{"id":"6f1854ba-d22c-48e1-b565-e1b526e01845","arxiv_id":"2607.27741","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Combining time-of-flight, photon energy, and the QED decay law in a Bayesian posterior localizes simulated three-photon positronium decays to 1.62 cm mean error, roughly twice as good as timing-only reconstruction.","lead":"TRIO is a new algorithm that estimates where an ortho-positronium atom decayed inside a PET scanner by fusing the three photons' arrival times, energies, and quantum-mechanical decay correlations into one Bayesian calculation. In simulations modeled on a clinical scanner it localizes decays to about 1.6 cm, roughly twice as well as timing alone, without needing a prompt photon.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"QED prior is neither correctly derived (Eqs. 25–26 omit Jacobian) nor ablated; the claimed factor-of-two could come solely from timing×energy likelihood fusion, so the central novelty claim is unsupported.","rationale":"The paper's strongest claim has two parts: an empirical improvement (1.62 cm vs 3.05 cm) and a novelty based on the QED prior. The empirical gain from fusing time and energy information is plausible and the algebraic core of the likelihoods is checkable, but the attribution to the QED prior is not tested. The reader's weakest assumption identified exactly this: the missing Jacobian, normalization, and the untested marginal contribution. My stress-test agrees with that assessment. The concern is load-bearing because the abstract explicitly markets the prior as the new contribution; if the prior is inert or incorrectly derived, the central novelty claim is unsupported even though the method may still work as a timing×energy fusion. I kept the verdict UNCHANGED because the simulation result is not invalidated by the missing ablation—it only becomes conditional on a component the paper does not isolate. The proposed concrete test is minimal and would settle whether the prior matters; if it does not, the paper should be reframed or the prior should be corrected. If it does, the Jacobian and normalization issues still need fixing before the 'probabilistic framework' claim can stand. The paper is honest about simulation-only validation, point sources, and true coincidences, which is credit, but those limitations do not substitute for an ablation of the headline component.","tokens_in":14878,"tokens_out":7340,"duration_ms":82897,"concrete_test":"Rerun the exact Monte Carlo pipeline of Sec. 3 (500,000 events per source, same Quadra-like geometry and fminsearch settings) with Eq. (9) modified by setting P(x) ≡ constant while keeping P(x_t|x) and P(x_E|x) unchanged. If the mean ε over the 21 point sources remains 1.62 cm within statistical error, the QED prior contributes nothing to the headline improvement and the 'for the first time' claim collapses. If ε worsens substantially, repeat with the Jacobian-corrected prior P(x) = P(˜E(x)) |∂E/∂x| to determine whether the current mis-specified form is responsible for the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that TRIO's improvement comes from the first physics-informed QED prior within a unified Bayesian framework. But the paper never compares Eq. (9) with P(x) against the same posterior with P(x) constant. In Sec. 5 the factor-of-two improvement is attributed to the complementarity of the timing and energy likelihoods, not to the prior. Thus the reported 1.62 cm mean error may be entirely due to fusing P(x_t|x) and P(x_E|x), a known-ingredient combination, with the QED prior contributing nothing. The prior itself is also not derived as a spatial density: Eq. (25)–(26) set P(x)=P(˜E(x)) without the Jacobian |∂(E1,E2)/∂(x,y)| needed to transform the QED energy-space PDF into a position-space PDF. As printed, Eq. (25) has no normalization constant κ and diverges at the simplex boundary like 1/E_j^2; this divergence would push the MAP estimate toward the triangle edges, which is not a demonstrated QED property. Finally, the 'prior' is computed from the measured hit positions x_i via Eq. (22), so it is not a data-independent Bayesian prior. Each of these flaws is fixable, but the decisive gap is that the prior's marginal contribution is untested, which leaves the novelty claim without empirical support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TRIO, a Bayesian maximum-a-posteriori reconstruction algorithm for the vertex position of ortho-positronium three-photon decays in PET. The posterior combines three ingredients: a time-based trilateration likelihood P(x_t|x), an energy-based reconstruction likelihood P(x_E|x), and a prior P(x) derived from the QED energy-angle correlations in o-Ps decay. The authors derive closed-form likelihood models under Gaussian measurement noise, then use Nelder-Mead optimization to find the MAP estimate. In Monte Carlo simulations modelled on the Siemens Biograph Quadra, TRIO reports a mean position error of 1.62 cm, versus 3.05 cm for time-only trilateration and about 18 cm for energy-only reconstruction. The paper claims this is the first unified probabilistic framework to include a physics-informed QED prior for three-photon imaging.","tokens_in":1859,"tokens_out":2355,"duration_ms":51239,"significance":"If the central claim were fully supported, the