{"id":"2767e7bf-02fd-4b21-8046-49f3e06aec8d","arxiv_id":"2607.13074","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For fixed crack geometry, combined tension–bending–bearing load is identifiable from the SIF profile iff the three elementary load profiles are affinely independent; recovery error is controlled by the minimum singular value of the profile matrix, not the condition number.","lead":"This paper proves that, for a fixed crack geometry, the mix of tension, bending and bearing loads is recoverable from a stress-intensity-factor profile exactly when the three single-load profiles are not collinear, and that the instability of recovery is governed by the smallest singular value of the profile matrix rather than by its condition number. It applies this to the public SIFBench finite-element dataset, where a learned surrogate and a simplex-constrained estimator s","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical identifiability split rests on learned surrogate M̂ without validation against true FEM κ_s; Table 2 shows 30% vs 17.5% ill-posed fractions, suggesting possible bias.","rationale":"The reader's weakest assumption correctly identifies the learned-surrogate sensitivity as the primary soft spot. The theoretical result (Theorem 1) is self-contained linear algebra and appears correct; the paper is also honest about missing real data and the Newman–Raju failure. However, the empirical identifiability split—the paper's most visible quantitative contribution—is computed exclusively from M̂, and the paper itself provides no comparison to the true FEM-based κ_s. My own reading of Table 2 strengthens this concern: the ill-posed fraction drops from 30% (Test A) to 17.5% (Test B) when the classifier switches from true to learned M, which is exactly the kind of surrogate-induced bias that would invalidate the headline numbers. The σ_min correlations, while plausible, inherit the same risk. This does not change the verdict: CONDITIONAL remains appropriate, because the concern is addressable by a straightforward sensitivity analysis on existing data and does not affect the theorem. I therefore mark the verdict as unchanged, and I agree with the reader that this is the weak assumption to test.","tokens_in":17009,"tokens_out":6814,"duration_ms":64435,"concrete_test":"On the 1000 held-out geometries, compute κ_s and σ_min from both the learned M̂ and the true FEM M (the latter is already used for Test A). Plot κ_s(M̂) against κ_s(M_true), and report: (i) the median bias in κ_s; (ii) the fraction of geometries whose ill-posed classification (κ_s>50) changes between true and learned; (iii) the recomputed median κ_s and the 30% ill-posed fraction using the true M. If the ill-posed fraction shifts by more than a few percentage points, the §7.3 empirical split must be treated as surrogate-dependent until the same analysis is run on the exact profiles.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The theoretical core (Theorem 1 and Propositions 2–3) is sound and model-free, but the empirical claims that motivate the paper—median κ_s≈20 and ≈30% ill-posed geometries—are computed from the learned surrogate M̂(g), not from the actual FEM profiles. §7.3 states: \"The quantity κ_s(M) is computed from the learned M(g) over the test geometries.\" No comparison is reported between κ_s(M̂) and κ_s(M_true) on the same geometries, even though the true FEM profiles are available (they are used in Test A of §7.4). This is a serious gap because the 70/30 well/ill-posed split is a headline quantitative outcome, and it is precisely what surrogate bias would corrupt.\n\nEvidence that the bias is real: in Table 2, the ill-posed stratum (κ_s>50) contains 6000/20000 draws (30%) for Test A using the true map, but only 3500/20000 (17.5%) for Test B using the learned M̂ on the same 1000 test geometries classified by their respective κ_s values. The paper does not discuss this discrepancy. A regression-to-the-mean effect in the surrogate would shrink the differences among K_T, K_B, K_P, reducing the functional area A(g) and inflating κ_s, or, conversely, smoothing could mask true ill-posedness; either way the reported split is not established.\n\nSimilarly, the σ_min-based stability correlations (corr(log σ_min, L1) ≈ −0.47 for Test A′ and −0.44 for Test B) are presented as confirming Proposition 3 on the learned operator. If σ_min(M̂) is a biased estimate of the true margin, the agreement could be an artifact rather than a property of the identifiability geometry. The paper provides no sensitivity analysis of κ_s or σ_min to surrogate error, and the released pipeline is promised, not yet available. This concern is empirical, not a flaw in the theorem, but it undermines the paper's central applied claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the inverse problem of recovering the tension/bending/bearing load mixture from a stress-intensity-factor profile along a crack front, using SIFBench finite-element data. For fixed geometry the forward map is exactly linear, and the paper's main theoretical contribution (Theorem 1) characterizes identifiability by the functional area A(g) of the triangle spanned by the three elementary load profiles in L^2: Φ is injective on the affine hull of the load simplex iff A(g)>0, equivalently σ_min(MB)>0, with quantitative identities σ_min(MB)σ_max(MB)=A(g)/√3 and κ_s=√3σ_max(MB)^2/A(g). Propositions 2 and 3 identify the degenerate direction and show that recovery error under noise is controlled by σ_min(MB), not by κ_s alone. A transformer-based crack-front operator is trained to supply M(g), and an inverse estimator with Dirichlet credible regions is described. The empirical sections report that on SIFBench corner-crack geometries the typical geometry is well-posed (median κ_s≈20) while ≈30% are ill-posed, that the noise behaviour follows the σ_min law, and that Test A (true forward map) recovers loads exactly at zero noise. The paper explicitly disclaims forensic use and states that credible-region coverage on the learned map is a falsifiable prediction rather than a reported result.","tokens_in":17384,"tokens_out":5566,"duration_ms":51982,"significance":"The theoretical core is sound and genuinely useful. Theorem 1 and Propositions 2–3 are derived cleanly from the linearity of the forward map; they give a closed-form, model-free criterion for identifiability and correctly identify σ_min(MB) as the intrinsic stability margin, explaining why κ_s alone can mislead when σ_max varies. The paper is also commendably honest: it reports that Corollary 4 (the Newman–Raju shallow-crack collapse) does not transfer to the FEM data, that the forward operator is behind RFR on the bending channel, and that no forensic validation is claimed. If the empirical claims about the SIFBench geometries — the 70/30 well/ill-posed split and the σ_min-controlled stability law — can be supported against the true FEM profiles rather than only the learned surrogate, and if the calibration claim is either verified or explicitly demoted, this would be a solid and citable contribution to inverse problems in fracture mechanics.","major_comments":[{"comment":"The headline empirical identifiability statistics (median κ_s≈20.3, ≈30% ill-posed, Fig. 2) are computed from the learned surrogate M̂(g), not from the FEM profiles, as stated in §7.3. No comparison is reported between κ_s(M̂) and κ_s(M_FEM) on the same geometries, even though the true profiles are available (and are used in Test A). The discrepancy is visible in Table 2 / §7.4: Test A (true map) has 6000/20000 = 30% ill-posed draws, while Test B (learned M̂) has 3500/20000 = 17.5%; the text calls these 'matching' without discussing the factor-of-two difference. Because surrogate bias can systematically shift the κ_s distribution, the paper's central empirical claim that 'the typical geometry is well posed while a sizable minority is genuinely ill-posed' is not established for the actual SIFBench geometries. Please report κ_s(M_FEM) on the same test set and provide a sensitivity analysis","section":"§7.3 and Table 2"},{"comment":"The central calibration claim is promised but not delivered. §6 states: 'the claim, verified below, is that the dispersion of the predicted posterior tracks the true non-identifiability measured by κ_s ... so that credible regions retain nominal coverage uniformly over the identifiability range.' Yet §7.4 says this claim 'is stated as a falsifiable prediction of the construction rather than reported here,' and §9 repeats that 'full credible-region coverage on the learned map is stated as a falsifiable diagnostic rather than reported.' This directly contradicts the abstract's 'calibrated uncertainty' and the 'verified below' in §6. Since calibrated uncertainty is a stated contribution (C3 and the inverse-estimator sections), this is a load-bearing omission. The authors should either add the coverage experiment on synthetic data (which they note is fully possible) or explicitly revise the","section":"§6 vs §7.4/§9"},{"comment":"The claim that 'the stability law survives surrogate approximation' (corr(log σ_min(M̂), L1) ≈ −0.44 in Test B) is weakened by the same surrogate-bias issue: σ_min(M̂) is computed from the learned operator, and no true-map counterpart is given for the same geometries. Moreover, the negative finding about Corollary 4 — that κ_s shows no dependence on a/t on FEM data — is based on the surrogate-computed κ_s. If the surrogate smooths or distorts the depth dependence of the profiles, that negative conclusion could be an artifact. The true FEM profiles are available; a direct computation of κ_s and σ_min from M_FEM on the 1000 test geometries would settle both points and should be added.","section":"§7.4 Test B / §7.3 Corollary 4 test"}],"minor_comments":[{"comment":"The phrase 'calibrated uncertainty' in the abstract and C3 overstates what is delivered; §7.4 explicitly defers coverage. Please align the wording with the actual content.","section":"Abstract and C3"},{"comment":"The first sentence says κ_s is computed 'over the test geometries,' but Figure 2 and the text then say 'over n_geom=4000 training geometries.' Please clarify which set is used.","section":"§7.3"},{"comment":"The column 'B (M̂)' lacks a clear header for the counts; the n(A) column is explained only in the text. Also, the Test B stratum counts (12480/4020/3500) appear only in prose; adding them to the table would make the discrepancy visible and easier to discuss.","section":"Table 2"},{"comment":"The statement 'the degenerate direction of Proposition 2 converges to (1,−1,0)' is stronger than the proof: the proof shows that (1,−1,0) is a near-null direction, not that the least singular vector converges to it. Please soften or provide a limit argument.","section":"Corollary 4"},{"comment":"The text says the pipeline is 'released' (§3.3) but the final section says it 'will be deposited ... upon publication.' Please clarify the current availability; a preprint with a DOI or repository link would help reproducibility.","section":"Data and code availability"}],"recommendation":"major_revision","confidential_remarks":"The theoretical contribution is real and the paper is written with unusual honesty about its limitations. The two load-bearing gaps are both fixable: (i) validate the empirical κ_s/σ_min statistics against the true FEM profiles, and (ii) either run the planned coverage experiment or revise the calibration claims. If the authors supply those, the paper could be acceptable. The self-citation to the INCRT architecture is peripheral and not problematic; the main risk is that the headline 30% ill-posed fraction and the 'calibrated uncertainty' promise overstate what is currently demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the theorem is correct and worth knowing; the applied packaging overreaches in two places that are fixable but real.\n\nThe genuinely new thing here is Theorem 1: an exact characterization of when the load mix is recoverable from the SIF profile, via a Gram determinant / functional area, plus the sigma_min vs kappa_s distinction in Proposition 3. The proof is clean, model-free, and the singular-value identity (sigma_min sigma_max = A/sqrt(3)) is a nice observation. The empirical demonstration that kappa_s alone anti-correlates with recovery error under noise, while sigma_min tracks it, is convincing on the true forward map (Test A'). I also give them credit for reporting the Newman-Raju failure on FEM data and explicitly bounding the scope to synthetic validation.\n\nThe soft spots are real, though. First, the calibration claim. The abstract and Section 6 say the estimator 'returns calibrated uncertainty' and that the claim is 'verified below'; Section 7.4 then says credible-region coverage is 'stated as a falsifiable prediction' rather than reported. That is a direct mismatch. The paper does not instrument the set-valued head, so the headline 'calibrated uncertainty' is not actually tested. It's fine to defer it, but the abstract should say so.\n\nSecond, the empirical identifiability split — median kappa_s ~20, ~30% ill-posed — is computed from the learned surrogate M-hat, not from the true FEM profiles, and no sensitivity analysis is given. Table 2 actually shows a warning sign: for the true map, 6000/20000 draws (30%) fall in the ill-posed stratum; for the learned map on the same geometries, only 3500/20000 (17.5%) do. That is a sizable discrepancy the paper does not discuss. Since kappa_s is built from differences between load profiles, surrogate bias on those differences could easily shift the 70/30 split. The authors have the true FEM profiles for the test set; they should have compared kappa_s(M-hat) with kappa_s(M_true) directly. This does not touch the theorem, but it does mean the dataset-level claim is not yet established.\n\nOverall: this is a good, useful paper for anyone working on inverse fracture diagnostics or SIF surrogates, and it deserves a careful referee. The theory is solid, the exposition is honest about most limitations, and the sigma_min result is transferable. But the abstract overclaims calibration, and the empirical split needs a surrogate-bias check. I'd recommend peer review, with major revision.","headline":"Solid identifiability theorem; the empirical 70/30 split and the calibration claim are both softer than the abstract suggests.","tokens_in":17932,"tokens_out":2942,"would_cite":true,"duration_ms":28514,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65J20","65F35"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a known crack geometry, a combined tension–bending–bearing load is recoverable from a stress-intensity profile exactly when the three unit-load profiles form a non-degenerate triangle in function space; the paper proves this equivalence","keywords":["stress-intensity factor","inverse problems","identifiability","crack front","load mix recovery","simplex conditioning","finite-element benchmark"],"falsifier":"Compute κ_s and σ_min directly from the exact finite-element profile matrices (not the learned surrogate) on the same held-out geometries and re-run the noise study; if the median κ_s, the ill-posed fraction, or the negative correlation between σ_min and error are not reproduced, the empirical transfer from surrogate to real profiles fails.","tokens_in":16874,"feed_emoji":"📐","tokens_out":10694,"duration_ms":88686,"temperature":0.7,"pith_summary":"The paper asks when the relative magnitudes of tension, bending, and bearing loads acting on a crack can be recovered from the stress-intensity-factor profile along the crack front. For a known geometry the forward map is exactly linear, so identifiability reduces to a single geometric question: whether the three elementary load profiles are linearly independent as functions along the front. The paper proves a closed-form characterization: a functional area that vanishes exactly in the unidentifiable regime, and shows that under noise the recovery error is controlled by the smallest singular value of a stability margin, not by the conditioning number alone. On a public finite-element benchmark of corner cracks, the typical geometry is well-posed (median simplex-restricted conditioning ≈ 20) while roughly 30% of geometries are genuinely ill-posed; a point estimate is therefore reliable on the majority and provably uninformative on the rest. The paper also contributes a forward surrogate and a calibrated set-valued inverse estimator, validated on controlled synthetic noise, and explicitly makes no forensic claim on real components.","feed_headline":"A zero-area triangle makes crack loads unidentifiable","feed_subtitle":"A zero-area triangle means tension, bending, and bearing are indistinguishable; ~30% of geometries are unidentifiable.","key_machinery":"The profile matrix M(g) = [K_T | K_B | K_P] whose columns are the three unit-load profiles along the crack front, together with the functional area A(g) = √det G of the Gram matrix of the difference profiles. A(g) is twice the area of the triangle the three profiles span in function space; Theorem 1 ties it to the singular values of MB — σ_min·σ_max = A/√3 — so this single scalar simultaneously controls injectivity, conditioning, and noise-robustness. The same M(g) also serves as the differentiable forward map inside the inverse estimator, closing forward and inverse problems on one object.","core_discovery":"The central result, Theorem 1, is a four-way equivalence for a fixed geometry: the forward map is not injective on the affine hull of the load simplex; the functional area A(g) = √det G equals 0, where G is the Gram matrix of the difference profiles D1 = K_T − K_P and D2 = K_B − K_P; the three elementary load profiles are affinely dependent in L²(μ); and the smallest singular value of MB vanishes, where M is the profile matrix and B is an orthonormal basis of the sum-zero plane. Quantitatively, σ_min(MB)·σ_max(MB) = A(g)/√3 and the simplex-restricted conditioning κ_s(M) = √3 σ_max(MB)²/A(g), so κ_s diverges exactly as A(g) → 0. Proposition 3 proves that the recovery error under noise obeys ‖","pith_inferences":["The same A(g) criterion could serve as a cheap a priori screening tool: compute the profile triangle area once per geometry to flag load mixes that are structurally undecidable before any expensive inverse solve or experimental campaign.","Adding a second crack front or an independent measurement channel would likely restore identifiability in the ill-posed minority, because it increases the effective rank of M(g) and shrinks the degenerate direction — a natural candidate for the paper's proposed twin-crack extension.","Because the theory is model-free, it transfers to any linear-elastic superposition problem with an unknown coefficient vector — for example separating residual-stress contributions or mixed-mode SIF components — whenever the forward operator is exactly linear.","The empirical failure of the analytical shape-function prediction suggests that surrogate-based identifiability maps, computed from learned M(g), could be used to audit other analytic shape-function models, turning the released pipeline into a validation loop for the forward side."],"forward_implications":["A point estimate of (α,β,γ) is meaningful precisely when A(g) > 0; if A(g) = 0, no estimator can separate the loads and any point value is an artefact of the solver.","Under noise, recovery error scales as σ/σ_min(MB); the conditioning number κ_s can even anti-correlate with error across real geometries because σ_max varies, so uncertainty must be reported via σ_min.","On the public corner-crack data, the typical geometry is well-posed (median κ_s ≈ 20) but roughly 30% are ill-posed (κ_s > 50); the calibrated estimator widens its credible region along the degenerate direction in exactly those cases.","The shallow-crack collapse predicted by an analytical shape-function model is not observed systematically in the finite-element data; identifiability is governed by the full profile geometry, not relative depth alone.","With the true forward map, the inverse is exact at zero noise (error ~10⁻¹⁵ in every κ_s stratum), and the same σ_min-controlled ordering persists when the learned surrogate supplies M(g), showing the stability law survives surrogate approximation."],"fun_headline_variants":["Zero-area triangle makes crack loads unknowable","When SIF triangle collapses, loads become a guess","A flat triangle dooms crack load identification","Crack loads vanish from SIF when triangle area is zero","Unidentifiable crack loads: the zero-area triangle effect"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The empirical claims about the real dataset — the median κ_s ≈ 20, the ~30% ill-posed fraction, and the σ_min-controlled error correlations — are computed from the learned surrogate profiles, and the paper gives no sensitivity analysis establishing that κ_s and σ_min are stable to surrogate error.","fun_headline_variants_meta":{"raw":{"variants":["Zero-area triangle makes crack loads unknowable","When SIF triangle collapses, loads become a guess","A flat triangle dooms crack load identification","Crack loads vanish from SIF when triangle area is zero","Unidentifiable crack loads: the zero-area triangle effect"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":1893,"prompt_tokens":832,"completion_tokens":1061,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":996}},"tokens_in":576,"tokens_out":1061,"duration_ms":11033,"temperature":1.0,"reasoning_tokens":996,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:04:31.626866+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute κ_s and σ_min directly from the exact finite-element profile matrices (not the learned surrogate) on the same held-out geometries and re-run the noise study; if the median κ_s, the ill-posed fraction, or the negative correlation between σ_min and error are not reproduced, the empirical transfer from surrogate to real profiles fails.","supporting_citations":[],"review_version":1}