Neutrino experiments never see the neutrino. They reconstruct its energy from the debris it leaves after hitting a nucleus. To get from debris to energy they rely on event generators: large simulation codes such as GENIE, NuWro, NEUT and GiBUU that model how a neutrino scatters inside a carbon or argon nucleus. These generators disagree with data in scattered, recurring ways. They mis-model nucleon binding energy, underpredict events at intermediate momentum transfer, and fail badly on multi-neutron emission. Upcoming experiments such as DUNE expect generator uncertainties to become their largest systematic error. Nobody has asked a simple structural question about these failures. Do they concentrate where the nuclear physics is intrinsically hard, or where the community has published few measurements, so that generator tuning never had anything to chase?
We propose to answer that question quantitatively. First, we build a literature-derived coverage map of generator validation. We apply keyword co-occurrence analysis, the bibliometric practice of treating terms that appear together in papers as conceptually linked, to the measurement literature. The map records which combinations of nuclear effect, final-state topology, target nucleus and kinematic regime have actually been measured and compared. Second, we test whether that coverage, computed only from papers published before a cutoff year, predicts where generators disagree with measurements published after it. We also ask whether it adds anything beyond the obvious physical coordinate, momentum transfer relative to the Fermi momentum. That coordinate plays the role that elevation plays in terrain modeling: a single dominant predictor that everyone reaches for first.
Core Claim
The claim is falsifiable and has two parts. The first part concerns coverage. Once physical difficulty is controlled for, how sparsely a kinematic cell has been co-studied in prior literature should predict the size of later generator–data tension. If it does, generator failures partly reflect where the community looked, not only intrinsic difficulty. The second part concerns method. The prediction has to be judged on the extremes, the handful of cells where every generator fails, and not on average fit. Nonlinear regressors on a dominant covariate are known to compress exactly those extremes.
No paper in our search material connects these directions. The closest bibliometric work falls into two groups. Longitudinal co-word mapping, such as the Alzheimer's neuroimaging study by Yao et al., found bridging concepts through network centrality. Automated survey generation, such as SurveyG, organizes citation graphs so that a language model can write reviews. Both describe a literature. Neither checks its map against an external physical ground truth. Here the ground truth is the generator–data disagreement itself.
Research Program
The unit of analysis is a cell. A cell is a target nucleus (carbon/hydrocarbon or argon), a final-state topology (for example muon plus protons with no pions, single neutral pion, or multiple neutrons), a bin of three-momentum transfer $|\vec q|$ scaled by the Fermi momentum, and an observable family (lepton kinematics, transverse kinematic imbalance, or hadronic energy).
Building the coverage map. We extract author keywords, title terms and abstract noun phrases from the accelerator cross-section literature. We add the physics-model vocabulary that appears in generator comparisons. The co-word hypothesis (Mei Liu et al.) justifies reading joint appearance as joint study.
The known weakness of co-word analysis is terminology drift. Following the normalization practice in Yao et al., we hand-build a synonym table, since this field is full of synonyms: "2p2h", "MEC" and "multinucleon" often denote the same mechanism, and "CC0π", "mesonless" and "quasielastic-like" overlap. Each paper is then tagged with the cells it touches. The simplest coverage score for a cell is the recency-weighted count of prior papers that jointly mention all four of its elements. The richer version uses the full term network, including betweenness centrality, which measures how often a term lies on shortest paths between clusters. The network version tests a specific hypothesis: combinations that bridge otherwise separate clusters, such as neutron detection and cascade models, are the systematically neglected ones.
Building the ground truth. For each measurement with a public data release, we run GENIE (older and version 3 configurations), NuWro, NEUT and GiBUU through a single shared comparison harness. For every cell and generator we compute a tension score: the $\chi^2$ per degree of freedom of the prediction against the published covariance, restricted to that cell's bins.
The measurements named in our material give the anchor set:
- Lu et al., MINERvA transverse imbalance at 3 GeV, where GENIE and NuWro show a deficit at intermediate momentum transfer.
- Cai et al., binding energy and the unexplained proton left–right asymmetry $A_{LR}=-0.05\pm0.02$.
- Ascencio et al., low momentum transfer inclusive scattering.
- Bashyal et al., antineutrino quasielastic-like scattering.
- Olivier et al., multineutron production, where all models fail.
