies error rates and γ thresholds, but neither of these touches the blocking delay. So the reported Q range of 0.000 across 21 combinations is not evidence of a robust threshold phenomenon; it is a direct consequence of the binary experimental design. To show a genuine cliff, the authors would need to vary blocking delay magnitude continuously and demonstrate that success rate collapses sharply at a critical delay value rather than degrading smoothly. They do not do this. The reader's concern about the U(5,30) SMS delay distribution being unvalidated is real but secondary — even with a validated distribution, the structural problem remains. The paper's contribution survives as a formalization and as a qualitative architectural argument, but the quantitative cliff-detection results are tautological as designed. This paper is for researchers and practitioners working on authentication in low-resource mobile systems. The formalization and the blocking-delay-as-mechanism argument have value independent of the flawed cliff detection. It deserves a serious referee who can push the authors to redesign the experiment with continuous blocking-delay variation, which would either confirm or refute the cliff as a genuine non-linear phenomenon. I recommend peer review.","headline":"Success Cliff framing is novel but cliff detection is structurally guaranteed by experimental design","tokens_in":11522,"tokens_out":1234,"would_cite":false,"duration_ms":90489,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"SMS OTP triggers Success Cliff in USSD sessions","keywords":[],"falsifier":"Run the same simulation with an empirically measured SMS OTP delivery distribution from a real deployment context. If the median delivery time is under ~5 seconds or the distribution is right-skewed with a thin tail rather than uniform, the blocking delay may not consume enough of the per-step budget to trigger user response timeout as a secondary failure mode, and the cliff would not materialize. The cliff's existence is structurally forced by the specific U(5,30) assumption.","tokens_in":10931,"feed_emoji":"📱","tokens_out":904,"duration_ms":185032,"temperature":0.7,"pith_summary":"This paper formally defines and demonstrates a phenomenon called the Success Cliff: a sharp, non-linear collapse in session completion rate that occurs in time-constrained USSD authentication systems when an out-of-band blocking delay—specifically SMS OTP delivery—is introduced. The authors model USSD sessions as sequences of interaction steps, each consuming time from a fixed ~120-second session budget, and run 50,000-trial simulations across four authentication complexity configurations (C1 through C4), three network latency regimes, three abandonment models, and sensitivity sweeps on error rates and detection thresholds. The central finding is structural: authentication complexity alone produces gradual, manageable degradation (2–3 percentage points across the full complexity range), and network latency alone produces linear degradation driven by a single mechanism (application timeout). But when SMS OTP blocking delay—modeled as a stochastic wait drawn from U(5,30) seconds—is introduced at the C3-to-C4 transition, session success collapses abruptly, dropping to as low as 75% under high latency, with the cliff detected invariantly across all 21 parameter combinations tested. The mechanism is architectural: the blocking delay consumes a significant portion of the per-step time budget before the user even begins responding, triggering user response timeout, session timeout, and abandonment as concurrent rather than sequential failure pathways. The cliff is therefore not a product of complexity or latency acting independently but emerges from the interaction between an uncontrollable external delivery wait and a fixed session time budget.","feed_headline":"SMS OTP triggers Success Cliff in USSD sessions","feed_subtitle":"Out-of-band blocking delay—not complexity or latency alone—collapses session success to 75%, threatening financial inclusion in mobile money","key_machinery":"The formal Success Cliff definition (Eq. 12): Q = min{C : Δ(C) ≥ δ and Γ(C) ≥ γ}, where Δ(C) is the consecutive drop in success rate between complexity levels and Γ(C) is the acceleration of that drop. The session time model (Eq. 1–2): t_k = U_k + R_k + D_k + H_k, where H_k is the blocking delay that is nonzero only for SMS OTP steps. The blocking delay model (Eq. 4–5): H_k = W_SMS if step k is an SMS OTP step, 0 otherwise, with W_SMS ~ U(5,30) seconds. The complexity metric (Eq. 6): C_j = E[U_j] / E[U_baseline], anchoring all step times to a 4-digit PIN baseline calibrated against empirical measurements from 34-participant field studies.","core_discovery":"The Success Cliff is detected exclusively at the C3-to-C4 transition—where SMS OTP blocking delay is introduced—and this detection is perfectly robust across all three abandonment models, all three error-rate regimes (baseline, low, high), and all four gamma detection thresholds (12 of 12 combinations), with a cliff-location range of 0.000. Without blocking delay, no cliff is detected under baseline conditions (0 of 12 gamma combinations, 0 of 9 error combinations except the artificially inflated high-error regime). The blocking delay creates a dual failure pathway: cumulative session timeout through total time exhaustion, and per-step user response timeout through compression of the per-30s","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["SMS OTP blocking delay causes Success Cliff in USSD","Out-of-band SMS OTP delay triggers Success Cliff in USSD","Blocking delay, not latency, creates USSD Success Cliff","SMS OTP out-of-band delay collapses USSD session success","USSD Success Cliff driven by SMS OTP blocking delay"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The SMS OTP delivery delay is modeled as uniformly distributed between 5 and 30 seconds (U(5,30)), reflecting what the authors describe as normal network conditions. This distribution is not empirically validated; the authors acknowledge field measurement is needed. The entire cliff phenomenon depends on this delay consuming enough of the per-step time budget to trigger concurrent timeout mechanisms. If real-world SMS delivery is faster—say, a median of 3–5 seconds ratherthan","fun_headline_variants_meta":{"raw":{"variants":["SMS OTP blocking delay causes Success Cliff in USSD","Out-of-band SMS OTP delay triggers Success Cliff in USSD","Blocking delay, not latency, creates USSD Success Cliff","SMS OTP out-of-band delay collapses USSD session success","USSD Success Cliff driven by SMS OTP blocking delay","Time-constrained USSD sessions collapse under SMS OTP delay"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1044,"prompt_tokens":453,"completion_tokens":591,"prompt_tokens_details":null},"tokens_in":453,"tokens_out":591,"duration_ms":20776,"temperature":1.0,"reasoning_tokens":546,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T03:38:27.756958+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Run the same simulation with an empirically measured SMS OTP delivery distribution from a real deployment context. If the median delivery time is under ~5 seconds or the distribution is right-skewed with a thin tail rather than uniform, the blocking delay may not consume enough of the per-step budget to trigger user response timeout as a secondary failure mode, and the cliff would not materialize. The cliff's existence is structurally forced by the specific U(5,30) assumption.","supporting_citations":[],"review_version":1}