Neutrino experiments infer the energy of an invisible incoming particle from what comes out of a heavy nucleus, such as carbon or argon. The nucleus distorts that picture in three ways. Its nucleons move and are bound, some interactions strike two correlated nucleons at once, and outgoing pions and protons rescatter or get absorbed on their way out. Event simulators such as GENIE expose these effects as a few dozen tunable parameters. The main failure is not poor fit quality but degeneracy: different parameter combinations reproduce the same published spectra, so the field cannot tell which correction is right. Breaking that degeneracy needs measurements in channels that individual collaborations have little reason to prioritize, such as multi-neutron or neutral-pion final states. Each collaboration chooses its next analysis to shrink its own oscillation systematics, and the global tuning effort only fits whatever gets published.
This proposal borrows one tool from the economics of learning, Bayesian incentive compatibility (BIC). A recommendation is BIC if, after hearing it, a self-interested agent who knows how recommendations are produced still prefers to follow it. Here the recommender is a coordinator holding the pooled posterior over generator parameters, and each experiment receives a recommendation of which measurement channel to pursue next. The proposal asks two linked questions. Does the archive of existing cross-section measurements already contain enough diversity to act as the "seed data" that makes posterior sampling BIC? And if so, does compliance-constrained planning break the nuclear-parameter degeneracy much faster than each experiment acting alone, at no cost to any experiment's own systematic budget? The answer is testable in simulation, and a negative answer would itself be informative.
Core Claim
Incentivized-exploration theory has a well-known unsolved practical step: posterior sampling becomes incentive-compatible only after a warm-up dataset with enough "spectral diversity" has been collected. Spectral diversity means the data constrain every direction in parameter space, measured by the smallest eigenvalue of the information matrix. Incentivized Exploration via Filtered Posterior Sampling proves the required warm-up is a horizon-independent constant of order $1/\epsilon^2$. It assumes this warm-up is bought or mandated. Neutrino physics inverts that situation, because the warm-up already exists as decades of published T2K, MINERvA and MicroBooNE data.
The nuclear-physics open problem is a statement about the same matrix. Parameter degeneracy means the smallest eigenvalue of the Fisher information in generator-parameter space is small along specific directions, notably the trade-off between short-range-correlation fraction and pion absorption or charge exchange identified by First combined tuning on transverse kinematic imbalance data with and without pion production constraints. The claim is that one quantity, the archive's information matrix projected onto those directions, decides two things at once. It decides whether channel recommendations drawn by posterior sampling are compliant for each experiment, and it decides how quickly the degenerate directions get resolved. The deliverable is a measurement-planning mechanism whose compliance can be checked numerically on GENIE's real parameter space.
Defining the game honestly
Physicists are not adversaries, and this proposal does not treat them as such. The "self-interest" is concrete and benign. Each experiment $i$ has a target nucleus, a flux, and an oscillation analysis whose dominant nuclear systematic is the bias in reconstructed neutrino energy. Its utility for measuring channel $a$ is
$$U_i(a) = V_i(\text{current}) - \mathbb{E}\left[V_i(\text{after measuring } a)\right] - c_i(a),$$
where $V_i$ is the posterior variance of experiment $i$'s energy-reconstruction bias, propagated through the generator, and $c_i(a)$ is a cost proxy in exposure and analysis effort for channel $a$. A channel recommendation $\sigma$ is BIC for experiment $i$ when $\mathbb{E}[U_i(a) - U_i(a') \mid \sigma = a] \ge 0$ for every alternative channel $a'$, the expectation taken over the parameter posterior.
The coordinator, modeled on a joint tuning group, has an information advantage because it holds the pooled cross-experiment likelihood and covariances. An individual experiment sees mainly its own data. That advantage is what pays for exploration, and no grants or mandates are involved.
The action set is a finite list of channels. These are $0\pi$ and $1\pi^+$ transverse-imbalance measurements, the $\pi^0$ channel, low-available-energy multi-neutron events as in the MINERvA multineutron measurement, argon inclusive, quasielastic-like and neutral-current $\pi^0$ channels as in MicroBooNE (inclusive, CCQE-like, NC $\pi^0$), and hadron-beam pion data like ProtoDUNE-SP. Rewards are correlated across arms through the shared generator parameters.
This structure shapes the choice of method. Smooth action sets allow fast bootstrapping, as shown in Geometry Meets Incentives, but that result does not transfer to a discrete menu of detector channels. Combinatorial menus can hit exponential barriers (Incentivizing Combinatorial Bandit Exploration, Sellke 2023). The finite-arm, correlated-prior structure is exactly what Filtered Posterior Sampling handles. Its "semantic map" is the rule that turns an internal posterior draw into the message the agent sees, and here it becomes "the channel whose predicted spectrum best discriminates the sampled parameter point." The method is therefore borrowed from that paper. What is new is the forward model, the utility, and the use of an existing archive as the seed.
Research Program
The first component is an emulator of the physics. We reuse the 14-parameter GENIE configuration of Li et al., built from the spectral-function nuclear model with correlated nucleons and the effective one-step pion-rescattering cascade, together with its Professor-style polynomial emulator. That emulator is extended with channels beyond the four transverse-imbalance datasets. For each channel we compute an expected-data Fisher matrix under realistic exposures and published covariances.
The second component is the seed audit. We compute the archive's information matrix, its smallest eigenvalue, and the eigenvector that achieves it, then translate these into the compliance precision $\epsilon$ that the Kalvit et al. guarantee would certify. Because their constants are conservative, we also check BIC directly by Monte Carlo: we sample recommendations, condition on them, and estimate each experiment's compliance gap.
The third component is the mechanism itself, Filtered Posterior Sampling over channels. When a recommendation fails the numerical compliance check for some experiment, we compare two fixes. One pools that experiment's recommendation with its own preferred channel, in the style of Mansour et al. and Cohen & Mansour. The other falls back to the experiment's solo choice, in the spirit of B2IR.
The fourth component is prior robustness. BIC theory assumes everyone shares one prior, but collaborations deliberately center their priors on different generator tunes. We give each simulated experiment a prior centered on its own default model and measure how compliance degrades. The Kalvit et al. paper flags this regime as open.
Baselines
The naive reference is uncoordinated greedy choice. Each experiment picks the channel maximizing its own $U_i$ under its own posterior, publishes, and the global fit updates. Every method must beat this on resolving the degenerate directions, and it represents current practice. It is run on the identical emulator, exposure budget and number of rounds.
The strongest existing tuning method is the combined Professor fit of Li et al.. On the four T2K/MINERvA transverse-imbalance datasets it lowered total $\chi^2$ from 461.25 (default G24-0) to 385.94 (G24-t), and lowered the MINERvA $\pi^0$ $\chi^2$ from roughly 70/12 to 36/12. The proposal does not try to beat that $\chi^2$, since it plans measurements rather than fitting existing ones. Its role is to supply the starting posterior and the end-of-program refit, with the same parameter set and emulator for every arm.
The strongest planning competitor is centralized Bayesian experimental design, which recommends the channel with the largest expected information gain on the global parameters and ignores compliance. No reported result was found on Lacuna for expected-information-gain channel selection in neutrino generator tuning, and no reported result for any BIC mechanism on this benchmark.
The ablation that isolates the key ingredient is Filtered Posterior Sampling with the pooled archive replaced by each experiment's own data. This removes the information asymmetry and tests whether pooled information, rather than posterior sampling itself, is what buys compliance.
Beating the baselines means two things at once. First, reaching a target posterior width along the archive's weakest eigen-direction in fewer simulated measurement rounds than greedy choice. Second, keeping the compliance gap non-negative for every experiment and holding each experiment's final $V_i$ at or below its greedy value. Centralized design should resolve degeneracy fastest. We count the mechanism a success if it comes within a small factor of that speed while centralized design violates individual compliance in a substantial fraction of rounds.
Why not just…?
