{"id":"133695a3-87f5-40bd-9670-426a44a48171","arxiv_id":"2608.07356","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Residual ECS spread in CMIP6, unexplained by the Cox fluctuation statistic, is significantly aligned with a finite CO2 response coordinate, yielding a 3.00 K HadCRUT5 conditional ECS estimate.","lead":"The authors test whether equilibrium climate sensitivity information lost from a historical variability statistic reappears in the finite CO2 response. They find strong complementary residual correlation in CMIP6 models and use it to update a HadCRUT5-based ECS estimate to 3.00 K.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Observed C6=0.593 implies Q≈0.54 via the paper's Eq. (10), so ΔR is not dominated by the common hidden coordinate; the Section II D precondition for the HadCRUT5 correction is not met and Γ6 is biased.","rationale":"The reader's weakest assumption identifies the common-hidden-coordinate requirement behind the correction. My pass sharpens that concern using the paper's own Eq. (10): the observed C6=0.593 maps to Q=0.542, so the finite-response residual is not dominated by r_amp; the stated precondition for Eq. (11) fails by the paper's own quantitative criterion. This is an internal consistency check rather than an external critique. The central empirical result — that CMIP6 ΔR organizes ΔE after conditioning on Ψ — can survive this concern, so the CONDITIONAL verdict remains appropriate. I flag it as a specific reason to treat the HadCRUT5 estimate and the 'first-order correction' language with caution. No change to the reader's verdict is needed; the concern strengthens the condition but does not move the verdict.","tokens_in":7844,"tokens_out":10237,"duration_ms":97246,"concrete_test":"Simulate synthetic CMIP6-like ensembles from the two-mode model (Eq. 7-10) with parameters calibrated to reproduce C6=0.593 (hence Q=0.542) and the observed Ψ-ECS relation, then run the identical HadCRUT5 correction pipeline (Eq. 11-12) on an observation drawn from a known E⊕. If the median recovered E⊕ deviates from the true value by more than the Monte Carlo width of the reported interval, the correction is biased in the exact regime the paper reports.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II D states that 'To use this alignment as a correction, the residuals must be dominated locally by a common unresolved forced-response coordinate δη' (Eq. 5). The two-mode model makes this quantitative: Eq. (10) gives C_pred = sqrt(Q/(1+Q)), where Q=||r_amp||^2/||r_kin||^2. Substituting the headline CMIP6 value C6=0.593 yields Q=0.542. Because Q<1, the finite-time kinetic/recovery-factor component r_kin has larger variance than the response-amplitude component r_amp in ΔR. The empirical complementarity is therefore not in the regime the paper identifies as required for the correction: ΔR is majority kinetic contamination, and the ensemble-level Γ6 = Cov/Var is attenuated by the variance fraction Var(r_amp)/Var(ΔR)=Q/(1+Q)=0.351, up to the conversion scale from r_amp to ΔE. Consequently the HadCRUT5 corrected coordinate (3.00 K, 66% interval 2.17–3.83 K, Eq. 11-12) is biased toward the Ψ-only reference E0(Ψ⊕)=3.184 K. This does not overturn the qualitative complementarity claim—C6 still indicates some shared residual information—but the operational claim that the residual direction 'defines a first-order correction' depends on a dominance condition that the paper's own calibrated model says is not met.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper tests whether the weakening of the Cox et al. fluctuation-memory constraint on equilibrium climate sensitivity (ECS) in CMIP6 is compensated by information contained in a finite CO2-response coordinate. After residualizing both ECS and a 55-year ARX-estimated CO2 response against the fluctuation-memory coordinate Ψ within each ensemble, the authors find weak residual alignment in CMIP5 (C5 = 0.154, p = 0.541) and strong alignment in CMIP6 (C6 = 0.593, p = 0.00288). They interpret this contrast through a two-mode stochastic response model in which hidden response-amplitude spread and kinetic/time-scale contamination jointly control the complementarity coefficient, and they use the CMIP6 residual operator Γ6 to construct a HadCRUT5-based corrected coordinate, yielding a 66% conditional interval of 2.17–3.83 K. The paper concludes that the weakened scalar fluctuation constraint does not imply loss of ECS information, because part of that information reappears in a complementary finite-response projection.","tokens_in":8203,"tokens_out":6125,"duration_ms":57229,"significance":"If the complementarity result holds, it sharpens the interpretation of emergent constraints: a scalar variability statistic may miss susceptibility information that is visible in a finite forced-response projection. The central empirical correlation is tested against a permutation null, the analysis is transparent about data and code availability, and the paper explicitly discusses physical candidate mechanisms such as cloud feedbacks, pattern effects, and ocean heat uptake. The main caveat is that the operational correction and the HadCRUT5 interval rely on a dominance condition that, under the paper's own two-mode model, is not satisfied by the headline CMIP6 value; this makes the quantitative observational estimate less secure than the qualitative complementarity claim.","major_comments":[{"comment":"The precondition stated in Section II D for using residual alignment as a correction is that the residuals must be dominated by a common unresolved forced-response coordinate δη. The paper's own two-mode model makes this quantitative: Eq. (10) gives C_pred = sqrt(Q/(1+Q)). Inserting the headline CMIP6 value C6 = 0.593 yields Q = 0.542, so Var(r_kin) > Var(r_amp) in ΔR; the finite-time kinetic component has larger variance than the response-amplitude component. Consequently ΔR is not dominated by the common hidden coordinate, and the ensemble-level operator Γ6 = Cov/Var is attenuated by the variance fraction Var(r_amp)/Var(ΔR) = Q/(1+Q) ≈ 0.351, up to the conversion from r_amp to ΔE. The HadCRUT5 corrected coordinate in Eq. (11) is therefore biased toward the Ψ-only reference E0(Ψ⊕) = 3.184 K, and the claim that the residual direction 'defines a first-order correction' is not supported by the stated dominance condition. The complementarity result itself can stand, but the observational correction and the 2.17–3.83 K interval in Eq. (12) need either a revised estimator that does not require dominance or an explicit bias/uncertainty quantification in the Q = 0.54 regime.","section":"Section II D, Section II E, Section III B"},{"comment":"The validation of the corrected coordinate inside CMIP6 is in-sample: the same CMIP6 data are used to estimate E0, R0, and Γ6 and then to compute the correlation r = 0.675 between E_corr and ECS (Fig. 4a). This does not demonstrate predictive skill for the HadCRUT5 application, because the calibration and evaluation sets are not separated. A leave-one-model-out cross-validation or a split-ensemble check would substantially strengthen the claim that Eq. (11) is a usable first-order correction rather than a re-description of the calibration data.","section":"Section III D, Fig. 4a"},{"comment":"The agreement between C_pred and C in Fig. 3b (r = 0.955) is an internal consistency check, not an out-of-sample validation. The toy ensembles are generated from the same two-mode model that defines Q, so the high correlation verifies the algebraic relation in Eq. (10) rather than testing whether the model predicts the CMIP6 complementarity. The paper should not present this as independent mechanistic support; it should be framed as a derivation check within the model class.","section":"Section III C, Fig. 3b"}],"minor_comments":[{"comment":"The ARX model in Eq. (2a) is described only tersely; please specify the estimation procedure, the treatment of the constant c, and how T_step(55) and T_control(55) in Eq. (2b) are computed from the fitted model, so that the response feature is reproducible.","section":"Section II C, Eq. (2)"},{"comment":"The term 'conditioning' is used to mean linear residualization by OLS fits E0,g(Ψ) and R0,g(Ψ). Since 'conditional' can also imply nonparametric or Bayesian conditioning, adding an explicit sentence that the main analysis uses ordinary least squares would avoid ambiguity.","section":"Section II D"},{"comment":"The permutation null is described as two-sided, but the text reports only one p-value per ensemble and Fig. 2c shows only the upper tail. Please state explicitly whether the quoted p-values are two-sided and whether they are adjusted for the two ensembles or the four marginal correlations tested in Section III A.","section":"Section II D, Fig. 2c"},{"comment":"The analysis uses common effective radiative forcing histories for all models; this is acknowledged, but a quantitative sensitivity test (e.g., using model-specific forcing estimates where available, or omitting individual forcing components) would help assess how much of the residual alignment C6 could be an artifact of imposing a shared forcing path.","section":"Section II A"},{"comment":"The response-information ratio Q = ||r_amp||^2/||r_kin||^2 is defined only in the sentence preceding Eq. (10); adding an explicit definition of the norm and of r_amp and r_kin in the main text would make the derivation more accessible.","section":"Section II E, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of physics.ao-ph and addresses a contested question in emergent constraints. I do not see a citation-pattern or novelty problem. My recommendation of major revision is based on the internal inconsistency between the stated dominance precondition in Section II D and the value of C6 implied by Eq. (10). The complementarity finding is not invalidated, but the quantitative HadCRUT5 estimate should be reframed as conditional on a model assumption that the paper's own model suggests is not satisfied, or the authors should provide a bias-corrected estimator. The paper's transparency about code and data is a clear strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central empirical result here is real: after conditioning on Ψ, CMIP6 ECS residuals and finite CO2-response residuals align at C6=0.593 (p=0.00288 by permutation), while CMIP5 does not show this. That is a genuinely useful diagnostic, and the two-mode stochastic model gives a plausible mechanism for why a scalar fluctuation constraint could weaken while response information reappears in a finite-response coordinate. The paper is well-written, honest about the precondition in Section II D, and ships code, which is more than many submissions in this area.\n\nThe soft spot is the operational correction. The stress-test math checks out. The paper's own Eq. (10) maps C6=0.593 to Q=0.54, meaning the finite-response residual is majority kinetic/time-scale contamination, not the common hidden forced-response coordinate δη. Section II D states that using the alignment as a correction requires the residuals to be dominated by δη. The paper's own calibrated toy model says they are not. So Γ6 is attenuated, and the HadCRUT5 corrected coordinate (3.00 K, 66% interval 2.17–3.83 K) is pulled toward the Ψ-only reference. The paper does not acknowledge this tension. This does not overturn the qualitative complementarity claim, but it does mean the abstract's 'first-order correction' language is overstated.\n\nOther issues are minor by comparison: 18 and 23 models are small, the ARX response coordinate relies on common forcing histories that may not match each model's actual forcing, and the HadCRUT5 correction is calibrated in-sample. These are limitations, not fatal flaws.\n\nWho should read this? Anyone working on emergent constraints, ECS estimation, or fluctuation–response theory. The diagnostic deserves a serious referee. A good referee should ask for an explicit decomposition of ΔR into amplitude and kinetic parts, or a separate out-of-sample test of the correction, before the HadCRUT5 number is taken seriously. I'd engage with the paper, cite the diagnostic, and not quote the corrected ECS estimate.","headline":"The residual-complementarity diagnostic is new and the CMIP6 result is credible, but the paper's own model implies the HadCRUT5 correction is biased; worth a serious referee, not worth quoting the corrected ECS.","tokens_in":8746,"tokens_out":3557,"would_cite":true,"duration_ms":29865,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims the CMIP6 weakening of the fluctuation constraint on climate sensitivity is a relocation, not a loss, of information: the ECS residual aligns with a finite CO2-response residual ($C_6 = 0.593$), and the correction applied…","keywords":["equilibrium climate sensitivity","emergent constraints","fluctuation-memory coordinate","finite CO2 response","residual complementarity","linear response theory","CMIP6","HadCRUT5"],"falsifier":"Run the identical residual pipeline on an independent ensemble with known sensitivities, for instance single-forcing historical simulations or the next generation of CMIP historical runs, and check whether the finite-response residual aligns with the ECS residual after conditioning on $\\Psi$. The mechanism predicts $C_g$ should again be strongly positive wherever hidden response-amplitude spread is poorly projected onto $\\Psi$; observing $C_g$ near zero there would falsify the claim that finite-response complementarity recovers the lost sensitivity information.","tokens_in":7605,"feed_emoji":"🌡️","tokens_out":13516,"duration_ms":105812,"temperature":0.7,"pith_summary":"The paper asks whether climate-sensitivity information that a well-known variability-based constraint fails to capture in modern models reappears in another observable. The authors construct a finite-time CO2 response coordinate $R$ from historical temperature and common forcing histories, without using any model's sensitivity label, and test whether the ECS left unexplained by the fluctuation-memory coordinate $\\Psi$ is organized by the part of $R$ that $\\Psi$ also leaves unexplained. In CMIP5 the residuals are nearly uncoupled ($C_5 = 0.154$, $p = 0.541$), but in CMIP6 they align strongly ($C_6 = 0.593$, $p = 0.00288$), meaning the information missed by $\\Psi$ in CMIP6 is visible in a finite response. A two-mode stochastic response model shows that hidden forced-response amplitude spread poorly projected onto $\\Psi$ produces exactly this alignment, while unrelated kinetic or timescale variation suppresses it. Applying the CMIP6 residual direction to the observed HadCRUT5 coordinates yields a conditional ECS estimate centered at 3.00 K, so the much-discussed weakening of the constraint need not mean the susceptibility information is gone.","feed_headline":"Lost climate-sensitivity signal returns through CO2 response","feed_subtitle":"After the CMIP6 constraint weakens, residual alignment 0.593 recovers HadCRUT5 ECS near 3.00 K.","key_machinery":"The machinery is a residual-complementarity test built from two projected coordinates. $\\Psi$ is the fluctuation-memory coordinate, the window-averaged ratio $\\sigma_w/(-\\log \\alpha_{1,w})$ of detrended standard deviation to one-year lag memory in rolling 55-year windows, following the original emergent-constraint construction. $R_{\\mathrm{CO2}}$ is a finite CO2 response coordinate: a CO2-doubling-equivalent 55-year step response of a scalar ARX/state-space model fitted to each model's annual temperature with common radiative forcing histories, estimated without ECS labels. After conditioning both ECS and $R$ on $\\Psi$ by linear regression, the central objects are the residual alignment $C_g = \\mathrm{corr}_g(\\Delta E, \\Delta R)$ and the residual operator $\\Gamma_g = \\mathrm{Cov}_g(\\Delta E, \\Delta R)/\\mathrm{Var}_g(\\Delta R)$, which converts the alignment into a first-order correction. A two-mode stochastic response model (an observed temperature coordinate coupled to a hidden slow state under common forcing) supplies the mechanism: the hidden susceptibility amplitude $H$ has cross-model spread partly orthogonal to $\\Psi$, generating complementary residuals, and the response-information ratio $Q = \\|\\mathbf{r}_{\\mathrm{amp}}\\|^2/\\|\\mathbf{r}_{\\mathrm{kin}}\\|^2$ predicts $C$ via $C_{\\mathrm{pred}} = \\sqrt{Q/(1+Q)}$.","core_discovery":"The paper's central claim is a statement about complementarity under projection: a scalar fluctuation statistic cannot encode the whole forced response, but the part of the equilibrium climate sensitivity it misses can reappear in an observable finite-time CO2 response. After regressing ECS and $R_{\\mathrm{CO2}}$ on $\\Psi$ separately within each ensemble, the residual complementarity coefficient $C_g = \\mathrm{corr}_g(\\Delta E, \\Delta R)$ is essentially zero in CMIP5 ($C_5 = 0.154$) and strongly positive in CMIP6 ($C_6 = 0.593$, $p = 0.00288$), with residual operator $\\Gamma_6 = 0.899$ converting the alignment into a correction. The contrast is reproduced by a minimal two-mode stochastic model in which a hidden slow response pathway with cross-model amplitude spread partly orthogonal to $\\Psi$ generates the alignment, and the ratio $Q$ of response-amplitude to kinetic contamination in $\\Delta R$ predicts $C$ through $C_{\\mathrm{pred}} = \\sqrt{Q/(1+Q)}$. At the observation-facing level, the corrected coordinate $\\hat{E}_{\\mathrm{corr}} = E_0^{(6)}(\\Psi) + \\Gamma_6 [R_\\oplus - R_0^{(6)}(\\Psi)]$ for HadCRUT5 is 3.004 K, with a 66% conditional diagnostic interval of 2.17–3.83 K. The paper concludes that the weakened CMIP6 fluctuation constraint is not evidence that susceptibility information is lost; it is evidence that the information is carried by a complementary finite-response projection.","pith_inferences":["If the complementarity mechanism is general, the same residual diagnostic should work prospectively: before an ensemble's sensitivity labels are used, one could predict that the finite-response residual organizes the ECS residual, and a pre-registered test on a future CMIP generation would separate the mechanism from a CMIP6-specific coincidence.","The two-mode model implies that the controlling quantity $Q$ can be estimated directly from real model behaviour (for instance from the ratio of each model's 55-year response to its converged response), which would let $C$ be predicted without computing any residual statistics.","The corrected 3.00 K value inherits the CMIP6 ensemble mean through $E_0(\\Psi)$; an out-of-sample split, estimating $\\Gamma$ on one half of CMIP6 and applying it to the other half, would reveal how much of the shift is a genuine correction rather than a rescaling of the ensemble.","Extending the test from global-mean temperature to spatial patterns or ocean-heat-content observables would identify what the hidden coordinate $\\delta\\eta$ physically is — something the paper proposes but does not demonstrate."],"forward_implications":["The documented weakening of the $\\Psi$-based ECS constraint from CMIP5 to CMIP6 does not close the possibility of constraining sensitivity from historical variability: in CMIP6 the residual ECS signal is carried by the finite CO2 response ($C_6 = 0.593$).","A first-order correction along the CMIP6 residual direction places HadCRUT5's conditional ECS at about 3.00 K, with the 66% diagnostic interval 2.17–3.83 K overlapping the commonly quoted assessment range.","Emergent constraints should be evaluated as paired projections: a scalar coordinate is informative only if the residual it leaves in ECS is organized by an independent observable's residual, making $C_g = \\mathrm{corr}(\\Delta E, \\Delta R)$ a general diagnostic.","Model-ensemble design should aim spread at physically interpretable response directions — cloud feedbacks, pattern effects, ocean heat uptake — because the value of an ensemble is whether it spans response directions that could contain the real system, not just how large it is.","The finite CO2 response coordinate does not replace $\\Psi$; it measures the forced-response part that the scalar fluctuation-memory projection fails to resolve."],"supporting_citations":[{"why":"Supplies the fluctuation-memory statistic $\\Psi$, the rolling-window variability-to-memory ratio that the whole analysis conditions on.","marker":"[5]"},{"why":"The direct CMIP6 test documenting the weakening of the $\\Psi$-based constraint; provides the empirical contrast the paper explains.","marker":"[19]"},{"why":"Traces the CMIP5/CMIP6 difference to a scalar assumption that failed; the assumption the complementarity framework replaces.","marker":"[20]"},{"why":"Projection/memory formalism showing how hidden coordinates survive when degrees of freedom are projected out; motivates the hidden-response coordinate.","marker":"[16]"},{"why":"Sets the requirement that emergent constraints rest on testable assumptions, which is what the residual-complementarity test supplies.","marker":"[4]"},{"why":"Attributes CMIP6 sensitivity spread to cloud-feedback differences; the physical mechanism class invoked for the hidden response direction.","marker":"[21]"},{"why":"The HadCRUT5 temperature dataset used for the observation-facing corrected ECS estimate.","marker":"[25]"}],"fun_headline_variants":["CMIP6 weak constraint? CO2 response pulls back ECS signal","Residual alignment 0.593 recovers climate sensitivity info","Complementary finite CO2 response restores missing ECS","Hidden response spread shows up in CO2 response, not lost","Weakened fluctuation constraint? ECS reappears via CO2 response"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole correction rests on the assumption that the leftover spread in ECS and the leftover spread in the finite CO2 response are dominated by one common hidden forced-response direction; if the finite-response residual is instead largely unrelated kinetic or timescale variation, the operator $\\Gamma_6$ and the 3.00 K estimate are biased.","fun_headline_variants_meta":{"raw":{"variants":["CMIP6 weak constraint? CO2 response pulls back ECS signal","Residual alignment 0.593 recovers climate sensitivity info","Complementary finite CO2 response restores missing ECS","Hidden response spread shows up in CO2 response, not lost","Weakened fluctuation constraint? ECS reappears via CO2 response"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2202,"prompt_tokens":1056,"completion_tokens":1146,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":1057}},"tokens_in":672,"tokens_out":1146,"duration_ms":8723,"temperature":1.0,"reasoning_tokens":1057,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:37:22.054823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical residual pipeline on an independent ensemble with known sensitivities, for instance single-forcing historical simulations or the next generation of CMIP historical runs, and check whether the finite-response residual aligns with the ECS residual after conditioning on $\\Psi$. The mechanism predicts $C_g$ should again be strongly positive wherever hidden response-amplitude spread is poorly projected onto $\\Psi$; observing $C_g$ near zero there would falsify the claim that finite-response complementarity recovers the lost sensitivity information.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fluctuation-memory statistic $\\Psi$, the rolling-window variability-to-memory ratio that the whole analysis conditions on."},{"cited_title":"Schlund, A","cited_arxiv_id":null,"evidence_quote":"The direct CMIP6 test documenting the weakening of the $\\Psi$-based constraint; provides the empirical contrast the paper explains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Traces the CMIP5/CMIP6 difference to a scalar assumption that failed; the assumption the complementarity framework replaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sets the requirement that emergent constraints rest on testable assumptions, which is what the residual-complementarity test supplies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Attributes CMIP6 sensitivity spread to cloud-feedback differences; the physical mechanism class invoked for the hidden response direction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The HadCRUT5 temperature dataset used for the observation-facing corrected ECS estimate."}],"review_version":1}