FedStrategist: Adaptive Federated Coordination

• FedStrategist is a coordination framework that enables decentralized learning by aggregating local model updates or beliefs while preserving data privacy.
• It employs methods such as FedAvg, secure aggregation, and Bayesian soft-decision fusion to handle heterogeneous and regulated data across multiple agents.
• The framework adapts to nonstationary trust and data regimes through robust aggregation and adaptive control strategies that trade off accuracy, communication, and security.

FedStrategist denotes a family of federated and multi-agent coordination designs whose common purpose is to perform learning, inference, or decision support across distributed actors without centralizing raw data. In the supplied arXiv materials, the label is used or invoked in four closely related senses: a cross-agency federated AI design pattern for regulated government data, a federated inference system based on soft-decision fusion, a hypothetical tool for simulating Federal Open Market Committee deliberation, and a meta-learning framework for adaptive robust aggregation in federated learning (Verma et al., 2018, Kayaalp et al., 2023, Hou et al., 5 Dec 2025, Haque et al., 18 Jul 2025).

1. Scope and conceptual usage

The supplied literature uses FedStrategist as a design label for decentralized coordination under privacy, trust, heterogeneity, and security constraints. The unifying motif is that local actors retain their private observations or datasets while a coordinating layer aggregates either model updates, posterior beliefs, or agent votes.

Context	FedStrategist role	Defining elements
Cross-agency government AI	“FedStrategist”-style solution	Agencies $A_1\ldots A_K$, no raw data may leave an agency, FedAvg, secure aggregation, DP, auditing
Federated decision making	Federated inference system	local Bayesian soft-decisions, arithmetic averaging, geometric averaging, MAP decision
Monetary-policy simulation	hypothetical FedStrategist tool	CrewAI pipeline, five role-based agents, debate, voting, Chain-of-Draft
Robust federated learning	FedStrategist framework	LinUCB contextual bandit, $\{\text{FedAvg},\text{Median},\text{Krum}\}$, cost-aware reward, risk tolerance parameter

This range of usage suggests that FedStrategist is less a single algorithm than a recurrent architectural pattern: local autonomy is preserved, while a strategic coordination layer governs aggregation, fusion, or committee synthesis (Verma et al., 2018, Kayaalp et al., 2023, Hou et al., 5 Dec 2025, Haque et al., 18 Jul 2025).

2. Cross-agency federated learning under regulatory constraints

In the government-oriented formulation, the problem is a multi-agency environment in which agencies $A_1\ldots A_K$ each hold proprietary, regulated data $D_1\ldots D_K$, including examples such as personally identifiable information, critical infrastructure logs, or medical records. The stated constraint is that no raw data may leave an agency, with restrictions arising from FISMA, CUI/ICAP controls, inter-agency MOUs, and privacy statutes. The goal is to train a single global model $w^*$ that performs as if trained on $D=\cup_k D_k$, but without centralizing raw data (Verma et al., 2018).

The mathematical objective is defined by local empirical risks

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$

with $N=\sum_k n_k$, and a global objective

$$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$

A common federation algorithm is FedAvg. In round $t$, the server broadcasts $\{\text{FedAvg},\text{Median},\text{Krum}\}$0, each client initializes $\{\text{FedAvg},\text{Median},\text{Krum}\}$1, performs $\{\text{FedAvg},\text{Median},\text{Krum}\}$2 local epochs of SGD over mini-batches $\{\text{FedAvg},\text{Median},\text{Krum}\}$3,

$\{\text{FedAvg},\text{Median},\text{Krum}\}$4

and returns $\{\text{FedAvg},\text{Median},\text{Krum}\}$5. The server then aggregates

$\{\text{FedAvg},\text{Median},\text{Krum}\}$6

The design also allows an optional proximal term $\{\text{FedAvg},\text{Median},\text{Krum}\}$7 in each $\{\text{FedAvg},\text{Median},\text{Krum}\}$8 to stabilize heterogeneous data.