Why not just run centralized Bayesian experimental design and ask collaborations to follow it? This is the real threat. If the information-gain-optimal channel is almost always also good for each experiment's own systematics, the incentive layer adds nothing. The experiment settles this directly by logging, for every round, whether the centrally optimal channel satisfies each experiment's compliance condition. If the violation rate is near zero, we will report that coordination problems in neutrino cross-section physics are informational rather than incentive problems. That conclusion is useful, and the work would then be a planning study. Cohen & Shao (Incentivized Collaboration in Active Learning) give a reason to expect violations: greedy collaborative query rules can raise one participant's burden from $O(1)$ to $\Omega(n)$. Their result is for noise-free thresholds, not for correlated generator parameters, so whether it carries over is an empirical question.
Why not just keep tuning globally with Professor? Global tuning fits only data that already exist, and Li et al. themselves report that nuclear-model and pion-rescattering parameters compensate for one another after the fit. A better minimizer does not add information along a direction the data do not constrain. The seed audit will quantify exactly which directions the archive leaves flat.
Why not just let each experiment optimize its own systematics? That is the greedy baseline. Its failure mode is visible in the file. Multi-neutron production was overpredicted by about 50% at the peak and its high muon-transverse-momentum fall-off was missed by every model (Olivier et al.). Binding-energy implementations fail to place the quasielastic peak, as seen in the transverse-imbalance data of Cai et al.. Neither channel need be any single experiment's best immediate investment. The simulation measures how much degeneracy greedy play leaves unresolved after a fixed budget.
Why not just fund the neglected channels directly? Payments are institutionally hard across national collaborations. More to the point, Cohen & Mansour show that information asymmetry alone can implement welfare-optimal compliant exploration in finite-arm settings. The proposal tests whether that free currency is large enough here. If the compliance gap cannot be closed without payments, we will report the required subsidy as a number.
Proposed Experiments
The decisive first check is the seed audit on real data. Using the Li et al. datasets plus the MicroBooNE and ProtoDUNE-SP measurements, we compute the archive's smallest-eigenvalue direction in the 14-parameter space. We then confirm that this direction aligns with the correlation–absorption trade-off Li et al. report, which validates that the information matrix captures the known degeneracy.
The second check is the main simulation over 20–50 sequential measurement rounds. Fake-data "truths" are drawn from parameter points consistent with current data but split along the degenerate direction, and from alternative-generator configurations where emulation allows. Every arm runs on the same emulator and exposure budget: greedy, centralized information gain, Filtered Posterior Sampling with pooling or fallback, and the no-asymmetry ablation. Metrics are rounds needed to halve the posterior width along the weakest direction, per-experiment compliance gap, the argon energy-reconstruction bias variance relevant to a DUNE-like experiment, and the final refit $\chi^2$ against held-out channels.
A held-out target test checks transfer from carbon to argon. Planning uses carbon-only information, and the result is scored on argon predictions, because the file's open-problem list identifies hydrocarbon-to-argon extrapolation as a known failure.
The final check is the mismatched-prior stress test. Each experiment's prior is centered on its own default generator tune, and we track how compliance degrades.
The proposal is falsified in its strong form if centralized design is nearly always compliant. It is also falsified if Filtered Posterior Sampling, under realistic prior mismatch, cannot keep every experiment's compliance gap non-negative while beating greedy play on the weakest direction.
References
- Incentivized Exploration via Filtered Posterior Sampling
- First combined tuning on transverse kinematic imbalance data with and without pion production constraints
- 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
- 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
- First Measurement of Differential Charged Current Quasielasticlike <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>-Argon Scattering Cross Sections with the MicroBooNE Detector
- 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
- Measurement of Exclusive $π^+$--argon Interactions Using ProtoDUNE-SP
- Geometry Meets Incentives: Sample-Efficient Incentivized Exploration with Linear Contexts
- Incentivizing Combinatorial Bandit Exploration
- Incentivizing Exploration with Linear Contexts and Combinatorial Actions
- Optimal Algorithm for Bayesian Incentive-Compatible Exploration
- Incentivized Collaboration in Active Learning
- Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions