{"id":"001eb7a2-da3d-49a0-beba-ebd011c2bfae","arxiv_id":"2607.23358","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Fanpy 2.0 ships modular Python classes for dozens of coupled-cluster, geminal, CI, and neural-network wavefunction ansätze, with a FANPT continuation tool to aid convergence.","lead":"Fanpy 2.0 is an open-source Python release that adds new wavefunction ansätze and a perturbation-theory continuation tool for prototyping quantum chemistry methods. A quantum chemist who wants to test a new wavefunction form without rewriting a production codebase might use this as a fast prototyping platform.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FANPT's advertised convergence benefit is unsupported: no numerical demonstration or Jacobian-conditioning check appears in the paper, and the only worked example reports no computed energy.","rationale":"The reader's weakest assumption identifies FANPT's effectiveness as the most exposed part of the central claim. I agree: the paper makes a concrete practical claim about FANPT as a convergence aid, but gives only an algorithmic description and a usage snippet. The absence of any numerical result is especially notable because this is a software-release paper; the central claim is that the platform works, and the natural evidence would be converged energies on test systems. The paper does not provide that. This is a missing-validation concern rather than a claim of internal inconsistency: the code is open source, there is an automated test pipeline, and the modular design is clearly described. Those count in the paper's favor. But they do not replace a numerical demonstration. Since the reader already assigned CONDITIONAL, my analysis does not move the verdict; it reinforces it. A single targeted FANPT run on a small strongly correlated system with condition-number monitoring would settle whether the advertised benefit is real.","tokens_in":18649,"tokens_out":4952,"duration_ms":51990,"concrete_test":"Run the FANPT driver on H4 in a square geometry (STO-3G, 4 electrons, 8 spin-orbitals) with sr-CCSD(0), stepping lambda=0 to 1 in 10 steps at second order, and print the final projected energy and the condition number of J(lambda) for every intermediate lambda. Compare with the same calculation initialized directly from HF parameters, as in Section III. If FANPT does not converge to the same or lower energy, or if J(lambda)'s condition number becomes large enough to make Eq. (6) numerically unreliable, the convergence-aid claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim that Fanpy 2.0 is a working, extensible platform, a reader needs to see that the newly added ansatze and FANPT actually produce converged results. The manuscript does not supply that evidence. Section II.B asserts FANPT is 'particularly useful as a convergence aid' for strongly correlated systems, but provides no convergence study, no error bound, and no numerical example. The implementation solves J(lambda) x^(n) = b^(n) at each lambda step (Eq. 6); if J(lambda) is singular or ill-conditioned anywhere along the homotopy path, the propagated guess is undefined or numerically unstable, and the paper does not report condition numbers, step-size checks, or fallback behavior. The only worked example (Section III) walks through a BeH2/STO-3G AP1roGSD calculation but prints no total energy and makes no comparison to FCI or pCCD, so it demonstrates code execution, not correctness. The open-source repository and automated tests are real mitigating evidence, but they do not by themselves establish that the advertised features perform as claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes version 2.0 of Fanpy, an open-source Python library for prototyping multideterminant wavefunctions based on the FANCI formulation. The authors claim new wavefunction families (coupled-cluster-inspired geminals, seniority-restricted coupled cluster, extended-hierarchy CI, and neural-network ansätze), a FANPT perturbative-continuation module, redesigned PySCF and new PyCI interfaces, an analysis module, and an improved development environment. The only worked example is a BeH2 AP1roGSD calculation presented as a code listing; no total energy, convergence diagnostic, or comparison to reference results is reported.","tokens_in":18938,"tokens_out":4877,"duration_ms":45375,"significance":"If the implementation lives up to its description, Fanpy 2.0 would be a valuable platform for quickly prototyping wavefunction methods. Strengths of the manuscript are the modular FANCI architecture, the explicit design that reduces new ansatz implementation to writing an overlap function, the availability of source code on GitHub, and the automated test pipeline. However, the paper's central claim that these features are working and usable is not supported by any numerical evidence in the manuscript. The FANPT convergence-aid claim in particular is purely asserted. Because the manuscript is a software paper, the absence of benchmarks or even a printed energy from the BeH2 example prevents the reader from verifying correctness, and the theoretical content is otherwise a summary of the authors' prior work.","major_comments":[{"comment":"The only numerical demonstration is a code listing that ends with the least_squares call. No total energy, gradient norm, or comparison to FCI or pCCD is printed. A software paper claiming 'working' implementations must show at least one complete, reproducible calculation with a verifiable output. Please add the computed energy and a reference comparison for this example, and ideally a small benchmark for each new wavefunction family.","section":"III, Example calculation"},{"comment":"FANPT's advertised role as a convergence aid for strongly correlated systems is not demonstrated. Each step requires solving J(lambda)x^(n)=b^(n); if J is singular or ill-conditioned along the homotopy path, the propagated guess is undefined or unstable. No condition numbers, step-size selection, or fallback behavior are reported, and no numerical comparison of optimization with vs. without FANPT is given. Provide a concrete case study and report these diagnostics.","section":"II.B, Eq. (6)"},{"comment":"The four new wavefunction families (geminal CC-inspired, sr-CC, ehCI, NN) are described with equations and code snippets, but no validation is presented. For each family there is no test calculation showing that overlaps/derivatives are correct or that optimization converges to a sensible energy. Since the paper's thesis is that users can rely on this infrastructure to prototype new ansätze, at least one representative numerical test per family is required to support the claim.","section":"II.D, New wavefunctions"}],"minor_comments":[{"comment":"'The overall workflow is shown in Fig 2' should refer to Fig. 3; Fig. 2 is the bitstring illustration.","section":"II.B, Figure reference"},{"comment":"Typos: 'easly' -> 'easily', 'flexibilty' -> 'flexibility', 'sensivity' -> 'sensitivity', 'safe guard' -> 'safeguard', 'numpyarrays' -> 'numpy arrays'.","section":"I, II.A, II.E"},{"comment":"The claim that the multiplicative readout 'ensures size consistency' is stated without proof or demonstration; at minimum give a short explanation or a numerical check.","section":"II.D.4"},{"comment":"Reference 61 is incomplete: 'The Journal of Physical Chemistry A0, null (0)' lacks volume and page numbers.","section":"Note 61"}],"recommendation":"major_revision","confidential_remarks":"This is a software paper. The core issue is the absence of any numerical validation, which is fixable within the manuscript's scope. The manuscript heavily cites the authors' own theoretical papers for the implemented methods; this is not a circularity problem for the software claim, but the paper should make clear which parts are new implementation versus prior theory. If the authors provide benchmarks, the paper could be suitable for publication in this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a genuine software release: the code is open source, the architecture is clean, and the modular FANCI design is a real strength. Second, the paper does not show a single converged energy, and the FANPT convergence claim is asserted, not demonstrated. That gap is the whole ballgame for a paper whose abstract promises a working platform.\n\nWhat is actually new: Fanpy 2.0 bundles several previously published wavefunction families (CC-inspired geminals, sr-CC, ehCI, NN ansatz) into one extensible framework, adds a FANPT continuation module, a PySCF/PyCI interface layer, an analysis module, and an automated test pipeline. The implementation is nontrivial and the modular design means a user can prototype a new ansatz by writing one overlap function and reuse everything else. That is a legitimate contribution, and the GitHub repository with tests is real evidence that the software exists and is maintained. I see no red flags of circular reasoning or misconduct; self-citation is heavy but appropriate for a paper that implements those earlier methods.\n\nThe soft spots are proportionate. The most serious: no numerical validation anywhere. The BeH2/STO-3G example walks through setup and optimization but prints no energy and compares nothing. The FANPT section claims it is 'particularly useful as a convergence aid' for strongly correlated systems, but there is no convergence study, no error bound, no condition-number check for J(lambda) in Eq. (6), and no fallback behavior when the Jacobian is singular. That is a load-bearing weakness because if FANPT fails or is ill-conditioned, the tool may not work as advertised. The secondary weakness is that all the wavefunction families are implementations of the authors' own prior papers, so the novelty is architectural, not scientific. That is fine for a software paper, but it means the paper's value rests on whether the implementations are correct and efficient, which the manuscript does not verify.\n\nWho is this for? Someone in the quantum chemistry method-development niche who wants to prototype geminal, sr-CC, or NN wavefunctions without rewriting infrastructure. A serious referee should ask for a benchmark section with a few FCI-comparable systems, at least one FANPT demonstration showing convergence improvement (or an honest statement of when it fails), and code-level tests that compare against known values. The paper does not deserve desk rejection; it deserves review with a request for revisions.\n\nI would bring it to a reading group only if the group cares about software design for wavefunction prototyping. I would not cite it in my own work yet without first running the code and seeing numerical results.","headline":"A credible, well-architected software release that would benefit from a validation section; the central claim (the code works) is plausible but not demonstrated numerically.","tokens_in":19428,"tokens_out":1215,"would_cite":false,"duration_ms":15431,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fanpy 2.0 makes implementing a new wavefunction ansatz as simple as writing a single overlap function, then reuses the same optimization and analysis tools for every method.","keywords":["Fanpy","FANCI","wavefunction prototyping","coupled-cluster geminals","seniority-restricted coupled cluster","FANPT","projected Schrödinger equation","quantum chemistry software"],"falsifier":"Run the documented BeH2 example exactly as written: if the new AP1roGSDGeneralized wavefunction class, the FANPT driver, or the analysis printers fail to run or the energies do not match the reported nonvariational projected-Schrödinger solution, the paper's central claim of a working, extensible platform fails. Alternatively, apply FANPT to a molecule with a known strongly correlated regime and check whether the propagated parameters actually satisfy the projected Schrödinger equation at λ=1; singular Jacobians or diverging energies would refute the convergence-aid claim.","tokens_in":18621,"feed_emoji":"⚛️","tokens_out":4566,"duration_ms":39971,"temperature":0.7,"pith_summary":"The paper presents Fanpy 2.0, a Python library built on the FANCI idea that any multideterminant wavefunction is fully specified by its overlap with each Slater determinant. It claims that a researcher can turn a new mathematical ansatz into working code by writing just that overlap function (and its derivatives), reusing the same Hamiltonian, objective, solver, and analysis modules. Version 2.0 adds several new wavefunction families — coupled-cluster-inspired geminals, seniority-restricted coupled cluster, extended-hierarchy CI, and a neural-network ansatz — plus a perturbative continuation method (FANPT) for finding good starting parameters, a redesigned PySCF interface, an optional PyCI backend, and an analysis module. If true, the platform lowers the barrier to testing new wavefunction methods, especially for strongly correlated systems where initialization is hard.","feed_headline":"One overlap function now yields a working wavefunction method","feed_subtitle":"New geminal, seniority-restricted, and neural-network ansätze plug into one modular pipeline.","key_machinery":"The central object is the FANCI overlap function f(m,P)=⟨m|Ψ⟩ (Eq. 2), which defines a wavefunction by its overlaps with Slater determinants. Because objectives (variational energy, projected Schrödinger equation, local energy) are built from overlaps and Hamiltonian matrix elements alone, the entire optimization and analysis pipeline depends only on a wavefunction exposing get_overlap and parameter derivatives. The paper also introduces FANPT, which solves a linear response equation J(λ)x(n)=b(n) with J the Jacobian of the projected FANCI equations, to propagate parameters along a Hamiltonian interpolation; the Jacobian and response machinery are the mechanism that makes the perturbative co","core_discovery":"The central claim is that Fanpy 2.0 is a working, modular platform for prototyping arbitrary multideterminant wavefunctions. The FANCI formalism expresses any wavefunction as a parameterized expansion in Slater determinants whose coefficient function f(m,P)=⟨m|Ψ⟩ is the overlap with a determinant; implementing a new ansatz reduces to writing get_overlap. The release delivers four new wavefunction families, an implementation of FANPT as a continuation layer on top of projected Schrödinger objectives, an optional PyCI interface that offloads expensive determinant-space operations, and initial analysis tools such as natural-orbital computation and human-readable parameter printers. The BeH2 exa","pith_inferences":["The paper does not include numerical benchmark results for the new wavefunctions or for FANPT beyond a single example; if the new ansätze are accurate on strongly correlated test systems, Fanpy 2.0 would become a convenient testbed for comparing pair-based and excitation-based correlation treatments at scale.","FANPT's success hinges on the invertibility of the projected-equation Jacobian and on convergence of the perturbation series; if those fail in difficult regimes, the continuation strategy could be supplemented by homotopy or resummation methods, which the paper leaves open.","The analysis module's natural-orbital and parameter-printing tools lay groundwork for more interpretative features; one likely extension is automatic detection of dominant excitations or geminals across a calculation, which would make the workflow more useful for method development."],"forward_implications":["A researcher can implement a new wavefunction ansatz by writing one overlap function and immediately use all existing objectives, solvers, and analysis tools without modification.","The new coupled-cluster-inspired geminal family generalizes pCCD to broader excitation spaces and yields sixteen newly supported wavefunction forms via s_type and singles variants.","Seniority-restricted CC wavefunctions let users interpolate between pair-based and full CC expansions by filtering the cluster operator by excitation rank and seniority.","FANPT provides a systematic way to build initial guesses for projected Schrödinger equations by propagating parameters from a Fock reference to the full Hamiltonian, which should reduce initialization sensitivity for strongly correlated systems.","The optional PyCI backend allows computationally demanding determinant-space operations to be offloaded while keeping Fanpy's prototyping flexibility."],"fun_headline_variants":["One overlap function turns any ansatz into working code","From math to code: Fanpy 2.0 automates ansatz design","New wavefunction ansätze: write one overlap function","Fanpy 2.0: Plug in any ansatz via a single overlap function","Write get_overlap, get a working method: Fanpy 2.0"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"FANPT's usefulness as a convergence aid assumes that the Jacobian J(λ) stays invertible along the path and that the perturbation expansion converges for the strongly correlated systems where the tool is most needed; the paper offers no error bounds or numerical demonstration of either condition.","fun_headline_variants_meta":{"raw":{"variants":["One overlap function turns any ansatz into working code","From math to code: Fanpy 2.0 automates ansatz design","New wavefunction ansätze: write one overlap function","Fanpy 2.0: Plug in any ansatz via a single overlap function","Write get_overlap, get a working method: Fanpy 2.0"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000839,"raw_usage":{"total_tokens":3450,"prompt_tokens":656,"completion_tokens":2794,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":400,"completion_tokens_details":{"reasoning_tokens":2699}},"tokens_in":400,"tokens_out":2794,"duration_ms":20220,"temperature":1.0,"reasoning_tokens":2699,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:38:08.004922+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the documented BeH2 example exactly as written: if the new AP1roGSDGeneralized wavefunction class, the FANPT driver, or the analysis printers fail to run or the energies do not match the reported nonvariational projected-Schrödinger solution, the paper's central claim of a working, extensible platform fails. Alternatively, apply FANPT to a molecule with a known strongly correlated regime and check whether the propagated parameters actually satisfy the projected Schrödinger equation at λ=1; singular Jacobians or diverging energies would refute the convergence-aid claim.","supporting_citations":[],"review_version":1}