When several institutions jointly train a model without moving their raw data, and the resulting model is sold to outside buyers, someone has to decide how much each institution earns. The institutions here are hospitals, agencies or firms, and their separate data stores are what this note calls "silos." Existing incentive schemes make that decision under a quiet simplification: revenue is assumed to grow linearly, or at least smoothly, with model accuracy. Real buyers do not behave that way. A hospital pays for a skin-lesion classifier only once it clears its own accuracy bar, and data-market platforms sell tiered menus whose price jumps at quality thresholds. Under such stepped demand, a silo's contribution stops being a matter of "how many points of accuracy it added." It becomes "did it push the model over someone's threshold." Cheap payment rules such as proportional-to-data or leave-one-out then pay the wrong institutions. The exact fair rule, the Shapley value (each silo's average marginal contribution over all orders in which silos could join), needs a model retrained for every subset of silos.
We propose borrowing the structure of a journal editorial board to fix this. Silos are grouped under "associate editors," one per acquisition domain such as a scanner vendor, a hospital network or a publishing agency. Each associate editor certifies and scores its own small group using a local accuracy measure, where a linear approximation of revenue is reasonable. An "editor-in-chief" tier prices only the few group-level combinations against the true, non-linear buyer demand. The claim is that the hierarchy is a principled answer to a pricing question: it confines the expensive, threshold-sensitive valuation to the scale where demand is genuinely non-linear, and it applies linear approximation only where changes are small. As a side effect, it limits how much a silo can gain by cloning its data.
Core Claim
Let $v(S)$ be the revenue that the model trained on silo set $S$ earns from a buyer population under a given pricing menu. Silos are partitioned into $g$ domain groups $G_1,\dots,G_g$. The editorial allocation pays silo $i$ in group $G_k$
$$ \phi_i \;=\; \mathrm{Sh}_k\big(v_G\big)\cdot \frac{\psi_i^{(k)}}{\sum_{j\in G_k}\psi_j^{(k)}}, $$
where $v_G(H)=v(\cup_{h\in H}G_h)$ is the revenue game played among groups. $\mathrm{Sh}_k$ is group $k$'s Shapley share of that game, computed on true non-linear revenue. $\psi_i^{(k)}$ is silo $i$'s Shapley value inside its group, computed on accuracy measured on the group's own validation data. Retraining cost falls from $2^n$ subsets of $n$ silos to $2^g+\sum_k 2^{|G_k|}$, and the within-group terms use cheap local evaluations.
The note makes three claims, each of which can be falsified. First, the error of linear-demand payment rules relative to exact Shapley on true revenue is small when the pricing menu is linear and grows with the number of price steps. If it does not grow, the linear assumption is harmless and this proposal is unnecessary. Second, when groups follow real acquisition domains, editorial allocation stays close to exact Shapley. When groups are random, it does not. This means the board's composition, not hierarchy as such, carries the benefit. Third, under editorial allocation a silo that clones its data can only take revenue from members of its own group, never from other groups. Small groups also make near-duplicate checks by the associate editor tractable.
How This Relates to Prior Work
The motivating gap comes from two papers that each hold half of the problem. Optimal Pricing for Data-Augmented AutoML Marketplaces prices downstream model gains to buyers with a non-linear menu, built by a mixed-integer program, that segments buyers by willingness to pay. It explicitly leaves open how revenue should flow back to the data suppliers. We take its stepped menu as our demand model and supply the missing supplier-side rule. Incentivizing Data Contribution in Cross-Silo Federated Learning compares egalitarian, linearly proportional, leave-one-out and Shapley allocation across silos. Its profit function is synthetic rather than derived from buyer demand, it flags data cloning as unresolved, and it uses a flat single-server design. We replace its internal profit with market revenue and test its four rules as baselines.
Two hierarchical schemes constrain the design. Design of Two-Level Incentive Mechanisms for Hierarchical Federated Learning uses a device-edge-cloud hierarchy in which the cloud pays a linear price per unit of performance. Its intermediate tier only aggregates updates and manages bandwidth; it does not evaluate them. Hierarchical Incentive Mechanism Design for Federated Machine Learning in Mobile Networks combines contracts at the lower tier with coalition formation at the upper tier under a fixed concave revenue curve. It finds that the grand coalition is often not stable. That finding is a warning for us, so we measure whether any group would rather leave under our rule. Our upper tier differs from both because it is priced against an externally learned, stepped demand curve. Our middle tier differs because it certifies and scores contributions rather than merely aggregating them.
Equilibrium Pricing in Oligopolistic Data Markets shows that uniform linear pricing of non-rival data has no 1.364-approximate equilibrium, while piecewise-linear convex pricing restores a 2-approximate one. This is independent evidence that linear prices fail for data. That paper assumes data can be exported as shards, whereas our silos never release raw data. The conceptual justification for the two-tier split comes from Optimal Pricing for Distance-Based Transit Fares. It shows that point elasticity, a local linearization, is valid only for small price changes, while large changes require arc elasticity. Our hierarchy applies the same distinction to contributions. The editorial analogy is grounded in descriptions of journals in which central editors set policy and associate editors filter within their area (IEEE Transactions on Cybernetics publication information). Embodied LLM Agents Learn to Cooperate in Organized Teams finds that designated-leader structures keep communication growing linearly with team size. We use it only as evidence that fixed hierarchies avoid coordination blow-up. We do not use LLM agents.
Research Program
The first piece is a demand layer. We simulate a buyer population whose willingness to pay is a function of model quality, then sweep its shape. The sweep runs from a linear curve, through a smooth curve in the semi-parametric class of Semi-Parametric Contextual Pricing with General Smoothness, to stepped menus with 2, 4 and 8 tiers built by a mixed-integer program in the style of the AutoML-marketplace paper. The number of steps is our single knob for "distance from linearity."
The second piece is the valuation pipeline. For small instances we compute exact Shapley on true revenue by retraining every subset; this gives ground truth. For larger instances we compute editorial allocation, flat Monte Carlo Shapley and the cheap rules, each under a fixed retraining budget.
The third piece is the associate-editor tier. Each group node holds a small validation set from its own domain. It computes within-group accuracy-Shapley and runs a near-duplicate check on submitted updates, for example by comparing the similarity of model updates. The executive tier fixes both rules in advance so that an associate editor cannot favor its own members. This addresses the principal-agent risk that the governance literature in the file repeatedly raises.
Analytically, we will bound the gap between editorial allocation and exact Shapley. The bound is in terms of how far each group's local accuracy game departs from a linear restriction of revenue. If true, this formalizes "linearize only where changes are small."
Baselines
The naive reference point is egalitarian splitting together with linearly proportional allocation by sample count, as analyzed in Incentivizing Data Contribution in Cross-Silo Federated Learning. These rules need no retraining, and many deployments would default to them. The linear-demand baseline computes exact or Monte Carlo Shapley on accuracy, then scales by a single price. This corresponds to the $R\cdot$performance reward of Design of Two-Level Incentive Mechanisms for Hierarchical Federated Learning. It is the strongest competitor on the specific question of whether non-linear demand matters, because it does everything right except model the demand.
The strongest existing method overall is flat Monte Carlo Shapley on true menu revenue. All flat and hierarchical methods receive the same retraining budget, matched by count of model trainings and wall-clock time, and use the same federated optimizer, architecture and validation data. Leave-one-out on true revenue is also included because it is cheap and threshold-aware.
Two ablations isolate the key ingredient. Random-group editorial allocation keeps the hierarchy but discards the domain alignment. Editorial allocation with true revenue used inside groups as well tests whether local linearization is what saves compute without hurting accuracy.
For a published reference number, the closest is from Optimal Pricing for Data-Augmented AutoML Marketplaces. Its menu captures about 85% of total possible welfare in-sample, and its data-discovery component reaches average classification utility above 0.85 against about 0.70 for the next-best benchmark. That number is our demand-side sanity check, not an allocation baseline. For allocation quality against exact Shapley, there is no reported result on Lacuna for cross-silo revenue-sharing error under any demand model. On Fed-ISIC2019, FLamby reports about 65% balanced accuracy for pooled training, 50–55% for FedAvg, FedProx, Scaffold and FedOpt, and as low as 35% for local-only training. These numbers anchor what the jointly trained model is worth.