The system architecture assigns each agency a Data Server that holds raw $\{\text{FedAvg},\text{Median},\text{Krum}\}$9 under local policy and a Federated-Client that is signed and sandboxed. The client pulls the latest global model $A_1\ldots A_K$0 or an encrypted update, runs local training under resource and governance constraints, applies local privacy filters such as DP noise or encryption, and returns $A_1\ldots A_K$1 or $A_1\ldots A_K$2 to the orchestrator. Optional Trusted Execution Environments such as Intel SGX enclaves are included for secure computations. The Federation Orchestrator, located at Agency 1 or a neutral party, maintains a Model Store, Aggregator Service, Secure Aggregation Module, and Workflow Manager. All RPCs run over mutually authenticated TLS, with key exchange via agency PKI or a cross-domain certificate authority; gRPC+Protobuf is an optional communication substrate for high-performance streaming of model parameters.

Privacy and security are layered. Secure aggregation is described in the Bonawitz et al. style: each client splits $A_1\ldots A_K$3 into shares, exchanges shares peer-to-peer, and the server only recovers $A_1\ldots A_K$4, preventing the server from learning any individual $A_1\ldots A_K$5. A homomorphic-encryption option uses Paillier-style aggregation: $A_1\ldots A_K$6 Differential privacy clips each update to norm $A_1\ldots A_K$7,

$A_1\ldots A_K$8

then adds Gaussian noise

$A_1\ldots A_K$9

with a per-round $D_1\ldots D_K$0-DP guarantee via

$D_1\ldots D_K$1

Policies and auditing are enforced locally, while all messages are time-stamped, signed, and logged in an append-only audit store.

The experimental setup uses an MNIST-style digit-classification task with $D_1\ldots D_K$2 agencies and a CNN comprising Conv1 $D_1\ldots D_K$3 ReLU $D_1\ldots D_K$4 MaxPool, Conv2 $D_1\ldots D_K$5 ReLU $D_1\ldots D_K$6 MaxPool, FC1 $D_1\ldots D_K$7 ReLU, and FC2 10-way Softmax. Hyper-parameters are $D_1\ldots D_K$8 global rounds, $D_1\ldots D_K$9, batch size $w^*$0, learning rate $w^*$1 decayed $w^*$2 per 10 rounds, and clipping norm $w^*$3 with DP noise $w^*$4 for $w^*$5. Three partitioning scenarios are reported: a random split across $w^*$6 agencies, a class-partitioned setup in which each agency holds only one digit class, and a class-partitioned setup plus limited data exchange of 128 examples per class. The random split yields 98.8% accuracy for a centralized baseline and 98.3% for FedAvg by round 100, with loss nearly overlapping centralized training after 50 rounds. Under class partitioning, FedAvg alone fails and stalls at approximately 10%. Sharing 128 examples per class, described as 2% of each local set, restores convergence: FedAvg+exchange reaches 97.5% by round 50, and even 1% of data exchange yields 95%+ accuracy. Communication overhead is approximately 10 MB per round per client for a model of about 1 MB, totaling approximately 10 GB for 100 rounds over 10 clients. The reported lesson is explicit: class skew must be addressed, DP and encryption add 20–50% overhead, and synchronous FedAvg converges faster per round but requires orchestration whereas asynchronous variants tolerate stragglers (Verma et al., 2018).

3. Federated inference and fusion of soft decisions

A second FedStrategist formulation concerns federated decision making rather than gradient aggregation. There are $w^*$7 edge agents and a finite hypothesis set $w^*$8. Agent $w^*$9 begins time $D=\cup_k D_k$0 with belief $D=\cup_k D_k$1, observes $D=\cup_k D_k$2 drawn from its unknown marginal $D=\cup_k D_k$3, and forms an intermediate belief or soft-decision by a local Bayesian update: $D=\cup_k D_k$4 Agents never share raw data; they share only beliefs (Kayaalp et al., 2023).