work would have practical relevance for event-by-event localization of three-photon o-Ps decays in conventional TOF-PET scanners. The manuscript also has clear strengths: the likelihood derivations in Secs. 2.1 and 2.2 are mostly explicit and checkable, the simulation setup includes realistic detector geometry, DOI and resolution parameters, and comparisons against TReco and EReco baselines are provided. However, as discussed below, the contribution of the QED prior—the claimed novelty—is not established by the reported experiments, and the prior as formulated in Sec. 2.3 has a mathematically non-normalizable form and an unjustified change of variables. The paper is therefore of moderate significance at best until these load-bearing points are addressed.","major_comments":[{"comment":"The prior transformation x -> E(x) omits the Jacobian determinant |∂E/∂x|. Even if P(E) is the QED energy-space density, the correct spatial density on the decay plane is P(x) = P(E(x)) · |∂E/∂x|. Without this factor, the printed P(x) is not the push-forward of the Ore–Powell density, so the claim that the prior 'encodes' the QED correlations is not mathematically justified. Additionally, Eq. (25) as written has no explicit normalization constant and the summand diverges like 1/E_j^2 at the simplex boundary E_j -> 0; thus P(x) is not normalizable over the domain of feasible positions. This non-integrability may push the MAP estimate toward the triangle edges in an uncontrolled way. The authors should present the correct Jacobian-transformed density and demonstrate that it is normalizable (or explicitly state that they use an unnormalized heuristic).","section":"Sec. 2.3, Eqs. (25)–(26)"},{"comment":"The paper's central novelty is the physics-informed QED prior, but the reported factor-of-two improvement is explicitly attributed in Sec. 5 to 'the complementary nature of the two likelihood terms', not to the prior. No experiment is reported that runs the same posterior with P(x) = constant and compares it to the full TRIO posterior. Without this ablation, the claim that the QED prior contributes any localization improvement is unsupported; the observed gain could come entirely from fusing P(x_t|x) and P(x_E|x), an ingredient that is not claimed as new. This is a decisive gap for the abstract's 'for the first time' claim and should be addressed with a quantitative comparison in the results.","section":"Sec. 5 and Fig. 2 (ablation missing)"},{"comment":"The prior P(x) is evaluated from the geometry of the event: the candidate position x and the measured hit positions x_i enter through the distances r_i and angles α_ij in Eqs. (15) and (22). Thus P(x) is not data-independent; it depends on the measured interaction positions of the same event. The text says the prior is 'independent of the measured times and energies', but that is not the same as being data-independent. If the hit positions are treated as fixed conditioning variables, then the posterior should be written as P(x | x_t, x_E, x_1, x_2, x_3) and the factor P(x | x_1, x_2, x_3) should be separately specified. As written, the notation P(x) and the claim of a 'pure physical constraint' conflate a data-conditioned geometry factor with a Bayesian prior, which risks double-counting the position information and weakens the statistical interpretation of Eq. (9).","section":"Sec. 2.3, Eq. (22) and usage of data in the prior"},{"comment":"Even aside from the Jacobian issue, the functional form in Eq. (25) is not self-consistent: κ is introduced as a normalization constant but never defined or computed. The expression Σ [(m_ec^2 - E_i)/(E_j E_k)]^2 is not integrable over the full simplex because of the 1/E_j^2 singularities. If the authors intend to use a truncated or regularized version, that must be stated explicitly and its effect on the reconstruction must be assessed. As printed, the discussion of a 'normalization constant κ' is not supported by the formula.","section":"Sec. 2.3, Eq. (25) normalization"},{"comment":"The reported errors are mean Euclidean distances over 500,000 simulated events per source point. The paper does not report standard errors, confidence intervals, or a statistical comparison across source positions. Given that the TRIO improvement versus TReco is roughly a factor of two and the simulation has very large N, the difference is probably significant, but the reader cannot check whether the improvement is uniform or driven by particular source locations. Reporting the distribution of errors (e.g., median, quartiles) or per-position values in tabular form would strengthen the claim, especially since the prior ablation requested above will require such statistics.","section":"Sec. 4 and Sec. 5, statistical assessment"}],"minor_comments":[{"comment":"'for the first time' appears in the abstract and Sec. 1. The phrase is strong and should be conditional on the corrected prior derivation and ablation; otherwise it is vulnerable to the issues in Sec. 2.3.","section":"Abstract"},{"comment":"There is a formatting typo: the formula has an open parenthesis in the text but the mathematical expression is clear. Please check the equation typesetting.","section":"Eq. (17)"},{"comment":"There is a minor spacing typo in '˜Ein' and the sentence ending in 'decay pointx.' should be rephrased for clarity.","section":"Sec. 2.3, first paragraph"},{"comment":"The top-left panel is labelled 'Prior distribution P(x)', but as noted in the major comments this is actually a data-conditioned geometry factor. The caption should be changed to reflect the correct interpretation.","section":"Fig. 4 caption"},{"comment":"The phrase 'with the NM optimization performed using the MATLAB fminsearch function from the Optimization Toolbox with the default parameter settings' is useful, but no convergence criteria or initialization strategy for the NM search is described. Adding this would improve reproducibility.","section":"Sec. 3"},{"comment":"Ref. [38] is described as a preprint; if this is the source of the closed-form energy solution, its status and availability should be clarified.