Skip to main content

9. Research Directions and Open Challenges

Based on use case analysis in Chapter 8 and a survey of existing research, we organize research directions toward CADL realization and development. For each direction, we state what the current state of the art already provides (named tools and results), what remains open for CADL, and the concrete research challenges that CADL should address. Where an existing tool already solves part of a problem, we say which part CADL would hand to that tool and which question the tool cannot answer.

The starting point is this specification (Version 0.3) and the reference implementation cadl 0.3.36. It checks the satisfiability of contracts with Z3 [de Moura & Bjørner, 2008], detects protocol deadlocks by structural analysis, supports the SoS-DSL extension (Appendix E: lifecycle:, monitors:, deadline:), generates Python, Solidity, OPA/Rego, and Unity C# code, drafts CADL from natural language with cadl ai (a parse / type-check / retry loop, and with --verify the checks of cadl verify), and reports a mapping to IEC 62853 (Chapter 10).

Statements below that no work addresses a given problem are based on a limited literature search (mainly 2023–2026). They mean that we did not find such work, not that none exists.

9.1 Algebraic Composition Theory for Institutions​

Establishment of an algebraic foundation for safely composing and decomposing multiple institutional descriptions is necessary. Benveniste et al. [Foundations and Trends in EDA, 2018] theorized parallel composition of contracts C₁ ⊗ C₂ in component design. Since then the algebra of assume-guarantee contracts has been largely completed: hypercontracts give closed forms for composition and quotient (the specification of a missing component) over hyperproperties [Incer+, NFM 2022], and Pacti computes the algebraic operations of contract theory efficiently [Incer+, arXiv 2023]. OCRA checks the refinement of contracts written in LTL, including metric real-time constraints [Cimatti+, ASE 2013]; Kind 2 checks that the modes of a contract cover its assumptions [Champion+, CAV 2016]; and contracts have been composed across agents with quantitative temporal requirements [Dewes & Dimitrova, AAAI 2025]. SysML v2 [OMG, 2025] gives requirements an assume constraint and a require constraint, and SAWS² connects a fragment of SysML v2 to OCRA, nuXmv, and xSAP [FBK, SAWS²].

What remains open for CADL is the institutional interpretation of these operators. The parties of a CADL contract are autonomous organisations whose compliance depends on incentives rather than on the designer's authority; composition must therefore be indexed by institutional parameters and its guarantees must hold in equilibrium. Switching between whole operating modes (regimes), including the transitional phase, must be safe, and the algebra must stay open to constituents joining and leaving. For example, when composing household chore-sharing institution with shared resource usage institution, a theory that algebraically guarantees absence of obligation conflicts and incentive interference is required. Mode-aware contracts address components, and the evolution of SoS configurations has been checked without a contract algebra [Etzien+, ISORC 2013]. We found no work that models normative concepts (obligation, permission, sanction) in SysML v2.

The structure of the institutional parameter space (α, β, λ) is a second open area. Parameter synthesis for parametric Markov chains and MDPs already partitions a parameter space into regions certified to satisfy or to violate a property [Junges+, FMSD 2024]; Storm implements it and can be called from Python [Hensel+, STTT 2022], and monotonicity checking proves, region by region, that a probability increases or decreases with a parameter [Spel+, ATVA 2019]. These tools answer whether one design satisfies a threshold on a region. They do not answer which design is best where, which is the question of a regime map. The stochastic model extension (Appendix H, stochastic 0.1) is a first step on RD-1.3: it lets a file state a parametric Markov model in which α, β, and λ appear, defines the regime map as a query whose result is a set of certified cells, and fixes an export to the PRISM language. The reference implementation computes certified cells by parameter lifting (H.14). The strategic model extension (Appendix I) adds the game reading of Section 9.7 to that model, and the grounding extension (Appendix J) lets a payoff term take the size of a sanction from a norm and multiply it by λ, a first step on RD-1.5 in which λ is a parameter of a model, not yet of a composition. Deriving the model from contracts and scaling beyond small models remain open.

IDChallengeContent
RD-1.1Institutional Composition AlgebraBuild parallel composition and merging on the existing hypercontract results [Incer+, NFM 2022], and concentrate CADL's own contribution on sequential and choice composition and on the institutional interpretation of the three layers (Institution, Protocol, Algorithm), with soundness proofs.
RD-1.2Institutional Refinement RelationsFormal proofs that overview level → design level → verification level refinements preserve institutional properties. Use OCRA's temporal refinement [Cimatti+, ASE 2013] and Pacti's refinement check [Incer+, arXiv 2023] as the means of checking refinement between description levels.
RD-1.3Certified Regime Maps
by Parameter Synthesis
Translate CADL descriptions into parametric MDPs (PRISM/JANI) in which α, β, and λ appear in transition probabilities and rewards; generate certified regions per regime with Storm [Hensel+, STTT 2022]; prove qualitative design laws (for example, that raising α does not lower the mission success probability in a region) by monotonicity checking [Spel+, ATVA 2019]; and define comparative region verification ("which regime is best where") as a new problem. Exact partitioning is practical only for a few parameters [Junges+, FMSD 2024], so the scale of this approach is an open question.
RD-1.4Safety of
Regime Transitions
Lift mode exhaustiveness (as checked by Kind 2 [Champion+, CAV 2016]) and proof obligations before and after a mode transition to switches between centralised and decentralised regimes, with an explicit contract for the transitional phase. Course D of the hands-on materials shows that switching without waiting for the shared road to clear breaks the safety invariant.
RD-1.5Incentive-parameterised
Composition
Treat λ as a parameter of the contracts and compute, through the quotient, the range of λ under which the composition of constituent contracts refines the system contract. Related to the equilibrium analysis of RD-7.2.
RD-1.6Interoperability
with SysML v2
Map actors to part def and contracts to requirement def with assume constraint and require constraint [OMG, 2025]; delegate the algebraic computation to SAWS²/OCRA [FBK, SAWS²] or Pacti; add normative concepts and regimes as a CADL model library, since SysML v2 does not define them.

9.2 Extraction and Institutionalization of Tacit Knowledge​

Establishment of technology to explicitly formalize institutional knowledge implicitly shared in household rules and community customs as CADL descriptions is required. IndoorWorld [Wu+, Findings of EMNLP 2025] realizes behavior simulation of LLM agents in indoor environments, but the inverse problem of extracting institutional rules from observed behavior patterns (institutional reverse engineering) remains unaddressed. For timed properties, procedures that mine signal temporal logic specifications from traces have been surveyed [Bartocci+, Inf. Comput. 2022].

LLM-based translation into formal specifications is accurate only for simple targets. A fine-tuned model reached more than 95% accuracy in translating natural language into temporal logic with abstracted propositions, while GPT-4 with few-shot prompting reached 77.7% [Chen+, EMNLP 2023]. For TLA+, the best of thirty LLMs produced semantically correct specifications in 8.6% of the cases [Bisharat+, ICSOFT 2026]. The LLM-based generation of Symboleo [Zitouni+, 2025] assessed 38 prompt configurations against 16 error types and found the results promising but with many remaining errors. cadl ai already loops over parsing, type checking, and retries, and with --verify runs the SMT and deadlock checks; this assures syntactic validity and consistency, not that the description is faithful to the source text. We found no benchmark for translating deontic content (obligation, permission, remedy).

IDChallengeContent
RD-2.1Institutional Inference from
Behavioral Observation
Algorithm to infer potential institutional rules (implicit authority structures, information sharing patterns, incentive structures) from IoT sensor data and behavior logs. Includes estimating candidate deadline values and guarantee thresholds from simulation and monitor logs by specification mining [Bartocci+, Inf. Comput. 2022].
RD-2.2LLM-based Institutional
Description Generation
and Faithfulness Checking
Technology to generate CADL descriptions from natural language interview records and meeting minutes, extending LLM-based Symboleo generation [Zitouni+, 2025] to general institutional descriptions. Beyond the syntactic checks of cadl ai, check faithfulness: compare round-trip translations (NL → CADL → NL), validate the generated contracts against scenario traces in the CADL simulator, and show the reviewer the clauses with low confidence.
RD-2.3Tacit Knowledge Ontology
for Institutions
Build ontology systematically classifying and structuring implicit institutional knowledge. Map Ostrom [1990] IAD framework [Ostrom, 2005] to CADL vocabulary.
RD-2.4NL→CADL BenchmarkA benchmark of natural-language institutional descriptions paired with reference CADL descriptions and scenario traces, including deontic content, to measure the accuracy and faithfulness of RD-2.2 and of other translators [Bisharat+, ICSOFT 2026].

9.3 Institutional Digital Twins and Assurance​

Developing the concept of Digital Twins of Organization [Lyytinen+, J. Organization Design, 2024], establishment of technology to build and simulate institutions themselves as digital twins is the next-generation research direction. Lyytinen et al. argue that a full twin of an organization may not be feasible because organizations involve agency, conflict, and emergence; a reply argues that a twin built for a narrow purpose from digital trace data is feasible [Winkler & Kremser, J. Organization Design, 2024]. A CADL description can serve as the normative layer of such a narrow twin: it mirrors the agreed rules, not the whole organization. vCity [BSC] and Social Digital Twin [Fujitsu] advance twinning of physical and social aspects, but twinning of institutions (governance structures, rules, incentives) is, to our knowledge, not their main target.

Dynamic assurance cases connect a model to runtime evidence. Starting from dynamic safety cases [Denney+, ICSE 2015], ACCESS evaluates an assurance case both at design time and at runtime, where the validity of the case depends on runtime data checked by a monitor [Wei+, JSS 2024], and Trusta checks LLM-drafted assurance cases against formal constraints [Chen+, SCP 2025]. The notations are GSN Community Standard Version 3 [SCSC-141C, 2021] and OMG SACM 2.3 [OMG SACM, 2023]. The divergence between a model and the running system is studied as models@run.time [Bencomo+, SoSyM 2019].

Three problems remain open for CADL. First, the dynamic assurance cases we found assure a single technical system; we found none whose claims are the obligations, permissions, and amendment procedures of several organizations and whose evidence comes from monitors owned by different parties. Second, research on twin synchronization assumes that reality corrects the model. For an institution the model is normative: a divergence is either a violation by a participant (evidence against an assurance claim) or a drift in which the rules actually in force changed without an amendment of the CADL description (the twin is stale), and observation alone does not tell them apart. Updating the model automatically to fit observed practice would ratify violations. Third, statistical evaluation of institutions on a simulator needs confidence bounds, and experiments on institutions change the behaviour of the people who take part in them.

IDChallengeContent
RD-3.1Institutional SimulatorTechnology to automatically generate institutional simulation models from CADL descriptions and what-if analyze impacts of institutional parameter changes. MAPF simulator integration is initial target. Wrap the simulator as a black-box oracle for statistical model checking, attach a confidence interval to each parameter point, and mark boundaries where the intervals of competing regimes overlap as statistically indistinguishable; export to JANI to reach both exact tools (RD-1.3) and statistical model checkers such as modes [Budde+, TACAS 2018].
RD-3.2Divergence Classification and
Connection to Amendment
Detect divergence between institutional operation and the CADL description by checking monitor traces against it, and classify each divergence as a violation or as a drift. A violation is handled by the contract (violation handling and sanctions) and counts as evidence against the assurance case. A drift is not corrected by updating the description automatically: it goes through the amendment procedure (in Course D, a constitution gate and a versioned record of amendments) and from there into the change management process of IEC 62853 [IEC 62853:2018].
RD-3.3Verified Institutional
A/B Testing
Technology to operate different institutional designs in parallel and compare their performance. Institutional parameters proposed by reinforcement learning, as in AI Economist [Zheng+, 2022], or by LLM agents are certified with the methods of RD-1.3 and RD-7.1 before they are deployed. The "rationality" of LLM agents is not the rationality assumed by game theory, and this must be stated as a precondition of any claim. Because experiments on institutions affect their participants, they need a framework of ethics and consent.
RD-3.4Assurance Cases
Generated from CADL
Derive the claims of an assurance case deterministically from contract elements, use design-time static checks and operation-time monitor verdicts as evidence, output the case in GSN [SCSC-141C, 2021] or SACM [OMG SACM, 2023], and use it as the dependability case of IEC 62853. Re-evaluate the case when monitors report, as ACCESS does for technical systems [Wei+, JSS 2024]. Open points: argument patterns for institutional claims (fairness, accountability, amendability), aggregating confidence over monitors owned by different organizations, and referring to the constituents' own assurance cases as away goals.

9.4 Scalability and Computational Complexity​

Scalability of formal verification and code generation of institutions is one of the most critical challenges toward practical deployment. To complete verification of SoS descriptions containing 100+ actors and 50+ contracts within practical time, theoretical analysis of computational complexity of verification methods and development of efficient algorithms are necessary. Some bounds are already known for the strategic questions of Section 9.7: checking LTL properties on the equilibria of probabilistic systems is 2EXPTIME-complete [Gutierrez+, KR 2021]; designing incentives so that a property holds on equilibria is in PSPACE for LTL and in NP or Σ₂ᴾ for GR(1) goals [Gutierrez+, LMCS 2024]; and verifying incentives with quantitative goals is 2EXPTIME-complete [Hyland+, KR 2024]. Practical tools will therefore need restricted fragments such as GR(1) [Piterman+, VMCAI 2006], finite traces, or bounded horizons.

IDChallengeContent
RD-4.1Computational Complexity of
Institutional Verification
Determine computational complexity classes for consistency verification, deadlock detection, safety verification, realizability of temporal contract sets, and equilibrium properties of CADL descriptions respectively, starting from the known results above. Identify decidability boundary conditions and the fragments (for example GR(1) [Piterman+, VMCAI 2006]) that keep them tractable.
RD-4.2Abstraction-based
Scalable Verification
Methods to efficiently verify large-scale SoS through hierarchical abstraction of institutional descriptions. Application of counterexample-guided abstraction refinement (CEGAR) [Clarke+, 2000] to institutional verification, and compositional assume-guarantee reasoning per constituent for the equilibrium checks of Section 9.7.
RD-4.3Incremental
Verification
Confirm consistency through differential verification rather than full re-verification when actors join or leave or institutions change. Formulate leaving as the quotient C_sys / C_rest (the specification that a replacement must satisfy given the remaining constituents) and joining as re-checking the refinement of the new composition, computable with Pacti [Incer+, arXiv 2023] for the fragments it supports. We found no work that formulates join and leave this way; earlier work on evolving SoS checks the reachability of critical configurations without a contract algebra [Etzien+, ISORC 2013].

9.5 Multi-Cultural and Multi-Jurisdictional Support​

For CADL to be applied to international SoS, support for different legal systems, cultural norms, and language environments is necessary. The EU Data Act (Regulation (EU) 2023/2854) sets essential requirements for smart contracts used for the automated execution of data sharing agreements, but institutional compatibility with other jurisdictions (Japan, US, China, etc.) remains an unresolved challenge. For data usage, ODRL is used as the policy layer of data spaces [Iannella+, W3C Blog 2025], and the draft ODRL Formal Semantics of the W3C Community Group defines deontic states for permissions, prohibitions, and duties and separates access control from monitoring; it leaves remedies unformalized and the handling of conflicts pending [W3C ODRL CG, Formal Semantics]. We found no reliable source that maps the obligations of the EU Data Act to ODRL policies, and CADL does not assume such a mapping.

IDChallengeContent
RD-5.1Cross-jurisdictional Institutional
Compatibility Verification
Technology to detect legal contradictions when composing CADL descriptions from different jurisdictions (e.g., differences between EU GDPR [Regulation (EU) 2016/679] and Japanese Personal Information Protection Act [Act No. 57 of 2003]). For data-sharing norms, refer to the conflict semantics of ODRL as it is formalized [W3C ODRL CG, Formal Semantics].
RD-5.2Parameterization of
Cultural Norms
Technology to model cultural dimensions (collectivism/individualism, power distance, uncertainty avoidance, etc.) as institutional parameters and handle cross-cultural governance differences quantitatively.
RD-5.3Multilingual Institutional DescriptionEnable CADL overview-level descriptions in multiple languages with AI verifying institutional equivalence across languages. This goes beyond NFR-7 (internationalization), which only requires that human-readable text can be written in any language.

9.6 Norms and Time​

CADL contracts describe normative agreements, but norms are not yet first-class objects with their own state over time. Symboleo [Sharifi+, 2020] treats obligations and powers as objects with lifecycles (creation, activation, suspension, resumption, fulfilment, violation, termination), SymboleoPC translates Symboleo to nuXmv and checks LTL and CTL properties [Parvizimosaed+, SoSyM 2025], and Symboleo specifications have been translated into smart contracts [Rasti+, MODELS 2022]. The normative programming language NPL handles obligations with deadlines and sanctions in multi-agent organisations [Hübner+, AMAI 2011], and constitutive norms (counts-as) create institutional facts from other facts, as distinct from regulative norms [Boella & van der Torre, KR 2004]. Norms can also be synthesized automatically for multi-agent systems [Alechina+, KR 2022].

On the side of time and monitoring, RTAMT computes the robustness of signal temporal logic online, and its interface-aware variant separates input signals from output signals, which matches CADL's split between assume and guarantee [Ničković & Yamaguchi, ATVA 2020] [Yamaguchi+, STTT 2024]. R2U2 compiles temporal specifications into hardware or software observers [Johannsen+, CAV 2023]. An assumption can make more properties monitorable by restricting the traces a monitor must handle [Henzinger & Saraç, RV 2020]. Deontic runtime verification identifies the party to blame [Kharraz+, ICTAC 2023] and distinguishes where a violation is detected from where it occurred [Colombo+, ISoLA 2024]. Monitoring has been decentralized [Francalanza+, 2018], and shields synthesized from temporal safety specifications block unsafe actions [Alshiekh+, AAAI 2018], also per agent without communication [Melcer+, NeurIPS 2022].

In the core language, assume and guarantee are state predicates without temporal operators, deadlines exist only on lifecycle transitions (deadline:), monitors are hand-written rule: predicates, and Z3 is used only for satisfiability. The norms extension (Appendix F, norms 0.1) is a first step on RD-6.1, RD-6.2, RD-6.3, and RD-6.5: it defines first-class norms, bounded temporal operators with four-valued verdicts, counts-as rules, and what happens to running norms on an amendment. The generated Python code monitors norms at run time, and so do the generated Rego, as a step function that is given the stored state, and Solidity, for the contracts they support (F.10), and the norms of a contract can be checked at design time, in one instance of the contract or in up to four instances together, which start at observations that the environment chooses, also when the contract is amended: F.12 translates them, with discrete time and classes of values, into an MDP for PRISM [Kwiatkowska+, CAV 2011] and asks whether each norm can be violated, whether every activation can be kept, whether two norms are in conflict [Vasconcelos+, JAAMAS 2009], and whether a power is ever exercised, each "yes" with a trace that the runtime semantics of F.5 confirms. The grounding extension (Appendix J, grounding 0.1) is a first step towards the link to α, β, and λ: it derives from the norms of a contract an index of information sharing over declared items and an index of the concentration of powers, compares the declared α and β with them, measures how many norms carry a sanction or a reward, and gives sanctions a numeric magnitude that λ scales; the same indices are also computed over all contracts of an SoS. The indices are counts and are evidence about a contract, not the meaning of the parameters. The formal meaning of α, β, and λ in terms of norms, design-time checking that F.12 does not cover (nonlinear arithmetic over a quantity without a given domain, which is checked only under a domain assumed from the literals of the predicate, groups of compared quantities whose elimination exceeds the limit, such as, from cadl 0.3.32, a leaf that reads many of them through count when one of its values never occurs, linear arithmetic over many ordered quantities, or members read both through max and by their indices, instances of different contracts together, more than four instances, sets of actors whose size is not fixed, continuous time, and which party can avoid a violation), monitoring in the Unity C# runtime, monitoring in the generated Rego and Solidity of what they refuse (quantifiers, aggregates, wildcard and range references in predicates, index expressions, and amendments; in Rego also /; in Solidity also *, /, literals with more than six decimals, temporal operators inside another one outside the content of a norm, and contracts whose creation code exceeds the limit of EIP-3860), a gas of an observation of the Solidity monitor that does not grow with the observations inside a window for until in a nesting and for an operand that combines a nested operator with other parts (decided per observation from cadl 0.3.32), weighted indices, norms per regime, and the derivation of model probabilities from norms remain open. We found no work that handles first-class norm objects together with the institutional parameters α, β, and λ, the composition of many contracts in an SoS, and amendment during operation. We also found no tool that connects deontic blame semantics with efficient monitoring engines, or that lets several organizations, each observing only its own interface with unsynchronized clocks, agree on a verdict and on the responsible party.

IDChallengeContent
RD-6.1First-class Norm ObjectsA candidate norms: block whose norms have a kind (obligation, permission, prohibition, power), a bearer, a counterparty, a trigger, a deadline, and the handling of violation (a contrary-to-duty obligation or a sanction), with the lifecycle of Symboleo [Sharifi+, 2020]. Check norms at design time by translation to nuXmv, as SymboleoPC does [Parvizimosaed+, SoSyM 2025], and export data-sharing norms to ODRL. The holder of a power corresponds to β, the observability of trigger and fulfilment to α, and the size of a sanction to λ; whether the probability of compliance can be analysed as a function of λ and α is an open question.
RD-6.2Temporal Assume-Guarantee
and Four-valued Verdicts
Allow bounded MTL/STL in contracts and give four verdicts: pending (open and within its deadline), fulfilled, violated, and assumption violated (the provider is not to blame) [Henzinger & Saraç, RV 2020]. Translate them to RTAMT for the Python backend [Yamaguchi+, STTT 2024] and to generated observers in the style of R2U2 for the Unity C# backend [Johannsen+, CAV 2023].
RD-6.3Counts-as Bridge
to Protocols
Rules of the form "message m in protocol P counts as institutional event e in contract C" [Boella & van der Torre, KR 2004], so that protocol traces change the state of norms.
RD-6.4Cross-organizational
Monitoring and Blame
Placement of monitors (provider, consumer, neutral third party) [Francalanza+, 2018], uncertainty of clocks, tamper-evident records, and tracing the cause of a violation upstream along the graph of delegated contracts, using the blame semantics of deontic runtime verification [Kharraz+, ICTAC 2023] [Colombo+, ISoLA 2024].
RD-6.5Continuity of Norms
across Amendments
Define what happens to obligations that are active, suspended, or violated when a regime transition or an amendment takes effect: carried over, discharged, or transformed, together with transitional provisions and the time a version takes effect. To our knowledge Symboleo, ODRL, and NPL do not define this. Course D of the hands-on materials amends rules while the system runs and lists the missing syntax for transitional provisions; its research exercise on the speed of amendment is a first experiment. From cadl 0.3.32 the amendment extension (Appendix K) gives the syntax for the time a version takes effect and for transitional provisions for running contract instances.
RD-6.6From Monitoring
to Enforcement
Synthesize per-party shields [Alshiekh+, AAAI 2018] [Melcer+, NeurIPS 2022] or Simplex-style fallback behaviour from CADL contracts, using the lifecycle state machines as the finite abstraction. The shields we found act on physical actions; we found none for institutional acts.

9.7 Strategic Constituents​

CADL assumes that constituent systems keep their operational and managerial independence, so they may not comply with a contract when compliance does not pay. Rational verification checks whether a temporal property holds on the Nash equilibria of a multi-agent system: EVE takes modules with LTL goals and reduces the problem to parity games [Gutierrez+, AI 2020], and the problem has been extended to probabilistic systems [Gutierrez+, KR 2021]. Recent work moved from checking to design: equilibrium design chooses incentives within a budget so that a property holds on equilibria [Gutierrez+, LMCS 2024], incentive design handles static and dynamic incentives with quantitative goals [Hyland+, KR 2024], and norms can be synthesized against strategic objectives [Alechina+, KR 2022]. PRISM-games verifies Nash and correlated equilibria of concurrent stochastic games with social-welfare and fairness criteria [Kwiatkowska+, TACAS 2022] and has been extended to psychological games, whose payoffs depend on beliefs [Kwiatkowska+, 2024]. Bayesian incentive compatibility of mechanisms has been reduced to model checking [Mittelmann+, AAAI 2023]. Learning-based institution design such as AI Economist [Zheng+, 2022] gives no formal guarantee.

What remains open for CADL is a map rather than a point. Equilibrium design finds one incentive scheme; CADL needs the range of λ (and of the motivation parameters ρ and κ, Appendix C) in which compliance is an equilibrium. We found no tool that combines parametric transition probabilities with equilibrium computation, and little work on equilibria when constituents observe only part of the state, which is what α controls.

The strategic model extension (Appendix I, games 0.1) is a first step on RD-7.1 and RD-7.2. It reads the model of Appendix H as a turn-based stochastic game and defines when a named memoryless deterministic profile, in particular the compliance profile, is an equilibrium against unilateral deviations, and the equilibrium map: the cells of a region of parameters, λ among them, on which the profile is certified to be such an equilibrium or not to be one. It does not compute the equilibria of a game. With the grounding extension (Appendix J) the sanction term of a payoff is, by definition, the magnitude of a norm multiplied by λ, and under the conditions of J.7 the set of λ at which compliance is an equilibrium is empty or an upper interval. The extension is implemented from cadl 0.3.23 (I.14). Concurrent moves, mixed strategies, properties that must hold on every equilibrium, the mediator and the value of centralisation of RD-7.2, the motivation parameters ρ and κ, and partial observation (RD-7.3) remain open.

IDChallengeContent
RD-7.1Verifying that
Compliance is Rational
Translate CADL constituent systems and their goals into the input of EVE [Gutierrez+, AI 2020] or into concurrent stochastic games for PRISM-games [Kwiatkowska+, TACAS 2022], and check whether the compliance properties hold on every equilibrium. Use restricted fragments (Section 9.4) to keep the checks tractable.
RD-7.2Parametric
Equilibrium Analysis
Compute the region of λ, ρ, and κ in which compliance is an equilibrium. Model a centralized mediator as a correlated equilibrium and decentralized operation as a Nash equilibrium, and measure the difference between their values at each parameter point as the value of centralization (β). Relates to RD-1.3 and RD-1.5.
RD-7.3Equilibria under
Partial Observation
Equilibrium analysis when each constituent observes only what the information sharing degree α allows, with compositional reasoning per constituent for scale.

9.8 Governance of AI Agents​

LLM-based agents are becoming constituent systems of SoS, and the protocols that connect them are being standardized. The current specification of the Model Context Protocol (MCP) has a stateless core, requires Mcp-Method and Mcp-Name headers so that gateways can route and meter requests without parsing their bodies, and adds a framework for extensions [MCP, 2026-07-28]. MCP is a project of the Agentic AI Foundation of the Linux Foundation [Linux Foundation, AAIF 2025], and the Agent2Agent (A2A) protocol was contributed to the Linux Foundation [Linux Foundation, A2A 2025]. A preprint that compares MCP, A2A, and three other protocols finds that none of them supports voting or preserving dissent, and concludes that community governance is a layer missing above them [Kang & Diponegoro, arXiv 2026]. Enforcement is converging on deterministic checks at the tool-call boundary; for example, Amazon Bedrock AgentCore Policy authorizes tool calls with Cedar policies and denies by default [AWS, AgentCore Policy]. A preprint review of 38 studies finds that writing the specification is the bottleneck, that runtime monitoring is the most mature approach but gives guarantees for single steps, and reports a study in which blocking 94% of unsafe actions still left fewer than 5% of tasks completed safely and correctly [Dantas+, arXiv 2026]. From the side of institutional design, a preprint enforced a public governance graph with an append-only audit log and reduced severe collusion among LLM agents from 50% to 5.6% of runs, while a prompt-only "constitution" showed no reliable improvement [Bracale Syrnikov+, arXiv 2026]. Regulation such as the EU AI Act, with its requirements on record-keeping and human oversight, is a further motivation for executable governance; this chapter does not claim a mapping from its articles to CADL.

What remains open for CADL is the step from organizational contracts (roles, obligations, the provenance of authority, incentives) to deterministic enforcement at the protocol level; we found no work that takes this step. The institutional designs above are built by hand for each case, and cadl ai assures only the syntactic validity and consistency of what it generates (Section 9.2). The agent governance extension (Appendix G, agents 0.1) is a first step on RD-8.1, RD-8.2, and RD-8.3: it defines the decision that the norms of a contract give for a message at the boundary, delegation with a provenance chain, and escalation after a denial. Governance elements such as voting and dissent (RD-8.4) and plan-level verification (RD-8.5) remain open.

IDChallengeContent
RD-8.1Compiling Contracts
to Boundary Policies
Map actors to MCP clients and servers or to A2A agents, and compile contracts and authority into Rego [Open Policy Agent] (already generated), Cedar [AWS, AgentCore Policy], and Datalog policies, with the Mcp-Method and Mcp-Name headers [MCP, 2026-07-28] as the point of enforcement.
RD-8.2Provenance of
Delegated Authority
Authority parameters that carry the chain of delegation, so that a guard checks where an authority came from and not only the local state, which cannot show whether an action was passed on through delegations its originator was not entitled to make.
RD-8.3Recovery after
a Blocked Action
Protocols that define the path after a guard blocks an action (escalation to a human, an alternative provider, renegotiation of the contract), in response to the low task completion reported after blocking [Dantas+, arXiv 2026].
RD-8.4Governance Elements
Missing from Protocols
Describe membership, voting, dissent, escalation to humans, and audit records as CADL protocols, and distribute them as MCP or A2A extensions [Kang & Diponegoro, arXiv 2026]. Course D of the hands-on materials already combines a vote of the companies with a recorded human decision.
RD-8.5Plan-level and
Multi-agent Verification
Verify properties of plans and of the interaction of several agents, not of single steps, using the strategic methods of Section 9.7, and evaluate the generated guards on public agent-security benchmarks against existing guards.

9.9 Research Roadmap​

We organize the above research directions along a time axis and present a short-, medium-, and long-term research roadmap. The order follows the current implementation: items that extend what cadl 0.3.36 already provides come first.

PeriodKey ChallengesContent
Short-term
(2026-2027)
RD-1.1, RD-1.2,
RD-2.2, RD-4.1,
RD-4.3, RD-6.1,
RD-6.2, RD-8.1
Basic theory of institutional composition (RD-1.1) and of refinement between description levels (RD-1.2), faithfulness checking for CADL generation via LLM (RD-2.2), computational complexity analysis of verification (RD-4.1), incremental verification when actors join or leave (RD-4.3), first-class norms and temporal contracts with four-valued verdicts as extensions of the existing deadline: and monitors (RD-6.1, RD-6.2), and compilation of contracts to boundary policies as an extension of Rego generation (RD-8.1).
Medium-term
(2027-2028)
RD-1.3, RD-1.4,
RD-1.6, RD-2.1,
RD-2.4, RD-3.1,
RD-3.2, RD-4.2,
RD-6.3-6.5,
RD-7.1, RD-8.2,
RD-8.3
Certified regime maps by parameter synthesis (RD-1.3), safety of regime transitions (RD-1.4), SysML v2 interoperability (RD-1.6), institutional inference including specification mining (RD-2.1), an NL→CADL benchmark (RD-2.4), institutional simulator with statistical model checking, starting with integrated experiments with a MAPF simulator (RD-3.1), classification of divergences connected to the amendment procedure (RD-3.2), scalable verification methods (RD-4.2), counts-as, cross-organizational monitoring, and continuity of norms across amendments (RD-6.3 to RD-6.5), verification that compliance is rational (RD-7.1), provenance of delegated authority and recovery after blocked actions (RD-8.2, RD-8.3).
Long-term
(2028-2030)
RD-1.5, RD-2.3,
RD-3.3, RD-3.4,
RD-5.1-5.3,
RD-6.6, RD-7.2,
RD-7.3, RD-8.4,
RD-8.5
Incentive-parameterised composition (RD-1.5), tacit knowledge ontology (RD-2.3), verified institutional A/B testing (RD-3.3), assurance cases generated from CADL (RD-3.4), multicultural and multi-jurisdictional support (RD-5.1 to RD-5.3), enforcement by shields (RD-6.6), parametric equilibrium analysis and equilibria under partial observation (RD-7.2, RD-7.3), governance elements for agent protocols and plan-level verification (RD-8.4, RD-8.5).