The central processor applies one of two canonical pooling rules. Under arithmetic averaging,

$D=\cup_k D_k$5

Under geometric averaging,

$D=\cup_k D_k$6

After fusion, each agent sets $D=\cup_k D_k$7 and continues. This yields a coordinated non-Bayesian social-learning loop with centralized fusion and decentralized sensing.

The analysis centers on the log-belief ratio

$D=\cup_k D_k$8

or equivalently the centered and scaled wrong-hypothesis log-belief

$D=\cup_k D_k$9

Under mild conditions—global identifiability of $$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$0, finite variances, and i.i.d. observations over time—both AA and GA admit asymptotic normality characterizations. For AA, the instantaneous mixture-log-likelihood is

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$1

with $$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$2 and $$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$3, and

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$4

For GA, the weighted log-likelihood ratio is

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$5

with

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$6

and

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$7

For a MAP decision

$$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$8

an error occurs when some wrong $$F_k(w)=\frac{1}{n_k}\sum_{(x,y)\in D_k}\ell(w;x,y),$$9 satisfies $N=\sum_k n_k$0. Using the Gaussian approximation,

$N=\sum_k n_k$1

and by a union bound,

$N=\sum_k n_k$2

The interpretation supplied in the source is that $N=\sum_k n_k$3 controls the mean drift and $N=\sum_k n_k$4 the diffusion, thereby determining the exponential-in-$N=\sum_k n_k$5 decay of error.

The simulation study comprises three experiments: independent Gaussians with $N=\sum_k n_k$6, $N=\sum_k n_k$7, and 500 Monte Carlo trials at $N=\sum_k n_k$8; independent exponentials; and correlated Gaussians with covariance $N=\sum_k n_k$9. In all cases, the normalized statistic $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$0 closely matched the $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$1 density, Shapiro–Wilk tests yielded $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$2-values $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$3, GA consistently achieved lower empirical error than AA, and in correlated data $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$4 increased while $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$5 remained unchanged. The stated trade-off is also explicit: GA enjoys a larger drift $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$6 by Jensen’s inequality and is generally preferable when local models are trusted, but AA is more forgiving of zero or small beliefs and is safer when one or two agents may be adversarial or highly mis-calibrated (Kayaalp et al., 2023).

4. Adaptive and robust aggregation as a control problem

In the 2025 federated-learning formulation, FedStrategist is a meta-learning framework that treats robust aggregation as a real-time, cost-aware control problem. The stated motivation is that federated learning is exposed to two sources of brittleness: non-IID data heterogeneity and adaptive adversaries, including a “stealth” attacker that normalizes its malicious update so that a static rule never flags it. The claim “no single rule is universally optimal” is central, and the framework is designed to choose among static defenses dynamically (Haque et al., 18 Jul 2025).

The federated aggregation model writes the server update as

$$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$7

where $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$8. FedAvg uses $$\min_w \; F(w)=\sum_{k=1}^K \frac{n_k}{N}F_k(w)+\lambda R(w).$$9; coordinate-wise median sets each coordinate to the median of $t$0; and Krum, assuming up to $t$1 Byzantines, selects the update $t$2 with the smallest sum of squared distances to its $t$3 nearest neighbors. FedStrategist introduces a state $t$4 with $t$5,

$t$6

an action set $t$7, normalized computational costs $t$8, $t$9, and $\{\text{FedAvg},\text{Median},\text{Krum}\}$00, and a reward

$\{\text{FedAvg},\text{Median},\text{Krum}\}$01

where $\{\text{FedAvg},\text{Median},\text{Krum}\}$02 is evaluated on a small proxy validation set held by the server.

The controller is a LinUCB-style contextual bandit. For each action $\{\text{FedAvg},\text{Median},\text{Krum}\}$03, the algorithm maintains $\{\text{FedAvg},\text{Median},\text{Krum}\}$04, $\{\text{FedAvg},\text{Median},\text{Krum}\}$05, and $\{\text{FedAvg},\text{Median},\text{Krum}\}$06. At round $\{\text{FedAvg},\text{Median},\text{Krum}\}$07, it computes the upper-confidence score

$\{\text{FedAvg},\text{Median},\text{Krum}\}$08

selects

$\{\text{FedAvg},\text{Median},\text{Krum}\}$09

applies the chosen aggregator, evaluates the reward, and updates

$\{\text{FedAvg},\text{Median},\text{Krum}\}$10

A stated proposition on controllability says that, under standard LinUCB concentration bounds, the asymptotic fraction of plays of each action $\{\text{FedAvg},\text{Median},\text{Krum}\}$11 is a continuous, monotonically nonincreasing function of $\{\text{FedAvg},\text{Median},\text{Krum}\}$12 for actions with cost $\{\text{FedAvg},\text{Median},\text{Krum}\}$13. The single hyperparameter $\{\text{FedAvg},\text{Median},\text{Krum}\}$14 is therefore presented as a direct “risk dial.”

The evaluation uses CIFAR-10 with $\{\text{FedAvg},\text{Median},\text{Krum}\}$15 clients and Dirichlet partitioning $\{\text{FedAvg},\text{Median},\text{Krum}\}$16. Two attack types are studied: standard “loud” poisoning, which scales a benign update by $\{\text{FedAvg},\text{Median},\text{Krum}\}$17, and stealth poisoning, which matches the average $\{\text{FedAvg},\text{Median},\text{Krum}\}$18 norm of benign updates. In the standard attack under $\{\text{FedAvg},\text{Median},\text{Krum}\}$19, final accuracies are 20.17% for FedAvg, 27.38% for Median, 10.00% for Krum, and 28.73% for Adaptive. In the stealth attack under $\{\text{FedAvg},\text{Median},\text{Krum}\}$20, final accuracies are 39.43% for FedAvg, 33.25% for Median, 21.41% for Krum, and 38.82% for Adaptive. In the Krum-favorable environment at $\{\text{FedAvg},\text{Median},\text{Krum}\}$21, final accuracies are 43.29% for FedAvg, 44.39% for Median, 39.97% for Krum, and 44.31% for Adaptive. The reported interpretation is deliberately cautious: under stealth attack, FedAvg achieves the highest raw accuracy but the model is stealthily corrupted, whereas FedStrategist learns to shift toward robust rules and accepts a small test-accuracy loss to preserve integrity.

The framework also reports an explicit policy shift with increasing $\{\text{FedAvg},\text{Median},\text{Krum}\}$22. In the stealth setting with $\{\text{FedAvg},\text{Median},\text{Krum}\}$23, $\{\text{FedAvg},\text{Median},\text{Krum}\}$24 yields 92% FedAvg, 4% Median, and 4% Krum; $\{\text{FedAvg},\text{Median},\text{Krum}\}$25 yields 80%, 15%, and 5%; $\{\text{FedAvg},\text{Median},\text{Krum}\}$26 yields 60%, 30%, and 10%; and $\{\text{FedAvg},\text{Median},\text{Krum}\}$27 yields 35%, 50%, and 15%. This makes the cost-security trade-off operational rather than merely qualitative. The computational overhead is also specified: LinUCB updates cost $\{\text{FedAvg},\text{Median},\text{Krum}\}$28 per chosen action, and the instrumentation metrics cost $\{\text{FedAvg},\text{Median},\text{Krum}\}$29 per round (Haque et al., 18 Jul 2025).

5. Multi-agent policy simulation and the FedSight-derived variant

A further usage appears in the discussion of a hypothetical FedStrategist tool for monetary-policy forecasting. The underlying system, FedSight AI, is implemented in CrewAI as a pipeline of LLM agents using OpenAI GPT-4o to simulate an FOMC meeting. It defines five role-based agents: an Analyst, an Economist, and three Member agents designated Regional Pragmatist, Academic Balancer, and Central Policymaker. Structured indicators and standardized unstructured text flow to the Economist, then to the Member agents, followed by debate, voting, and a final statement (Hou et al., 5 Dec 2025).

