{"id":"f063bedc-438d-44ff-8c6c-c3bdaaf800ba","arxiv_id":"2607.04253","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Cox–Ingersoll–Ross diffusion bridge with water-temperature-driven random start/end times is shown to be well-posed and applied to Ayu migration data.","lead":"The paper builds a stochastic model of fish migration counts where the start and end of migration are set by water temperature, not fixed dates. It proves the model has unique nonnegative paths and applies it to 20+ years of Ayu migration counts and to eDNA samples.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mathematical well-posedness is not the fatal issue; the load-bearing concern is that the WT thresholds in §3.2.2 are calibrated from the very start/end dates they are then used to explain, and the nominal stochastic model does not reproduce those calibration targets.","rationale":"The reader's verdict is CONDITIONAL, and the load-bearing concern is the same one the reader identified: the empirical WT thresholds are calibrated in-sample to the migration dates they are then used to explain. I do not see a fatal flaw in Proposition 1: the terminal-time argument can be made rigorous because the martingale term is L^2-bounded with vanishing variance, so it converges almost surely to 0, and the deterministic term also vanishes; this is a standard martingale-convergence argument rather than a real gap. The application, however, is vulnerable. The thresholds in §3.2.2 are averages of the observed start/end WTs, so in the deterministic WT case the model matches the observed mean dates by construction. In the stochastic nominal case, Table B2 shows that the model's mean start and end dates (13.5 and 138.4 days) differ noticeably from the observed values (about 20.7 and 148.5 days), so the calibration is not even internally consistent under the model's own stochastic dynamics. Moreover, the paper cites [16] as suggesting calendar date dominates over local environmental factors, but does not provide an independent test that would distinguish the WT-clock mechanism from a simple calendar-date model. This does not undermine the mathematical contribution, but it does mean the applied claim should remain conditional on out-of-sample validation. Thus the reader's CONDITIONAL verdict is appropriate and no change is needed.","tokens_in":30910,"tokens_out":20936,"duration_ms":223370,"concrete_test":"Use a temporal holdout: estimate the thresholds (and OU parameters) using only 2003–2020 data from Table 2 and Table B1, then simulate the model to predict the start and end dates for 2021–2025. Compare the predictive RMSE against a null calendar-date model that simply predicts the historical average start/end dates (20.7 and 148.5 days). If the WT-threshold model does not achieve lower RMSE, the randomized-time mechanism is not supported by the Nagara River data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main mathematical claim (Proposition 1) appears essentially sound: the terminal-pinning step is not a genuine gap, since the martingale part N^2_s is L^2-bounded with vanishing variance and hence converges a.s. to 0, while the deterministic part N^1_s→0 by the Assumption 2 exponents. The load-bearing weakness is in the application. In §3.2.2 the thresholds are set to the observed average water temperatures on migration start/end dates (w=9.07 °C, w̄=23.23 °C, Table 2) and then used in model (8) to define θ1 and θ2. Because daily WT is nearly linear (Table B1), this determines the deterministic-model start/end dates essentially by construction. Under the stochastic WT model actually used as nominal (Eq. 22, ω=2), Table B2 gives an average start date of 13.5 days and end date of 138.4 days, versus the observed averages of about 20.7 and 148.5 days. The nominal model does not reproduce the very statistics to which the thresholds were calibrated, and no out-of-sample or independent test is provided. The cited review [16] identifies calendar date as the dominant driver, and the paper itself notes this tension but does not resolve it. Thus the application does not independently validate the environmental-clock mechanism; it demonstrates calibration only.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Cox–Ingersoll–Ross (CIR) bridge whose initial and terminal times are random, generated by hitting times of the integral of a nonnegative environmental process M. The central theoretical result (Proposition 1, §2.2.2) asserts pathwise existence, uniqueness, nonnegativity, and pinning X_{θ1}=X_{θ2}=0 under Assumptions 1–2, via time-change to a standard CIR bridge with singular coefficients. The authors then apply the model to juvenile Ayu migration in the Nagara River, using water temperature (WT) thresholds to define the migration window and Ornstein–Uhlenbeck WT dynamics, and to eDNA time series in the Hii River through a linear/nonlinear ODE coupled to the bridge.","tokens_in":31335,"tokens_out":8985,"duration_ms":106563,"significance":"The mathematical construction is a genuine contribution: randomizing the bridge endpoints through a time-change preserves the affine structure of the CIR bridge while allowing environmental information to enter, and the proof gives an explicit