- MicroBooNE argon measurements of neutral-current $\pi^0$ (Abratenko et al.), inclusive electron neutrinos (Abratenko et al.) and protons with no pions (Abratenko et al.).
- T2K neutral-current $\pi^+$ (Abe et al.).
The model. We compare two nested predictors of tension $T_{c,g}$ for cell $c$ and generator $g$:
$$ T_{c,g} = f(z_c, a_c, g) \quad \text{versus} \quad T_{c,g} = f(z_c, a_c, g) + h(K_c^{<t}). $$
In these models, $z_c$ is the scaled momentum transfer (the "elevation" coordinate), $a_c$ is a detectability covariate (for example whether the topology contains neutrons or other neutral particles), and $K_c^{<t}$ is coverage computed only from papers before cutoff year $t$.
The terrain literature supplies two cautions for this design. Ishikawa's Mt. Fuji transects showed that a single elevation lapse rate misrepresents temperature once slope aspect interacts with elevation. By analogy, $z_c$ alone may miss target-by-topology interactions, so $f$ must be flexible. The AMSR-E downscaling study found that random forests using elevation as the top predictor underestimated the highest values and overestimated the lowest ones. Tension is heavy-tailed, so an averaged regressor would shrink exactly the failures we care about. We therefore fit gradient-boosted quantile models and evaluate on the upper tail.
Baselines
The naive reference ranks cells uniformly at random. It sets the chance level for top-decile recall of high-tension cells. We also report a second naive reference, "tension equals the cell's mean tension in the pre-cutoff data". It captures pure persistence: cells that were bad stay bad.
The physics-only baseline is the elevation analogue, $f(z_c, a_c, g)$ without coverage. It is the strongest honest competitor, because the physics literature already points to specific regions as hard. Lu et al. locate deficits above the Fermi momentum, and Olivier et al. trace failure to energy transfer into neutrons. It gets the same model family, the same hyperparameter search budget (a fixed number of cross-validated trials) and the same temporal split as the full model. Beating it means that adding coverage raises top-decile recall on post-cutoff cells by a margin outside the bootstrap interval over measurements. A calibrated gain on held-out tail cells also counts. A gain in mean squared error alone does not.
The count-only ablation replaces network coverage with a raw recency-weighted count of prior measurements per cell. This isolates whether co-occurrence structure matters beyond volume.
The language-model gap-finding baseline gives a retrieval-augmented model, and separately a SurveyG-style hierarchical-graph agent, the same pre-cutoff corpus. Each is asked to rank cells by expected generator failure. For survey writing, SurveyG reports citation recall of 91.4% against a human reference of 92.5% on computer-science topics (SurveyG). SurGE reports end-to-end citation recall below 10% for standard retrieval agents. For the actual task here, predicting generator–data tension, we found no reported result on Lacuna for any method.
Why not just…?
Why not just ask the neutrino physicists? Experts already know that neutrons and the region above the Fermi momentum are hard. That knowledge is what the physics-only baseline encodes. We will also collect a ranking from a small expert panel under the same pre-cutoff information. If experts match the coverage model on top-decile recall, the contribution shrinks to a reproducible, auditable version of expert intuition. We would report that plainly. The question the proposal adds is different: whether under-measurement, separately from difficulty, predicts failure. Experts cannot answer that from intuition, because the two are entangled in their experience.
Why not just count measurements per cell, with no co-occurrence network? This is a real threat, and we expect raw counts to carry most of the signal. The count-only ablation settles it. If network features add nothing, the main claim survives in its simplest form: coverage predicts failure. We would then drop the network machinery rather than defend it. The network earns its place only if bridge-type gaps predict tail tension that counts miss. Candidate bridge gaps include cells linking neutron observables with cascade-model terms, or the argon neutral-current pion channel where NEUT beat GENIE while GENIE v3 led on charged-current argon.
Why not just use the physical coordinate, as terrain modelers use elevation? If momentum transfer and detectability explain tension fully, then generators fail only where physics is hard, and coverage is irrelevant. The nested comparison answers this directly. One confound must be handled head-on: hard cells are also rarely measured, so difficulty and coverage correlate. We address it with the detectability covariate. We also examine cells of similar difficulty but different coverage, such as proton-only final states on carbon versus the same topology on argon, where there are fewer published comparisons.
Why not just let a language model read the literature and list the gaps? SurGE shows that fluent synthesis can mask poor grounding. SurveyG fixes attribution for surveys but has never been scored against a physical outcome. We run both as baselines under identical corpora, so their gap lists receive the same external test as ours.
Decisive Checks
The central test is retrodiction with a temporal split. We build coverage from papers before the multineutron and argon neutral-current pion results and ask whether those cells land in the predicted top decile of tension. Olivier et al. report failure across all models, which makes that cell an unambiguous positive. If the physics-only baseline already ranks it top and coverage adds nothing across several cutoff years, the coverage hypothesis is rejected.
A second check reverses the causal direction. Cells whose coverage rose after a cutoff should show falling tension in later generator versions. The GENIE v2-to-v3 transition, with the 2p2h additions motivated by Gran et al., is the natural test case.
A third check targets range compression. We compare top-decile recall of mean-regression versus quantile models. If the mean model systematically underranks the worst cells, the tail-focused evaluation is justified.
The deliverable is the coverage map itself: a ranked list of under-measured, predicted-high-tension cells for carbon and argon. Experiments preparing for DUNE could consult it when choosing which measurements to prioritize.
References
- Mapping longitudinal scientific progress, collaboration and impact of the Alzheimer’s disease neuroimaging initiative
- A Multi-Agent LLM Framework with Hierarchical Citation Graph for Automated Survey Generation
- Bibliometric analysis of the association between drinking water pollution and bladder cancer
- Measurement of Final-State Correlations in Neutrino Muon-Proton Mesonless Production on Hydrocarbon at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mo stretchy="false">⟨</mml:mo><mml:msub><mml:mrow><mml:mi>E</mml:mi></mml:mrow><mml:mrow><mml:mi>ν</mml:mi></mml:mrow></mml:msub><mml:mo stretchy="false">⟩</mml:mo><mml:mo>=</mml:mo><mml:mn>3</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi></mml:mrow></mml:math>
- Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions
- Measurement of inclusive charged-current <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mi>μ</mml:mi></mml:mrow></mml:msub></mml:mrow></mml:math> scattering on hydrocarbon at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mo stretchy="false">⟨</mml:mo><mml:msub><mml:mrow><mml:mi>E</mml:mi></mml:mrow><mml:mrow><mml:mi>ν</mml:mi></mml:mrow></mml:msub><mml:mo stretchy="false">⟩</mml:mo><mml:mo>∼</mml:mo><mml:mn>6</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi></mml:mrow></mml:math> with low three-momentum transfer
- High-statistics measurement of antineutrino quasielasticlike scattering at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>E</mml:mi></mml:mrow><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:msub></mml:mrow></mml:math> 6 GeV on a hydrocarbon target
- Measurement of the multineutron <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mi>μ</mml:mi></mml:msub></mml:math> charged current differential cross section at low available energy on hydrocarbon
- Measurement of neutral current single <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> production on argon with the MicroBooNE detector
- First measurement of inclusive electron-neutrino and antineutrino charged current differential cross sections in charged lepton energy on argon in MicroBooNE
- Measurement of differential cross sections for <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>ν</mml:mi><mml:mi>μ</mml:mi></mml:msub></mml:math>-Ar charged-current interactions with protons and no pions in the final state with the MicroBooNE detector
- First Differential Measurement of the Single <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:mrow> <mml:msup> <mml:mrow> <mml:mi>π</mml:mi> </mml:mrow> <mml:mrow> <mml:mo>+</mml:mo> </mml:mrow> </mml:msup> </mml:mrow> </mml:math> Production Cross Section in Neutrino Neutral-Current Scattering
- Soil Temperatures on Mt. Fuji during the Summer of 1974
- Downscaling of AMSR-E Soil Moisture over North China Using Random Forest Regression
- SurGE: A Benchmark and Evaluation Framework for Scientific Survey Generation
- First Measurement of Inclusive Muon Neutrino Charged Current Differential Cross Sections on Argon at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>E</mml:mi><mml:mi>ν</mml:mi></mml:msub><mml:mo>∼</mml:mo><mml:mn>0.8</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi></mml:math> with the MicroBooNE Detector
- Antineutrino Charged-Current Reactions on Hydrocarbon with Low Momentum Transfer