Policy-Coupled Coverage

• Policy-coupled coverage is a framework that defines adequacy relative to the behavior induced by a policy rather than relying on static data properties.
• It encapsulates three key couplings—distributional, decision-induced, and structural—each tailoring coverage to how policies influence trajectories, actions, and rule hierarchies.
• This approach addresses practical challenges such as data poisoning in offline RL, auditability in policy texts, and dynamic validity in counterfactual decision-making.

Policy-coupled coverage denotes a family of coverage notions in which adequacy is defined relative to a policy, decision rule, or induced behavior rather than as a policy-independent property of data, uncertainty, or text. Across recent work, this idea appears in several technically distinct forms: in offline reinforcement learning, coverage concerns whether logged data support the trajectories actually induced by a target policy; in counterfactual decision-making, it concerns whether prediction sets cover the realized outcome under the action chosen from those sets; in LLM policy design, it concerns whether a policy map makes covered and uncovered regions of model behavior visible and editable; and in uncertainty-aware control, it concerns whether a set of policies spans the uncertainty regimes that may arise at deployment (Zhou et al., 12 Jun 2025, Zheng et al., 2 Jul 2026, Lam et al., 2024, Ilboudo et al., 2024).

1. Conceptual structure of policy-coupled coverage

The cited works suggest three recurrent couplings. First, there is distributional coupling, where coverage is measured against the occupancy, sampling, or visitation distribution induced by a policy. Second, there is decision-induced coupling, where the coverage target depends on the action selected by the uncertainty object itself. Third, there is structural coupling, where coverage obligations are derived from the internal structure of a policy artifact, such as a rule hierarchy, a policy document, or a communication graph.

In the decision-theoretic formulation, policy-coupled coverage is explicitly self-referential. With action-indexed prediction sets $C=\{C(x,a)\}_{a\in\mathcal A}$, the induced max–min policy is

$\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$

and the central validity condition is coverage of the realized outcome under that induced action. In the preference-learning formulation, coverage is instead a relation between two policies through feature covariance,

$C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$

so that the informativeness of data collected by one policy is evaluated relative to the target neighborhood of another. In XACML testing, the coupling is structural: a rule is covered only through the full PolicySet–Policy–Rule chain encoded by the rule’s Target Set, not by the rule in isolation (Zheng et al., 2 Jul 2026, Kim et al., 13 Jan 2026, Lonetti et al., 2018).

A common consequence is that marginal or local notions of coverage are often insufficient. Coverage over isolated state-action pairs, isolated actions, or isolated policy clauses can fail to characterize what matters once the operative policy induces trajectories, selects actions, or activates hierarchical rules. This suggests that policy-coupled coverage is best understood as a shift from static support conditions to behavior-conditional support conditions.

2. Sequential decision learning and data support

In offline reinforcement learning, policy-coupled coverage enters through the mismatch between a fixed behavior-policy dataset and the occupancy distribution of a target policy. The sequence-level analysis in offline RL extends the classic concentrability coefficient

$$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$

to a sequence-level coefficient

$$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$

Because the trajectory ratio decomposes as

$$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$

the paper shows $C_\tau\le C^l$, and the discounted sequence-level $Q$-error bound scales accordingly. The central result is that insufficient sequence-level coverage can exponentially amplify the upper bound on estimation error with sequence length $l$. The same paper turns this into an attack surface via Collapsing Sequence-Level Data-Policy Coverage (CSDPC): state-action pairs are converted into clustered decision units, consecutive duplicate labels are removed to obtain compact decision patterns, rare patterns are identified by occurrence frequency, and small perturbations replace rare patterns with common ones under the stealth constraints $\|\zeta_t^s\|_\infty < \eta \cdot \|s_t\|_\infty$ and $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$0, with $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$1 typically set to $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$2. The reported effect is that poisoning just $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$3 of the dataset can degrade performance by up to $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$4 (Zhou et al., 12 Jun 2025).

The federated offline RL formulation weakens per-agent coverage requirements by replacing them with collaborative coverage. Instead of demanding that each local dataset cover the optimal policy’s support, it defines an average occupancy distribution $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$5 across agents and assumes an average single-policy clipped concentrability condition. Under this condition, FedLCB-Q achieves linear speedup in the number of agents, requires only $\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$6 communication rounds, and nearly matches the centralized-data sample complexity up to polynomial factors in the horizon length. The key point is that unsupported regions are judged collectively rather than agent-by-agent (Woo et al., 2024).