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's engineering contribution—a full Bayesian posterior combining timing and energy likelihoods—is plausible and the simulations appear carefully set up. However, the paper currently overclaims the novelty of the QED prior. The prior transformation is not a proper change of variables, the printed density is non-normalizable, and the experiments never isolate the prior's contribution. These are fixable within the scope of a revision: add the Jacobian factor, show normalization or explicitly use a regularized heuristic, and run a flat-prior ablation. Until those are provided, the central 'for the first time' claim should not be accepted as supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is better than its abstract. The simulation result—that multiplying a time-likelihood and an energy-likelihood gives sharper vertex localization than either alone—is plausible and clearly demonstrated. The authors are honest about the simulation-only scope, point sources, and true coincidences; the limitations are spelled out. The algebra in Sections 2.1 and 2.2 is checkable and mostly sound. If you work on positronium imaging, this is a useful new baseline.\n\nThe problem is the thing they sell as the novelty: the QED spatial prior. They define a map x → E(x) and then set P(x) = P_QED(E(x)) without the Jacobian |∂E/∂x|. That is not the push-forward of the QED density; it is a different function, and a MAP optimized under it is not a proper Bayesian MAP with that prior. This is a real mathematical slip, not a notational choice. Also, they never run a no-prior ablation. The paper's own discussion and conclusion attribute the factor-of-two gain to the complementarity of the two likelihoods, not to the prior. So the claim that the QED prior contributes is unsupported. It might contribute, might not, or might push solutions toward the triangle edges where the QED density diverges; we just don't know.\n\nThe other gaps are minor by comparison: the Nelder-Mead initialization is unspecified, no code or data is shipped, and the simulator is only described by name. Those should be addressed in revision, but they do not undermine the main empirical pattern.\n\nShould it go to peer review? Yes. The flaws are fixable and the core idea—event-by-event Bayesian localization using time and energy likelihoods on the decay plane—is worth giving to the community. The revision should fix the prior derivation or demote the prior to a heuristic weighting, and add the no-prior ablation. If the improvement stands without the prior, the paper still makes a contribution, but with more honest framing.","headline":"A useful Bayesian fusion of timing and energy for three-gamma PET, but the claimed QED prior is unproven and its derivation skips a Jacobian.","tokens_in":15788,"tokens_out":5714,"would_cite":true,"duration_ms":59235,"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":"This paper claims that a Bayesian combination of photon timing, energy, and quantum-electrodynamics decay-law correlations can localize three-photon ortho-positronium annihilations in a clinical PET scanner to about 1.6 cm, roughly twice as","keywords":["positron emission tomography","ortho-positronium","three-photon annihilation","Bayesian image reconstruction","time-of-flight PET","maximum a posteriori","QED prior","vertex reconstruction"],"falsifier":"Run TRIO twice on the same simulated events, once with the QED prior P(x) as printed and once with P(x) uniform; if mean position error stays at ~1.62 cm in both, the QED prior is not responsible for the gain. Separately, numerically integrate κ Σ_i [(m_e c²−Ẽ_i)/(Ẽ_j Ẽ_k)]² over the feasible decay plane with the change-of-variables Jacobian; a divergent integral would show the printed prior is not a normalizable density.","tokens_in":14631,"feed_emoji":"⚛️","tokens_out":5062,"duration_ms":50396,"temperature":0.7,"pith_summary":"TRIO treats each three-photon ortho-positronium decay as a Bayesian inference problem: the vertex position is estimated by maximizing a posterior that multiplies a time-based likelihood, an energy-based likelihood, and a prior derived from the QED amplitude for three-photon decay. The paper's central claim is that this fusion works at the modest energy resolution of a current clinical PET scanner and that the QED prior is a genuinely new ingredient, not just a repackaging of data. In simulations of a long-axial-field-of-view scanner, TRIO achieves a mean position error of 1.62 cm, versus 3.05 cm for timing trilateration and about 18 cm for energy-only reconstruction. If correct, this would make three-photon imaging practical with standard radionuclides such as 18F, because no prompt photon is required.","feed_headline":"Three-photon PET imaging reaches 1.62 cm precision","feed_subtitle":"Combining timing, energy, and QED decay physics beats timing-only reconstruction by about 2x.","key_machinery":"The load-bearing object is the Bayesian posterior of Eq. (9), built from three densities: a Gaussian time likelihood P(xt|x) comparing measured arrival times with times predicted from a candidate vertex; a Gaussian energy likelihood P(xE|x) comparing measured energies with energies predicted from the candidate vertex via the opening angles among the three hits; and the QED prior P(x) = κ Σ_i [(m_e c² − Ẽ_i)/(Ẽ_j Ẽ_k)]², evaluated at the vertex-predicted energies. Maximizing this product with a derivative-free simplex search explores the two-dimensional decay plane—the plane containing the three hits and the vertex, enforced by momentum conservation—for the most probable position. The QED pri","core_discovery":"On its own terms, the discovery is that the three measured ingredients of a triple-coincidence event—hit positions, arrival times, and deposited energies—are best combined multiplicatively rather than by choosing one reconstruction. The authors write the posterior P(x|xt∩xE) ∝ P(xt|x) P(xE|x) P(x), where P(xt|x) scores how well a candidate vertex explains the measured times, P(xE|x) scores how well it explains the measured energies, and P(x) assigns prior weight according to the QED prediction for the relative probability of the photon momentum configuration implied by the candidate vertex. They report that this product sharpens the localization in directions where the individual likelihoods","pith_inferences":["A correct Bayesian transfer of the QED energy-space density to position space requires the Jacobian of the map x→Ẽ(x); the paper does not include it, so the printed prior is not strictly the QED density pulled back to the plane. Without the Jacobian (or a normalization check), the 'physics-informed prior' label is stronger than the math shown.","The reported improvement could in principle come entirely from the timing×energy likelihood product; the paper does not run a control with the prior removed. A uniform-prior ablation would settle whether the QED term, not just the fusion, earns the gain.","For real scanners, hit-position uncertainty is not negligible as assumed; the conditional-independence claim weakens with depth-of-interaction blur. A joint likelihood over positions, times, and energies would be the next step.","If TRIO's vertex accuracy holds on experimental data, the same event-by-event localization could serve as a veto or selection tool in fundamental o-Ps symmetry tests, a direction the paper only sketches."],"forward_implications":["Three-photon o-Ps imaging no longer needs a prompt photon, so standard radiotracers like 18F-FDG can in principle supply this channel.","Combining time and energy information in the posterior gives about a twofold accuracy gain over timing-only trilateration at current clinical energy resolution.","Even when energy-only reconstruction is inaccurate (~18 cm error), the energy likelihood still constrains the posterior in a direction complementary to timing.","The framework automatically adapts to the relative quality of timing versus energy, so it should transfer to scanners with different resolution trade-offs.","Per-event reconstruction is independent and parallelizable; the observed ~61 ms single-CPU time is not a fundamental limit."],"fun_headline_variants":["Bayesian 3-photon PET localizes to 1.62 cm","Physics-informed prior sharpens triple-photon PET","Merging time, energy, and QED lifts PET precision 2x","Three-photon PET: Bayesian fusion beats timing-only"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim depends on the assumption that plugging the geometry-predicted photon energies into the QED energy-space distribution yields a valid probability density over decay positions, which requires a Jacobian and a finite normalization that the paper does not provide.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian 3-photon PET localizes to 1.62 cm","Physics-informed prior sharpens triple-photon PET","Merging time, energy, and QED lifts PET precision 2x","Three-photon PET: Bayesian fusion beats timing-only"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1111,"prompt_tokens":801,"completion_tokens":310,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":246}},"tokens_in":545,"tokens_out":310,"duration_ms":4073,"temperature":1.0,"reasoning_tokens":246,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:13:32.766133+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TRIO twice on the same simulated events, once with the QED prior P(x) as printed and once with P(x) uniform; if mean position error stays at ~1.62 cm in both, the QED prior is not responsible for the gain. Separately, numerically integrate κ Σ_i [(m_e c²−Ẽ_i)/(Ẽ_j Ẽ_k)]² over the feasible decay plane with the change-of-variables Jacobian; a divergent integral would show the printed prior is not a normalizable density.","supporting_citations":[],"review_version":1}