Evidence-Fed Delphi Panels as Priors for Nutrient-Limited Decomposition Models

Soil models predict how fast microbes turn dead plant material into carbon dioxide. Much of that rate depends on nutrients: when litter is rich in carbon but poor in nitrogen or phosphorus, microbes cannot build new cells, and decomposition stalls. The models that encode this depend on quantities no one can measure directly in the field, such as maximum microbial growth rate, the energy microbes spend just staying alive, and how efficiently they turn consumed carbon into biomass. The result is equifinality: many different parameter settings reproduce the same measured CO₂ curve equally well. The model fits the data but tells us little about the mechanism. The most promising fix so far adds ecological rules ("maintenance must be slower than growth") that rule out implausible solutions. Those rules are currently chosen by the modelers themselves.

We propose to generate those rules, and quantitative parameter ranges, through a Delphi study. A Delphi study is an anonymous, multi-round expert survey in which panelists rate proposals, see summarized group statistics, and revise their answers. Our version changes the feedback step. Between rounds, panelists also see how their proposed constraints fare against real incubation data. The claim is that expert agreement and disagreement, measured this way, can be turned into constraints and prior distributions. We expect those priors to reduce equifinality and to predict held-out experiments better than hand-picked rules, pooled-data priors, or literature extraction. If they do not, the study will say so, and it will show where expert knowledge adds nothing beyond the data.

Core Claim and What Is New

The closest existing work motivates and constrains this proposal. Constrained hybrid modelling to predict microbial dynamics and organic matter turnover in soil systems (HySoMi) couples a neural network with a differentiable process model. The network maps genomic traits to biokinetic parameters, and the process model is a differentiable ordinary-differential-equation solver. HySoMi penalizes violations of hand-written ecological inequalities. On its synthetic benchmark this cut hidden-state mean squared error by roughly 66% and parameter L1 error by nearly 30% relative to an unconstrained hybrid. On field data from the European LUCAS survey, unconstrained fits violated the rules while constrained fits did not.

That result shows constraints matter. It does not show that the particular constraints chosen were the right ones, or that they transfer across soils. The research file found no study that uses a formal expert-consensus procedure to set such constraints for microbial nutrient limitation.

The name "Delphi" in Eidos, INDRA, & Delphi: From Free Text to Executable Causal Models refers to something else: a Bayesian software engine that compiles machine-read causal statements into executable graphs. It involves no human elicitation, and its authors report opacity in how parameters are initialized and in how uncertainty propagates. MPLasso injects prior microbial knowledge into network inference through penalties on a statistical estimator. That knowledge concerns which taxa associate, not kinetic bounds, and it is not elicited from people.

What we add has three parts. The first is a systematic, bias-controlled way to produce the constraint library and the parameter priors. The second is a rule that turns panel spread into prior width, so that items experts disagree on get wide priors instead of being dropped. The third is a feedback loop in which data-derived evidence enters the Delphi rounds, which addresses the method's known weakness that consensus is never checked against physical ground truth.

Research Program

The panel will have about 35 to 45 members. They will be drawn deliberately from two groups: process modelers, and empirical soil and wetland microbiologists. Membership is recorded, not revealed. Three rounds follow standard anonymity and controlled-feedback practice, as used in An international, Delphi consensus study to identify priorities for methodological research in behavioral trials in health research and the consensus meeting reported in Conservative Management and Biological Treatment Strategies: Proceedings of the International Consensus Meeting on Cartilage Repair of the Ankle.

Panelists will answer three kinds of items. The first kind is inequality constraints of the HySoMi type, which are rated for validity and scope, for example whether a rule holds in all soils or only aerobic ones. The second kind is quantile estimates for key parameters: the 5th, 50th, and 95th percentiles for maximum growth rate, maintenance rate, and the half-saturation constant for mineral nitrogen uptake. Panelists will also estimate the bounds on carbon use efficiency as a function of litter carbon-to-nitrogen ratio. Carbon use efficiency is the fraction of consumed carbon kept as biomass, and it is the quantity made flexible in Modeling the effects of litter stoichiometry and soil mineral N availability on soil organic matter formation using CENTURY-CUE (v1.0). The third kind is structural questions where the literature conflicts. The main one is the conditions under which added nitrogen speeds decomposition by relieving stoichiometric limits, as reported for methane-oxidizing root microbiomes in Methanotrophy Alleviates Nitrogen Constraint of Carbon Turnover by Rice Root-Associated Microbiomes, versus slowing it by suppressing lignin-degrading enzymes.

Items are retained using the content validity ratio. This is the item-filtering statistic used in Feasibility, comprehensibility and acceptability of the VISION-Cog, a novel tool to assess cognitive impairment in visually impaired older adults: a cross-sectional pilot study in Singapore, and it is paired with the scale-level content validity index used in Follow-up after major traumatic injury: a survey of services in Australian and New Zealand public hospitals. Cutoffs are fixed in a preregistration, and we report sensitivity to the 70% versus 80% agreement choice, because such cutoffs are otherwise arbitrary.

Panel output becomes model input in two ways. Retained inequalities become penalty terms. Parameter quantiles, pooled across panelists, become priors whose spread is set by the between-panelist interquartile range. The new element is evidence-fed feedback. Between rounds 1 and 2, and again between rounds 2 and 3, we fit the model to the calibration incubations. For each proposed constraint and range, panelists then see whether the unconstrained fit violates it and how much predictive loss it imposes on calibration data. Held-out data are never shown.

Baselines

The naive reference point is the unconstrained model: the same process model and hybrid network with no penalties and flat priors. It shows the equifinality problem at its worst. It will be fit with identical optimizers, data splits, and random-seed counts.

The strongest existing method is HySoMi with the constraint set Collart and colleagues hand-curated. It is the only published approach that attacks equifinality with ecological rules and reports gains: the 66% hidden-state MSE reduction and the roughly 30% parameter L1 reduction on its synthetic benchmark. We will reimplement it on our data with the same tuning budget, so that the only difference is where the constraints and priors come from. Beating it means lower held-out prediction error on unseen incubation experiments together with narrower but still well-calibrated parameter posteriors. A win on only one of these two counts as a partial result, not a success. For the Recous et al. (1995) and Guenet et al. (2010) incubations, CENTURY-CUE reports only qualitative improvement over default CENTURY. No reported numerical result was found on Lacuna for any method on those datasets, nor for LUCAS field error, so we will establish those numbers ourselves.

The data-driven competitor is hierarchical Bayesian calibration that pools the fourteen incubation experiments and learns the priors from the data itself.

Literature extraction is represented by parameter bounds pulled by a language model from published studies, in the spirit of the Eidos/INDRA pipeline. It gets the same parameter list the panel receives.

Two ablations isolate the key ingredients. The first uses a single-round elicitation with no iteration and no feedback. The second is a standard three-round Delphi with statistics-only feedback, which isolates the effect of evidence-fed feedback. A third comparison fits separate constraint sets from the modeler subpanel and the microbiologist subpanel, to measure how much panel composition shapes the outcome.

Why not just…?

Why not just keep the hand-picked constraints? They are cheap and already work on synthetic data. However, they were validated only against a simulator whose rules their authors also wrote, which is circular. The leave-one-experiment-out test on real incubations is where this is settled. If hand-picked rules match panel rules there, the Delphi step has not earned its cost for inequality constraints. It may still matter for parameter ranges and for the nitrogen question.

Why not just learn priors from pooled data with hierarchical Bayes? This is the most serious threat. With many comparable incubations, the data may pin down priors without expert input. The honest prediction is that pooling wins where calibration data resemble test data, as with corn and wheat residues across a range of nitrogen levels. The panel should win where they do not, including phosphorus co-limitation, which CENTURY-CUE explicitly omits, and extrapolation to field soils. We will report results separately by regime and accept a null result in the data-rich regime.

Why not just extract ranges from the literature with a language model? Published values mix incubation temperatures, textures, and definitions of carbon use efficiency. Extraction averages across them without judging validity, and the Eidos/INDRA authors report that automated assembly obscures how parameters are set and how uncertainty propagates. The extraction baseline tests this directly. If it matches the panel, the cheaper method should be preferred.

Why not an ordinary expert workshop? Open meetings let senior voices dominate, which is the problem anonymity was designed to solve. We cannot afford a parallel workshop arm. We can, however, measure conformity pressure indirectly by tracking how much round-3 answers move toward the round-2 median without new evidence being presented.

Proposed Experiments

The primary test is leave-one-experiment-out prediction on the five Recous and nine Guenet incubations. The metrics are root-mean-square error of cumulative respired litter carbon and of mineral nitrogen trajectories, plus the coverage of 90% predictive intervals.

Equifinality is measured as the posterior volume of the key parameters after calibration. We will also measure it as the fraction of posterior mass lying in regions the other constraint sets rule out.

On LUCAS we will fit HySoMi with each constraint set and report two things: violation rates of held-out rules, meaning rules not used in training, and prediction error on held-back samples. We will treat the 15-hour respiration windows as a known limitation.

The decisive check is split in two. The panel approach succeeds if it beats hand-curated HySoMi and pooled hierarchical Bayes on held-out incubations without losing interval coverage. Evidence-fed feedback succeeds if the evidence-fed arm beats the statistics-only Delphi arm.

For the nitrogen question, the panel's consensus decision rule will be encoded as a switch between the relief and suppression formulations. It will be compared against both fixed formulations on the incubations spanning mineral-nitrogen levels.