In online preference learning, coverage is dynamic rather than fixed. The coverage improvement principle states that, with sufficient batch size, each update moves into a region around the target where coverage is uniformly better, so subsequent data become increasingly informative. In the contextual bandit setting with Bradley–Terry preferences and a linear softmax policy class, the paper shows that online DPO converges exponentially in the number of iterations once batch size exceeds a generalized coverage threshold, whereas learners restricted to offline samples from the initial policy are limited by the frozen coverage bottleneck. A hybrid sampler based on a novel preferential G-optimal design removes dependence on coverage and guarantees convergence in two rounds (Kim et al., 13 Jan 2026).

A related verifier-coupled variant appears in RLVR data selection. IRDS defines a difficulty weight

$\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$7

and a trainability weight

$\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$8

both derived from verifier success counts, and then constructs a whitened difficulty-over-trainability metric on a Sparse Autoencoder cluster basis. Subset selection is posed as D-optimal design,

$\pi_\RA(x;C):=\arg\max_{a\in\mathcal A}\inf_{y\in C(x,a)}u(a,y),$9

so the selected data cover failure-relevant, trainable, and non-redundant directions. Experiments on three instruction-tuned models and six math reasoning benchmarks show the highest overall accuracy, exceeding the strongest baseline by $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$0 percentage points on the two Qwen models and by $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$1 percentage points on Llama-3.1-8B, while running an order of magnitude cheaper than the trajectory-based baseline (Li et al., 27 May 2026).

3. Counterfactual decisions, uncertainty sets, and calibrated validity

The counterfactual decision-making formulation gives the most explicit formal definition of policy-coupled coverage. The setup has features $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$2, finite actions $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$3, potential outcomes $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$4, and utility $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$5. A utility certificate $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$6 for a policy $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$7 satisfies

$C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$8

Direct risk-averse optimization then maximizes the expected certificate subject to that high-probability constraint.

Within this framework, three coverage notions are contrasted. Per-action marginal coverage requires $C_{\pi\shortrightarrow\pi'} := \inf\{C>0:\ \VV(\pi')\preceq C\,\VV(\pi)\},$9 for each action $$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$0. Universal policy coverage requires coverage under every policy $$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$1. Policy-coupled coverage (PCC) requires coverage only for the action induced by the set-valued decision rule itself, that is, for the realized outcome under $$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$2. Theorem 1 shows that PCC makes the max–min policy minimax-optimal over the induced ambiguity class, and also yields a valid utility certificate

$$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$3

This is the paper’s basis for calling prediction sets a “lossless interface” between uncertainty and action (Zheng et al., 2 Jul 2026).

The set-optimization problem under PCC, denoted RA-CPO-1, is shown to be equivalent both to a stronger universal-coverage problem, RA-CPO-2, and to direct risk-averse policy optimization, RA-DPO. The population-optimal sets are characterized through conditional utility quantiles. Defining

$$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$4

the optimal action-indexed prediction sets take the explicit form

$$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$5

A two-stage conformal method, Policy-Coupled Risk-Averse Conformal Prediction (PC-RACP), approximates these optimal sets by fitting nuisance quantile models, learning a policy through a split sample, and then applying weighted conformal calibration only on calibration points whose logged action matches the learned action. The finite-sample theorem states that the resulting sets satisfy policy-coupled coverage for the test point. Simulations and a real email-marketing experiment show higher utility than existing approaches while maintaining valid coverage, and the paper argues that ignoring the counterfactual structure is suboptimal for both validity and utility (Zheng et al., 2 Jul 2026).

A recurrent misconception addressed by this formulation is that uncertainty can be specified independently of the decision rule. The paper’s central claim is the opposite: in counterfactual settings, the outcome distribution that matters is the one induced by the chosen action, so the coverage target must be action-dependent in exactly that realized sense.

4. Policy texts, policy maps, and auditable assessment

In document-centered settings, policy-coupled coverage takes a structural rather than an occupancy-based form. Coverage concerns which parts of a policy document, taxonomy, or rule system are represented, exercised, or made governable. The same basic pattern appears in LLM policy authoring, privacy-policy analysis, XACML testing, and medical coverage review, although the concrete objects of coverage differ substantially (Lam et al., 2024, Lindner, 2020, Lonetti et al., 2018, Pokharel et al., 3 Jan 2026).

Domain	Policy object	Coverage unit
LLM policy design	Cases, concepts, if-then policies	Cases matching concepts or policy conditions
Privacy-policy analysis	OPP-115 data-practice taxonomy	Segment-level class assignments
XACML testing	Rule target/condition hierarchy	Covered request/response traces
Medical coverage review	Governing passages and symbolic rules	Retrieved subsection, facts, fired rule

AI Policy Projector formalizes a progression from Cases to Concepts to Policies. Cases are input-output pairs $$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$6; concepts are user-defined abstractions such as “violence” or “graphic details”; policies are if-then rules over those concepts. The system provides a map visualization and an authoring flow, and uses Classify(), Match(), Act(), and Generate() to support both concept definition and policy execution. In a study with 12 AI safety experts, participants authored 24 new policies using 43 concepts, including 31 self-defined concepts; all 12 participants authored unique custom concepts that no other participant identified, and 28 of 31 concepts were distinct ideas. The paper’s claim is not that full coverage is attainable, but that policy refinement should be grounded in a visible map of covered and uncovered behavior regions (Lam et al., 2024).

In privacy-policy coverage analysis, coverage means “the extent to which a privacy policy contains segments that correspond to the set of OPP-115 data-practice categories.” The task is segment-level multiclass classification over ten categories, such as Data Retention, Data Security, and Do Not Track. Models receive tokenized segments and output a softmax distribution over classes, with $$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$7. Using stratified 10-fold cross-validation, the strongest models achieve micro-F1 about $$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$8–$$C = \sup_{(s,a)\in \mathcal I_\mu}\frac{d^\pi(s,a)}{\mu(s,a)}$$9, logistic regression performs surprisingly close to the deep models, and classes with fewer samples, such as Data Retention, are harder to predict well (Lindner, 2020).

In XACML testing, policy-coupled coverage is defined over hierarchical rule obligations. The paper introduces four inclusive coverage criteria: Rule Target True, Rule Condition True, Rule Target False, and Rule Condition False. Its key abstraction is the Target Tuple

$$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$0

and a rule’s coverage obligation is coupled to the ordered Target Tuples of the enclosing PolicySet, Policy, and Rule. A monitoring infrastructure based on Glimpse performs on-line tracing by correlating request events and PDP responses with policy traces. In the reported evaluations, Multiple Combinatorial achieves $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$1 across all criteria, while Simple Combinatorial achieves only $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$2 on Rule Condition True for the sample policies because some conditions require multiple subject values (Lonetti et al., 2018).

In medical coverage review, policy-coupled coverage assessment is implemented as a neuro-symbolic pipeline. A coverage-aware retriever, trained on about 20 certified coding SMEs’ annotations from 172 CoCs/SPDs, uses LongformerForMultipleChoice to identify governing passages rather than merely topically similar ones. The dataset contains 1.84 million labeled $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$3 pairs. Retrieved passages are converted into attributes and then into PyKnow rules so that a reviewer sees the passage, the explicit facts, the triggered rule, and the matching conditions. The abstract reports a 44% reduction in inference cost alongside a 4.5% improvement in F1 score, and the evaluation reports average performance of Accuracy $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$4 and F1 $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$5 for the rule-based system with the finetuned retriever (Pokharel et al., 3 Jan 2026).

5. Multi-agent, domain-uncertainty, and geometric control formulations

A distinct line of work treats coverage as the set of policies needed to span uncertain environments, spatial neighborhoods, or geometric service regimes. In these formulations, a single universally averaged policy is often replaced by a family of policies whose adequacy depends on the current domain belief, energy state, neighborhood coupling, or geometric position (Ilboudo et al., 2024, Meng et al., 2018, Dai et al., 5 Dec 2025, Rastgoftar, 5 Jan 2026, Baccelli et al., 2013).

The multi-domain domain-randomization formulation introduces the Pseudo-Multi-Objective MDP (PMOMDP), where each domain $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$6 is treated as a separate objective with return vector

$$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$7

For linear scalarization by an uncertainty vector $$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$8, the relevant object is the Convex Coverage Set (CCS),

$$C_\tau=\sup_{\tau\in \mathbb T_\mu}\frac{d^\pi(\tau)}{\mu(\tau)}.$$9

The paper’s argument is that standard domain randomization learns one conservative compromise policy for a fixed prior, whereas CCS learning yields the set of policies that are optimal for different uncertainty distributions and is therefore better aligned with deployment-time system identification (Ilboudo et al., 2024).

In multi-agent coverage with energy depletion and repletion, coverage quality is coupled to recharging decisions through a hybrid system with Coverage, To-charging, and In-charging modes. A guard

$$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$0

forces an agent to leave coverage before it can no longer reach the charging station, and the station acts as a centralized scheduler under First-Request-First-Serve (FRFS) or Shortest-Distance-First (SDF). The recharge-threshold vector $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$1 is optimized by Infinitesimal Perturbation Analysis (IPA), and the paper’s reported conclusion is that full recharging is optimal, that is, $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$2 for all agents (Meng et al., 2018).

The distributed scalable coupled policy framework, DSCP, addresses networked multi-agent reinforcement learning with interdependent rewards and locally coupled policy parameters. Each agent’s policy depends on its own parameter and those of its $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$3-hop neighbors, while rewards depend on local neighborhoods. The paper derives a coupled policy gradient through a neighbors’ averaged $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$4-function, estimates it by a geometric 2-horizon sampling method, and tracks neighbors’ parameters via a push-sum protocol. The resulting joint policy is proven to converge to a first-order stationary point. The accompanying robot swarm path-planning experiments show that larger $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$5 improves coordination and final performance but slows convergence (Dai et al., 5 Dec 2025).

A more explicitly geometric decentralized coverage framework appears in finite-state multi-agent ground coverage. Agents are partitioned into anchors and followers; each follower interacts with exactly three in-neighbors, forming an enclosing triangular communication structure. Each follower solves a local finite MDP whose state space is induced by barycentric discretization inside its communication triangle, and the time-varying cost is defined by the deviation from the centroid of the target subset contained within that triangle. Under the assumption of Anyway Output Controllability (AOC), the paper establishes decentralized convergence to a desired configuration that optimally represents the environmental target (Rastgoftar, 5 Jan 2026).

In a technically distinct wireless-network setting, geometric policies directly shape coverage contours. A user is served either by one base station or by two cooperating base stations according to

$$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$6

where $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$7. The no-cooperation probability is $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$8, and the resulting full-cooperation region expands toward cell edges. This is not a policy-coupled coverage formulation in the decision-theoretic sense, but it is a clear instance in which the policy parameter $$\frac{d^\pi(\tau)}{\mu(\tau)}=\prod_{t=0}^{l-1}\frac{\pi(a_t\mid s_t)}{\mu(a_t\mid s_t)},$$9 directly governs the geometry of the coverage set (Baccelli et al., 2013).

6. Common themes, misconceptions, and limitations

Taken together, these works suggest that policy-coupled coverage is less a single formalism than a recurring design principle. The recurring principle is that the object called “coverage” should be evaluated against what a policy actually induces: a trajectory family, a chosen action, a behavior cluster, a rule chain, a domain posterior, or a communication neighborhood. This is why single-step occupancy, per-action marginal validity, fixed taxonomies, or single averaged controllers repeatedly appear as inadequate baselines rather than satisfactory endpoints (Zhou et al., 12 Jun 2025, Zheng et al., 2 Jul 2026, Lam et al., 2024, Ilboudo et al., 2024).

One common misconception is that local guarantees automatically lift to policy-level guarantees. The offline RL sequence analysis shows that per-step ratios can multiply into $C_\tau\le C^l$0; the counterfactual decision analysis shows that per-action coverage does not ensure coverage for the chosen action; and the on-policy preference analysis shows that coverage is not fixed but evolves with the sampling policy. A second misconception is that coverage is only a generalization issue. The CSDPC results show that insufficient sequence-level coverage is also a security vulnerability, and the medical coverage-review pipeline shows that coverage assessment can be an auditability and cost problem as much as a predictive one (Zhou et al., 12 Jun 2025, Kim et al., 13 Jan 2026, Pokharel et al., 3 Jan 2026).

A third recurring issue is that exhaustive coverage is often impossible. Policy Projector explicitly adopts the mapmaking analogy because the space of LLM input-output pairs is unbounded. The counterfactual conformal framework addresses this by optimizing the interface between uncertainty and action rather than enumerating all cases. Multi-domain CCS learning addresses it by learning a set of sufficient policies rather than a single universally best one. This suggests that policy-coupled coverage is often a problem of selecting the right abstraction layer—sequences, concepts, policies, utility certificates, or convex coverage sets—rather than a problem of brute-force enumeration (Lam et al., 2024, Zheng et al., 2 Jul 2026, Ilboudo et al., 2024).

The limitations are correspondingly domain-specific. The offline RL poisoning analysis assumes a fixed behavior-policy dataset and shared transition dynamics for behavior and target policies. The federated offline RL guarantee is derived in a finite-horizon episodic tabular MDP. PC-RACP relies on sample splitting, bounded utility, and a known or estimated behavior policy. Policy Projector reports only a short study with participants from a single organization and does not provide production-grade compliance monitoring. The medical policy-review system attributes 73.5% of its incorrect cases to missing the correct attribute and 26.5% to incomplete rule generation. Multi-agent energy-aware coverage assumes simplified depletion and charging models, while the finite-state decentralized coverage framework depends on AOC and a prescribed triangular decomposition (Woo et al., 2024, Zheng et al., 2 Jul 2026, Lam et al., 2024, Pokharel et al., 3 Jan 2026, Meng et al., 2018, Rastgoftar, 5 Jan 2026).

Across these literatures, policy-coupled coverage reframes adequacy from a static property of data or text into a behavior-conditional property of the full policy loop. That reframing is the main source of its theoretical force and its practical consequences.