construction of the time-changed Brownian motion. The use of closed-form moment formulas from the underlying CIR bridge is a practical strength, and the paper is transparent about several limitations of the empirical analysis. However, the application section does not currently validate the environmental-clock mechanism: the WT thresholds are calibrated from the very start/end dates the model is said to explain, and the nominal stochastic model does not reproduce those calibration targets. The eDNA case study is even more clearly exploratory. If the empirical claims are reframed or independently tested, the manuscript could be publishable; as it stands, the application is a calibrated illustration rather than a demonstration that WT is the causal clock.","major_comments":[{"comment":"The thresholds w=9.07 °C and w̄=23.23 °C are the observed average WTs on the migration start/end dates in Table 2, and the deterministic WT (21) is anchored at the observed mean start date t_start=20.7 d. Thus the deterministic model reproduces the observed mean start/end dates by construction, not by independent mechanism. Under the stochastic WT model used as nominal (Eq. (22), ω=2), Table B2 gives mean start 13.46 d and mean end 138.4 d, against observed means ≈20.7 d and ≈148.5 d. So the nominal model does not even match the calibration targets, and no out-of-sample or cross-validation is provided. Please either identify thresholds independently, fit the full start/end distribution with uncertainty, or explicitly reframe this section as an illustration rather than validation.","section":"§3.2.2, Table 2, Eqs. (8) and (21), Table B2"},{"comment":"The duration and end-date statistics appear to be mixtures with a large degenerate component: the Fig. 4 caption reports a 47% point mass at T_emp=127.8 d, and Table B3 shows the duration standard deviation is 2×10^{-9} d for ω=1. Consequently, the reported means and standard deviations in Tables B2–B4 combine a continuous distribution with a point mass placed exactly at the calibration value. This must be stated explicitly, and its biological interpretation discussed; as presented, the histograms, particularly Fig. 4(c), obscure the structure of the model output.","section":"§3.2.3, Fig. 4, Tables B2–B4"},{"comment":"The eDNA analysis inherits the threshold circularity: the 2025 start (Mar 10) and end (Jul 14) dates are inferred from the same w̄=23 °C threshold, giving T_emp=127 d, and G,H,R are calibrated by least squares to the same weekly eDNA series. The allometric exponent H≈0.75 is reported without confidence bounds and rests on roughly 15 weekly samples; moreover, the paper itself notes that a transferred model produces R≈100, which is unphysical. I recommend presenting this as a proof-of-concept and adding uncertainty quantification, rather than as an estimated allometric relationship.","section":"§3.3.3, Table 3, Eq. (23)"},{"comment":"The exponent bookkeeping in the proof of the two limits in (29) is not transparent and appears garbled. In Eq. (30) the integrand is said to be O((1-u)^{α-1}), but the preceding estimate contains the factor (1-u)^{-r} from the exponential weight; the printed condition “if 1? ... i.e., α>-1” seems to drop the dependence on r. With the application value r=61.9 (Table 1), a bounded â (α=0) would not make ∫_0^1 (1-u)^{-r} a_u du finite, so a more careful argument or a corrected assumption is needed. Please rewrite this part so that the sufficiency of Assumption 2 can be checked.","section":"Appendix A, proof of Proposition 1, Eqs. (29)–(33)"}],"minor_comments":[{"comment":"The inverse gamma density in (15) appears misprinted: the normalization, the exponent of z, and the argument of the exponential need to be checked against the stated parameters (μ, λ).","section":"Eq. (14)–(15)"},{"comment":"Typos and wording: “This modal” should be “This model”; “CIR brides” should be “CIR bridges”; the Declaration says “The author used” while the paper has two authors; and in §3.2.3 “approximately 7 (day) and 10 (day)” should specify that these are standard deviations of the start/end dates, not mean shifts.","section":"Throughout"},{"comment":"The value r=6.190E+01 is very large relative to the usual CIR reversion scale; please clarify whether this is the same r as in Proposition 1 and why the moment formulas (37)–(38) remain valid for this value.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core is likely salvageable and could be a reasonable contribution to applied probability / SDE modeling. My main reservation is the empirical section, which currently claims more than it demonstrates. The authors should either provide an independent calibration of the WT thresholds or substantially weaken the language about WT being the driver of migration timing. The heavy reliance on the authors’ own prior papers for the base model and parameters is acceptable but worth monitoring in review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution here is the randomized initial/terminal time CIR bridge and the well-posedness result. The time-change construction is standard but carefully applied, and Proposition 1's proof checks out in substance. The reader's worry about an L2-vs-pathwise gap at the terminal point is overcautious: the variation-of-constants decomposition has a deterministic term tending to 0 and a martingale term whose variance vanishes, so L2-bounded martingale convergence gives the a.s. limit 0. Pathwise continuity at the endpoint follows by extension. The nonnegativity and the careful treatment of the singular coefficients under Assumption 2 are also solid. That part deserves serious attention.\n\nThe soft spot is the application, and it is genuine. In §3.2.2 the thresholds w and w̄ are set to the observed average water temperatures on the migration start and end dates (9.07 and 23.23 °C). With the deterministic linear WT model (21), this reproduces the observed average start/end dates by construction. But the nominal stochastic model (22) gives average start and end dates of 13.5 and 138.4 days, versus the observed 20.7 and 148.5. So the model does not reproduce the very statistics to which the thresholds were calibrated, and no out-of-sample test is offered. The paper honestly cites the literature suggesting calendar date is a dominant driver, but it does not resolve the tension. The eDNA example uses the same thresholds and a chosen endpoint of July 14, which is again circular in spirit.\n\nThat said, the sensitivity analysis is competently done and the parameter ranges are clearly reported. The mathematical contribution stands independently of the data application, and the authors do not hide the limitations. This is a paper with a solid theoretical core and a thin, partially circular empirical section.\n\nI would send it to peer review: the stochastic bridge result is new and should be accessible to the community. But the referee should push for a proper calibration or reframing of the applied work as an illustrative case study, not validation. The paper is worth engaging with, and a serious referee can help separate the durable math from the weak application.","headline":"The randomized-time CIR bridge is a genuinely new and essentially sound piece of stochastic modeling; the real weakness is the empirical calibration, which fits thresholds to the very data the model then fails to reproduce under stochastic WT.","tokens_in":31769,"tokens_out":2766,"would_cite":true,"duration_ms":35342,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","60G40","60J60","92D25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A temperature-driven biological clock randomizes the endpoints of a fish-migration bridge, and the resulting model is provably well-posed.","keywords":["CIR bridge","random initial and terminal times","time change","fish migration","Plecoglossus altivelis altivelis (Ayu)","water temperature","environmental DNA","stochastic differential equation"],"falsifier":"Hold out one or more years of Nagara River data; use the fitted temperature thresholds (9.07 °C, 23.23 °C) to predict the start and end dates in the held-out years. If the prediction error is no smaller than using the fixed average calendar dates (start Feb 20, end Jun 30), then the temperature-driven clock adds no predictive power. For the mathematical claim, a counterexample would be a nonnegative càdlàg M that is positive a.e. but for which the solution of (4) fails uniqueness or nonnegativity; checking edge cases where M vanishes on a set of positive measure would test whether the strict-p","tokens_in":30837,"feed_emoji":"🐟","tokens_out":8051,"duration_ms":78284,"temperature":0.7,"pith_summary":"This paper tries to establish that the Cox-Ingersoll-Ross (CIR) bridge, a diffusion bridge used to model the number of fish migrating past a fixed point, remains well-posed when its initial and terminal times are made random through a time-change driven by an environmental process. The authors define a 'biological clock' as the time integral of a nonnegative factor such as water temperature, and let migration start when that clock crosses a lower threshold and end when it crosses an upper threshold. The central result is a proof that the resulting stochastic differential equation has a pathwise unique, nonnegative solution that vanishes at both random endpoints. The empirical application calibrates the thresholds to observed water temperatures on Ayu migration start and end dates in the Nagara River, and demonstrates that total fish counts are largely robust to temperature noise. If the mathematical claim holds, this offers a tractable way to let environmental variation set migration timing without abandoning the closed-form statistical structure of the bridge.","feed_headline":"A temperature clock randomizes the start and end of a migration bridge","feed_subtitle":"Water temperature sets the migration window; the underlying equations are provably well-posed.","key_machinery":"The key machinery is the time-change (5): the biological clock τ_t = ∫0^t M_u du, where M is a nonnegative, càdlàg environmental process. The initial and terminal times θ_i are the first times the clock crosses thresholds T_i. Under biological time s = τ_t, the state Y_s = X_t satisfies the original fixed-time CIR bridge (6), so the well-posedness of the original bridge is transferred to the randomized-time model. The load-bearing identity is that the transformed noise is a standard Brownian motion, which requires strict positivity of M almost everywhere; this is what turns the formal time-change into a rigorous equivalence.","core_discovery":"The central claim, Proposition 1, is that under Assumptions 1–2 the randomized-time CIR bridge has a well-posed solution: the stopping times θ1 < θ2 are almost surely finite and strictly ordered, and the SDE (4) admits a pathwise unique, almost surely nonnegative solution on (θ1, θ2) with X_{θ1} = X_{θ2} = 0. The argument turns on the time-change (5): in biological time s = ∫0^t M_u du, the randomized-time bridge becomes the original fixed-time CIR bridge, provided the environmental process is almost everywhere strictly positive so that the time change is invertible. The proof constructs the driving Brownian motion explicitly and verifies it is a genuine Brownian motion via Lévy's characteri","pith_inferences":["If water temperature is only a correlate of calendar date, then the fitted thresholds might simply be an elaborate clock that re-labels the season: a decisive test is to compare start/end predictions in an anomalous-temperature year (e.g., a warm early March) against the fixed-date baseline – if the temperature clock predicts no better, the random-time machinery is not adding causal content.","The time-change construction is general: any diffusion bridge with an affine drift can be randomized this way by choosing M as a function of a covariate (e.g., discharge, salinity, day length), so the paper's Proposition 1 sets up a template for other environmentally timed animal movements.","The buffer-zone effect suggests a monitoring implication: fish-count sampling in the central migration window matters most for estimating total abundance, since the endpoints are insensitive regions – a testable prediction for survey design.","The eDNA attenuation constant R may serve as a river-specific fingerprint of mixing and degradation; the failure of parameter transfer is itself informative and points to collecting local eDNA decay experiments before applying the model to a new site."],"forward_implications":["Well-posedness (Proposition 1): the randomized-time CIR bridge has a unique, continuous, nonnegative solution on (θ1, θ2) that vanishes at both random endpoints, justifying simulation and estimation.","A 10% change in the migration-duration parameter T_emp changes the expected total fish count by about 10%; this parameter, not the temperature noise or the clock acceleration ω, is the dominant sensitivity in the Ayu application.","Total fish-count statistics are nearly insensitive to doubling or halving the water-temperature noise (relative differences around 0.5%), because the identified shape parameters m,n,p,q are large and create low-count buffer zones near the endpoints.","The eDNA model obtained by coupling the bridge to a degradation-accumulation ODE is well-posed; a sublinear allometric exponent H≈0.75 fits the Hii River eDNA data about 5% better than H=1 (RMSE), and the attenuation constant ≈0.4/day suggests the eDNA signal retains a memory of the fish count for a few days.","Direct transfer of parameters from the Nagara to the Hii River gives an unphysical attenuation constant (R≈100), so site-specific estimation is required."],"fun_headline_variants":["Temperature randomizes migration start and end: a well-posed CIR bridge","Provably well-posed SDE for migration times driven by temperature","CIR bridge with random endpoints models fish migration under temperature","Temperature-driven migration: a provably well-posed bridge model"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The empirical application assumes water temperature is the dominant driver of Ayu migration start and end; if calendar date or some other factor actually controls timing, then the temperature-threshold clock is fitted to the data it claims to explain, and the random-time mechanism has no causal content.","fun_headline_variants_meta":{"raw":{"variants":["Temperature randomizes migration start and end: a well-posed CIR bridge","Provably well-posed SDE for migration times driven by temperature","CIR bridge with random endpoints models fish migration under temperature","Temperature-driven migration: a provably well-posed bridge model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3474,"prompt_tokens":753,"completion_tokens":2721,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":2647}},"tokens_in":497,"tokens_out":2721,"duration_ms":17212,"temperature":1.0,"reasoning_tokens":2647,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:38:35.670604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out one or more years of Nagara River data; use the fitted temperature thresholds (9.07 °C, 23.23 °C) to predict the start and end dates in the held-out years. If the prediction error is no smaller than using the fixed average calendar dates (start Feb 20, end Jun 30), then the temperature-driven clock adds no predictive power. For the mathematical claim, a counterexample would be a nonnegative càdlàg M that is positive a.e. but for which the solution of (4) fails uniqueness or nonnegativity; checking edge cases where M vanishes on a set of positive measure would test whether the strict-p","supporting_citations":[],"review_version":2}