Editorial allocation beats flat Monte Carlo Shapley if, at an equal budget with at least 24 silos, it achieves lower mean absolute allocation error against exact or high-budget Shapley and a lower clone gain. It beats the linear-demand baseline if its error advantage grows with the number of menu steps.
Why not just…?
Why not just compute exact Shapley over silos? With six FLamby centers, exact Shapley needs only 64 retrainings, which is affordable. This is a real threat in small consortia, and we concede that regime. The proposal targets consortia where centers contain sub-units: departments, devices, or individual agency datasets in NYC Open Data. There $n$ reaches the dozens, $2^n$ is out of reach, and Monte Carlo variance under step-function revenue is high because most permutations show zero marginal gain. We report the crossover $n$ at which editorial allocation starts to win at equal budget. If no crossover appears below 50 silos, the computational argument fails, and only the clone-confinement result remains.
Why not just keep linear pricing for allocation, since it is simpler? If payments track accuracy share alone, a silo that is pivotal for crossing a buyer's threshold is paid like any other contributor of a few accuracy points. That undercuts the incentive to supply exactly the data buyers pay for. Equilibrium Pricing in Oligopolistic Data Markets already shows that linear prices destabilize data markets on the seller side. Our first decisive check measures directly whether the allocation error of linear rules grows with menu steps. If it stays flat, we will report that the simpler approach suffices.
Why not just detect clones with deduplication and keep a flat structure? Deduplication is complementary, not a substitute. Perturbed near-duplicates can evade global checks, and comparing every pair of updates across all silos scales quadratically. The hierarchy gives a structural guarantee even when detection fails, because a clone can only dilute its own group's fixed share. It also makes detection cheap by limiting comparisons to small groups. We test flat Shapley with global deduplication against editorial allocation, using clones perturbed with increasing noise.
Why not just use gradient-based contribution scores that avoid retraining? Such scores measure local, first-order influence on accuracy. That is precisely the linear approximation which fails at buyer thresholds. We include an influence-style score as a within-group option. We do not expect it to work at the executive tier, and we will test that expectation.
Proposed Experiments and Decisive Checks
The horizontal testbed is FLamby Fed-ISIC2019. Its six acquisition systems serve as groups, and each system is split into 4–8 sub-silos by patient ID, giving 24–48 silos. Fed-TCGA-BRCA and Fed-Heart-Disease are cheaper controls where federation reaches pooled parity. The vertical testbed uses classification tasks drawn from the NYC Open Data setting of the AutoML-marketplace paper. There, augmenting datasets are suppliers and publishing agencies are associate editors. Metrics are mean absolute allocation error against exact Shapley, computed on instances with up to 12 silos and on high-budget Monte Carlo Shapley beyond that. We also report the number of retrainings; the clone gain ratio, meaning the cloner's payoff divided by its honest payoff; the fraction of revenue that leaks to a cloner from other groups; and group-exit incentive, meaning whether any group earns more alone.
Three results decide the proposal. First, if the error of linear-demand rules does not rise from the 1-step to the 8-step menu, the premise fails. Second, if domain-aligned groups do not beat random groups by a clear margin, the editorial-board structure is decoration rather than mechanism. Third, if editorial allocation shows clone gain near 1 and zero cross-group leakage while flat Shapley shows gain above 1, the governance tier earns its place even where flat Shapley is affordable. We will also check, on Fed-ISIC2019, whether between-group revenue remains well defined when some group combinations diverge, as FLamby reports for other tasks. Divergence produces zero-revenue coalitions that stress every allocation rule equally, and we will report it rather than hide it.
References
- Optimal Pricing for Data-Augmented AutoML Marketplaces
- Incentivizing Data Contribution in Cross-Silo Federated Learning
- Design of Two-Level Incentive Mechanisms for Hierarchical Federated Learning
- Hierarchical Incentive Mechanism Design for Federated Machine Learning in Mobile Networks
- Equilibrium Pricing in Oligopolistic Data Markets
- Optimal Pricing for Distance-Based Transit Fares
- IEEE Transactions on Cybernetics publication information
- Embodied LLM Agents Learn to Cooperate in Organized Teams
- Semi-Parametric Contextual Pricing with General Smoothness
- FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings