Institutional Analysis and Development Framework

• Institutional Analysis and Development (IAD) Framework is a formal system that defines how operational rules shape collective action, organization, and governance.
• It categorizes seven key rule types—from boundary to transformation—that structure decision-making processes and institutional outcomes.
• The framework is implemented both computationally (using agent-based and game-theoretic models) and qualitatively for comparative policy analysis.

The Institutional Analysis and Development (IAD) Framework is a formal analytic system originating in the work of Elinor Ostrom and colleagues, designed to explicate and compare the structural determinants of human organization, collective action, and governance across natural, social, and increasingly, socio-technical systems. The framework distinguishes a set of universal variables—most centrally, operational rules in use—that govern the structure, conduct, and outcomes of all repeated interaction scenarios or "action situations." In contemporary research, the IAD framework is instantiated both qualitatively (e.g., context mapping for AI risk and oversight (Morgan et al., 2023)) and quantitatively (e.g., computational game-theoretic instantiation (Montes, 2021)), with tool-supported encoding of institutional statements (IG Parser (Frantz, 19 May 2025)), enabling systematic cross-domain policy analysis, comparative institutional diagnostics, and formal what-if evaluations of rule changes.

1. Seven "Rules in Use": Formal Typology and Instantiation

Ostrom's IAD framework delineates precisely seven types of operational rules, each of which structures participant behavior and institutional outcomes in an action situation:

1. Boundary Rules ($R_{b}$): Specify eligibility for entry or exit from decision positions. Formally, letting $P$ be positions and $\mathrm{Info}$ information sets (licenses, credentials), the set of eligible tuples is $$R_{b} = \{ (p, \mathrm{info}) \mid \mathcal{E}(p, \mathrm{info}) = \mathrm{true} \}$$.
2. Position Rules ($R_{\pi}$): Map positions to available action sets, i.e., $R_{\pi}: P \rightarrow 2^{A}$ for action set $A$.
3. Choice Rules ($R_{\chi}$): Partition the action set of a position according to permitted, obligatory, or forbidden status: $$R_{\chi}(p) = \{ a \in R_{\pi}(p) \mid \mathrm{status}(a) \in \{\mathrm{allowed}, \mathrm{required}, \mathrm{forbidden}\} \}$$.
4. Aggregation Rules ($R_{\alpha}$): Govern how individual choices are combined to produce collective outcomes, $R_{\alpha}: A^{n} \rightarrow O$, with outcomes $O$.
5. Information Rules ($R_{\iota}$): Define informational partitions—what each position can observe, $I(p) \subseteq \mathrm{Data} \cup (A \times P)$.
6. Payoff Rules ($R_{\rho}$): Assign distributional consequences (rewards, sanctions), $R_{\rho}: A^{n} \times O \rightarrow \Pi$.
7. Transformation Rules ($R_{\tau}$): Govern state transitions, $R_{\tau}: S \times A^{n} \rightarrow S'$, with $S$ the set of states.

Each rule can be instantiated and measured as a "contextual variable" using real-valued functions reflecting, for instance, the number of required credentials ($R_{b}(C)$), the proportion of forbidden acts ($R_{\chi}(C)$), or the observed adaptivity of system response ($R_{\tau}(C)$) (Morgan et al., 2023).

Rule	Definition	Clinical Example (Morgan et al., 2023)	Contextual Metric
Boundary	Entry/exit conditions	GMC license, Royal College exams	# of mandatory credentials
Position	Action-set assignment to roles	FY1 order labs, consultant sign-off	$|R_\pi(p)|$
Choice	Action permission/obligation/forbiddance	High-risk drug: consultant sign-off	% forbidden acts in log
Aggregation	Combining inputs to produce outcome	Consultant override in MDT	% overrides vs. joint decisions
Information	Control over data/explanation visibility	Who sees AI explanations	# datapoints/explanation features
Payoff	Outcome-linked rewards/sanctions	GMC investigation/praise	Probability of sanction
Transform.	Rules for state evolution	Model retraining, care plan change	# model updates per quarter

2. Action Arena Formalism and Rule Encoding

An IAD "action arena" is formalized by the tuple $$\mathcal{A} = (A_{\mathrm{context}}, P, \{R_{b}, R_\pi, ..., R_\tau\}, S)$$, where the context, set of actors/positions, operationalized rules, and current institutional state are enumerated.

In high-stakes AI oversight (clinical setting), actors $P$ are, for example, junior doctors and consultants. Inputs $X$ (e.g., lab results, images) receive AI recommendations $y = M(X)$ with explanation $E(y)$. The final team decision is $d = h(y, E(y), X; \theta_{\mathrm{team}})$, where $h$ encodes the "team-in-the-loop" logic parameterized by collective expertise $\theta_{\mathrm{team}}$. The process cascades through nested positions (e.g., junior proposes $d_j$, consultant confirms via $d_c = h_c(d_j, y, E(y))$), and transformation rules update the system state $(X, d_c) \mapsto X'$ (Morgan et al., 2023).

3. Computational Realization: Agent-Based and Game-Theoretic Models

The IAD framework is mechanized computationally by defining a domain-specific rule language—Action Situation Language (ASL)—whose grammar is tailored to the seven core rule types, enabling direct translation to extensive-form games for agent-based "what-if" analysis (Montes, 2021).

In ASL, agents, initial states, and rules are defined via:

• agent(·)
• initially(·)
• rule(id, type, priority, if Cond then Consq where Constraints)

Rule types, including boundary, position, choice, and control, map directly to IAD rule categories. The ASL engine processes these declarations into a game tree $G = (N, H, P, A, u, Z)$, with nodes annotated by current states, available actions, payoffs $u_i(z)$, and probabilistic transitions.

An example two-agent coordination scenario demonstrates how altering a single choice rule (forbidding an action for one player) immediately generates a new game and equilibrium structure, illustrating the IAD framework's capacity for policy simulation and consequence analysis (Montes, 2021).

4. Machine-Readable Institutional Grammar: The IG Parser

Institutional Grammar 2.0, operationalized by IG Parser, provides a formal syntax ("IG Script") and toolchain to encode and decompose institutional statements for downstream IAD analysis (Frantz, 19 May 2025). IG Script is defined via a BNF grammar supporting:

• Core (atomic: Attributes, Deontic, Aim, Objects, Context, OrElse)
• Extended (logical combinations, nesting: AND/OR/XOR)
• Logico (component/statement-level semantic annotations)

Each rule-like utterance is parsed into a tree of components (e.g., who may, must do, what, to whom, under what conditions), which can then be represented in tabular or network form for analysis across jurisdictions or systems. IG Parser automates error checking, sub-statement indexing, and output in various formats (CSV, JSON, visual trees), supporting both bulk API-based policy corpus ingestion and network-based institutional studies.

Typical IG Script mapping for a compliance rule:

A(officer) D(must) I(fine) Bdir(violator) Bind(authority) Cac(cond=violation) O(sanction)

This mapping supports typological filtering, network construction (e.g., Actor–Action bipartite graphs), and statistical aggregation of rule complexity or semantic clusters for comparative studies (Frantz, 19 May 2025).

5. Measurement, Contextual Scoring, and Policy Functions

Each of the seven rules functions not only as a binary or categorical constraint but can be formalized as a measured contextual variable with a real-valued scoring function $$R_{\mathrm{type}}: \mathcal{C} \rightarrow \mathbb{R}$$, enabling quantitative comparison of institutional features across settings (Morgan et al., 2023).

Examples include:

• Rigidity of entry/exit ($R_b$), measured by number of credentials.
• Action-set delegation ($R_\pi$), measured by size or restrictiveness.
• Data transparency ($R_\iota$), measured by number of features/outputs shared.

Such scoring enables direct "Map and Govern" alignment with the NIST AI Risk Management Framework, providing a systematic workflow:

1. Elicit local rules-in-use ($R_b, R_\pi, ..., R_\tau$)
2. Quantify them per context
3. Map to oversight weaknesses/strengths
4. Design targeted governance interventions (e.g., strengthen boundary rules, enforce explainability in information rules) (Morgan et al., 2023).

6. Applications, Generalization, and Impact

The IAD framework, particularly in its formalized and computable instantiations, has been applied across domains including high-stakes AI oversight in healthcare, public sector governance, and comparative policy diagnostics (Morgan et al., 2023, Montes, 2021).

Key insights from recent research include:

• Oversight structures in practice are polycentric and team-centric, not reducible to individual "human-in-the-loop" models.
• The seven-rule decomposition foregrounds institutional variation, oversight vulnerability, and the risk of deskilling where AI delegation exceeds organizational boundary or choice rules.
• Machine-readable, modular encoding (e.g., IG Parser) enables integration with statistical, network, and agent-based methods, facilitating cross-jurisdictional, temporal, or policy-variant analyses (Frantz, 19 May 2025).

A plausible implication is that by standardizing the encoding and quantitative evaluation of rules in use, policy-makers can transparently identify leverage points for institutional reform or oversight adaptation, and researchers can systematically simulate the organizational consequences of regulatory changes within or across action arenas.

7. Comparative and Computational Analysis: Current Directions

Contemporary institutional analysis leverages the IG Parser and ASL-based platforms to construct rule databases, network models, and formal "what-if" games. Analysts can:

• Filter and re-aggregate rules by Deontic, Aim, or semantic annotation (e.g., all choices with stringency=high).
• Build actor–action bipartite networks or context–sanction paths.
• Compare the complexity metric (e.g., Degree of Variability) or centrality across institutional systems.
• Employ the ASL engine to recompile institutional descriptions under alternative rule sets, yielding comparative evaluations of predicted behavioral equilibria or welfare outcomes (Montes, 2021, Frantz, 19 May 2025).

Such pipelines close the loop between qualitative rule extraction, quantitative comparative analysis, and computational, game-theoretic policy exploration of the range of institutional configurations enabled (or obstructed) by the "rules in use."