The formalization uses meeting index $\{\text{FedAvg},\text{Median},\text{Krum}\}$30 and agent index $\{\text{FedAvg},\text{Median},\text{Krum}\}$31. Structured indicators are $\{\text{FedAvg},\text{Median},\text{Krum}\}$32, unstructured inputs are concatenated as $\{\text{FedAvg},\text{Median},\text{Krum}\}$33, and an agent’s initial state is

$\{\text{FedAvg},\text{Median},\text{Krum}\}$34

At debate round $\{\text{FedAvg},\text{Median},\text{Krum}\}$35,

$\{\text{FedAvg},\text{Median},\text{Krum}\}$36

followed by

$\{\text{FedAvg},\text{Median},\text{Krum}\}$37

The final vote is

$\{\text{FedAvg},\text{Median},\text{Krum}\}$38

where $\{\text{FedAvg},\text{Median},\text{Krum}\}$39 correspond to cut, hold, and hike. The committee decision is then taken by equal-weight majority vote: $\{\text{FedAvg},\text{Median},\text{Krum}\}$40 with an implied tie-breaking default to hold.

A distinctive extension is Chain-of-Draft. Traditional Chain-of-Thought is described as verbose and expensive in multi-agent debate, whereas CoD enforces brevity with at most 30 words per draft step, then applies pruning and revision. The step-length constraint is

$\{\text{FedAvg},\text{Median},\text{Krum}\}$41

and pruning is defined by

$\{\text{FedAvg},\text{Median},\text{Krum}\}$42

The final draft is $\{\text{FedAvg},\text{Median},\text{Krum}\}$43. The reported hyperparameters include $\{\text{FedAvg},\text{Median},\text{Krum}\}$44 simulations per meeting, three historical meetings stored as long-term memory for the ICL variant, and a prune threshold $\{\text{FedAvg},\text{Median},\text{Krum}\}$45, empirically chosen. CoD yields approximately 20% fewer tokens.

The evaluation covers 16 scheduled FOMC meetings from February 2023 to December 2024, consisting of 9 holds, 3 cuts, and 4 hikes. Structured predictors include 15–20 macro indicators; unstructured sources include Beige Book text, dot-plot verbalizations, and FedWatch probabilities. The performance metrics are Total Accuracy, Agent Accuracy, Voting Stability, Semantic Similarity, Average Tokens, and Mean Absolute Error in basis-point magnitude. The main quantitative results report 87.50% Total Accuracy, 78.13% Agent Accuracy, 86.67% Voting Stability, 74.58% Semantic Similarity, 75,724 Average Tokens, and 0.0313 bp MAE for FedSight AI; 87.50%, 80.63%, 88.54%, 72.72%, 81,303, and 0.0313 bp for FedSight ICL; and 93.75%, 90.22%, 93.33%, 73.82%, 60,464, and 0.0156 bp for FedSight CoD. The details also report separate external-baseline comparisons in which MiniFed attains 75% accuracy with MAE $\{\text{FedAvg},\text{Median},\text{Krum}\}$46 bp, Ordinal Random Forest attains 62.5% directional accuracy, Linear Regression attains 31.25%, and FedSight CoD reaches 100% accuracy with 0 bp MAE in those comparisons. A rare misprediction is described: one 2024 cut was predicted as hold because late-arriving Beige Book details on regional labor softness were underweighted in the early draft. The blueprint is explicitly presented as informative for the design of a hypothetical FedStrategist tool, indicating a broader usage in which strategic orchestration concerns committee reasoning rather than federated gradient aggregation (Hou et al., 5 Dec 2025).

6. Recurring principles, limitations, and deployment guidance

Several principles recur across the FedStrategist variants. First, raw-data non-disclosure is fundamental. In the cross-agency setting, no raw data may leave an agency; in the fusion setting, only beliefs are transmitted; in the robust-FL setting, only model updates are observed; and in the monetary-policy variant, the system aggregates structured and unstructured evidence through agent states, messages, and votes rather than through centralized retraining (Verma et al., 2018, Kayaalp et al., 2023, Hou et al., 5 Dec 2025, Haque et al., 18 Jul 2025).

Second, heterogeneity is not treated as a peripheral nuisance but as the core design challenge. Class partitioning in digit classification causes FedAvg to stall at approximately 10%, and limited data exchange of 128 examples per class or even 1% of local data restores convergence. In the robust-FL setting, the optimal aggregator depends on whether the environment is highly non-IID, stealthy, or Krum-favorable; the framework’s premise is that no single static rule is universally optimal. In the fusion-theoretic setting, GA provides faster drift but AA is safer under mis-calibration or adversarial conditions. This suggests that FedStrategist, in its different instantiations, is fundamentally about adaptive coordination under nonstationary trust and data regimes rather than about a single preferred update rule (Verma et al., 2018, Kayaalp et al., 2023, Haque et al., 18 Jul 2025).

Third, raw predictive accuracy is not always an adequate operational criterion. The robust-FL study explicitly describes the paradox of the “Losing Winner,” in which FedAvg has the highest apparent accuracy under stealth attack while carrying a hidden backdoor. The government FL study likewise shows that a superficially standard FedAvg pipeline can fail under class skew. The policy-simulation variant identifies data latency and underweighted qualitative cues as a source of misprediction. These cases undermine a common misconception that decentralized systems can be evaluated solely by top-line accuracy without regard to integrity, skew, or provenance (Verma et al., 2018, Hou et al., 5 Dec 2025, Haque et al., 18 Jul 2025).

The deployment guidance is concrete. For cross-agency FL, the recommendations are to formalize inter-agency trust through MOUs, a joint governance board, and shared PKI; begin with a pilot of 3–5 agencies on a well-understood dataset; use secure enclaves or homomorphic aggregation; calibrate DP parameters to privacy budgets $\{\text{FedAvg},\text{Median},\text{Krum}\}$47; monitor convergence per agency and implement skew alerts; log provenance with agency ID, timestamp, and round number; compress model deltas through quantization or sparsification to cut bandwidth by $\{\text{FedAvg},\text{Median},\text{Krum}\}$48; provide a rollback mechanism for suspicious gradients; and integrate with TensorFlow-Federated or PySyft wrapped by the agency’s policy engine. For adaptive robust aggregation, deployment consists of adding instrumentation for $\{\text{FedAvg},\text{Median},\text{Krum}\}$49, cosine similarity, and mean-update norm; maintaining LinUCB bookkeeping; and choosing $\{\text{FedAvg},\text{Median},\text{Krum}\}$50 according to acceptable compute-risk trade-offs. For federated inference, one can estimate $\{\text{FedAvg},\text{Median},\text{Krum}\}$51 through a short pilot run and select the fusion rule that maximizes $\{\text{FedAvg},\text{Median},\text{Krum}\}$52 (Verma et al., 2018, Kayaalp et al., 2023, Haque et al., 18 Jul 2025).

The principal limitations are likewise explicit. The policy-simulation system is evaluated on only 16 meetings and depends on LLM behavior, data cutoff, and possible status-quo anchoring. The robust-FL framework uses a compact $\{\text{FedAvg},\text{Median},\text{Krum}\}$53 state, which is efficient but necessarily selective. The fusion-theoretic analysis assumes global identifiability, finite variances, and i.i.d. observations over time. The government FL pattern incurs communication cost, orchestration requirements, and 20–50% overhead from DP and encryption. Taken together, these limitations indicate that FedStrategist is best understood as an architectural family for privacy-preserving strategic coordination, with each instantiation making different trade-offs among robustness, communication, interpretability, and institutional compatibility (Verma et al., 2018, Kayaalp et al., 2023, Hou et al., 5 Dec 2025, Haque et al., 18 Jul 2025).