evidence/source_audit.md

Paper, verdict, and scope audit

Retrieved on 2026-07-23 with the explicit user agent OpenResearch-Reproduction/1.0 (paper audit; contact via GitHub MachineLearning-Nerd).

Source	Immutable locator	SHA-256
Judged paper	https://arxiv.org/pdf/2602.21565v1	3a8c57d9f739df274b21a65461db0cce87197d989721c67d118398649b5c9ac3
Current paper	https://arxiv.org/pdf/2602.21565v3	c126395ba290db35973a5bd3952daf1016696c203aaa4f67de805b99cb2abf41
Current HTML	https://ar5iv.labs.arxiv.org/html/2602.21565	52684605ba5a9992ca16ee1f9ecf56d75c92e2b152d5242f48a3424e3d012543
Live verdict dataset	ICML-2026-agent-repro/verdicts@main, filtered only by space_id == "DineshAI/qbe18sZPWS"	ff8bcc5eee67ae172d74d292df6f87691ef324380db0f3cc467a3e965efc77bc

The judged claims match arXiv v1. The current v3 changed several empirical numbers and added uncertainty estimates. For example, v1 Table 3 reports QM9 GAP-SA reward 0.876 (ours), 0.816 (MOGFN), and 0.805 (HN-GFN); v3 reports 0.873±0.003, 0.799±0.022, and 0.742±0.035. This campaign tests the exact v1 claims used by the judge and records v3 only as version context.

Exact anchors and quantifiers

• Claim 1: v1 Proposition 4.1, Section 4.2, and Appendix A.1. Given k GFlowNets on a shared state graph with terminal distributions p_i(x) ∝ R_i(x) and non-negative weights ω_i, Eq. (4) at β=1 realizes p_M(x) ∝ Σ_i ω_i R_i(x). Weights need not sum to one.
• Claims 2 and 5: v1 Section 5.1–5.2 and Table 1. The domain is the full 32×32 HyperGrid. For each k∈{2,3,4,5}, results average L1 over 128 evenly spaced simplex preferences. Ours is 0.003; MOGFN is 0.021/0.027/0.042/0.048; HN-GFN is 0.017/0.021/0.032/0.035; the no-reaching ensemble is 0.117/0.098/0.113/0.111.
• Claim 3: v1 Section 6.1–6.2 and Table 3. Atom-based QM9 GAP-SA uses β=32, 10 evenly spaced two-objective preferences, 128 generated candidates per preference, and mean scalarized reward of each preference's top 10. The reported aggregate is 0.876 for ours, 0.816 for MOGFN, and 0.805 for HN-GFN.
• Claim 4: v1 Section 6.2–6.3, Tables 4–5. Logical operators use 5,000 samples for target-bin accuracy. Timing averages 1,000 samples. Fragment logical composition takes 24.70 ms (2 objectives) and 27.02 ms (3) for ours versus 1789.11 ms and 1974.64 ms for classifier guidance, with Table 4 reporting higher target-bin percentages for ours in every listed fragment and QM9 composition.
• Claim 6: v1 Section 5.3 and Figure 3 (extended in Figure A6). The distortion is δ(x)=u_M(x)/N_M(x). The paper says it is not exactly constant for nonlinear operators but stays close to 1/Z_M where the composition value G(·) is large; deviations should concentrate where G(·) is small. Figure 3 defines outliers using [Q1−1.5·IQR, Q3+1.5·IQR].

Existing accepted evidence

The existing exact tabular suite is retained as a theorem-level regression. It enumerates all 1,024 terminal states and triangulates forward rollout with an independent edge-flow certificate. It does not substitute for the paper's trained-neural or molecule experiments.

Environment reconciliation

The official README describes two independent Python 3.10 environments. Its molecule dependency gflownet@v0.2.0 pins torch-geometric==2.4.0, whereas the grid dependency torchgfn==2.4.1 requires torch-geometric>=2.6.1. The campaign requirement is exactly one repository .venv, so the lock uses torch-geometric==2.6.1 for both codebases. This is an explicit compatibility override, not an experimental variable; import and behavior checks must pass before molecule evidence is accepted.

The same README says torch==2.1.2, but the published torchgfn==2.4.1 package metadata requires torch>=2.6.0. The single lock therefore pins torch==2.6.0 and treats the paper's upstream GFlowNet commit as the fixed code source. This dependency-only deviation is tested against the full accepted grid regression and explicit molecule import/checkpoint smoke tests.