Glossary
Working vocabulary for CADL and the surrounding research programme. Terms are grouped by topic. Links point to the chapters where each term is introduced in depth. Where it helps, an entry ends with an example.
Core language
- CADL — Contract Architecture Description Language. A DSL for specifying the Institution / Protocol / Algorithm layers of a System of Systems in a single, verifiable document. See the Specification introduction. Example: CADL targets a wide range of institutions, from household chore-sharing rules to the governance of large-scale urban transport.
- Institution layer — The layer that states who participates and
under which rules:
actors:,contracts:, andtransitions:(the conditions for switching between regimes). The same definition is used in Section 2.2 and Appendix A, §A.4. In the IR of the reference implementation,institutionholds the actors and contracts, and the transitions are a separate top-level list. Each contract states its parties,assume/guarantee,authority,information,responsibilities,incentives, andviolation, and carries the governance parameters α, β, λ. - Protocol layer — Coordination procedures: message passing, compute steps, conditionals, parallel blocks, and barriers between actors. Defined in Section 5.2.4 and Appendix A, §A.5.
- Protocol — The interaction procedure between actors, with timing constraints: who sends what to whom, and in which order. Example: (1) the customer orders → (2) the dispatcher assigns a robot → (3) the robot delivers → (4) the customer is notified.
- Algorithm layer — Reference to central or local algorithms
consumed by the protocol steps. Each algorithm is keyed by a function
name and has a
centraland alocalentry. - Institution — A general term for the rules, authorities, role assignments and reward mechanisms between people or systems, from codified laws and contracts to implicit household rules. CADL describes an institution through three aspects: authority structure, information sharing and incentives. Example: "Children finish their chores before dinner" and "they receive 100 yen when done".
- Actor — A constituent system, agent, or role participating in the
SoS. Identified by its
id, possibly parameterised by a range. Defined in Section 5.2.2 and Appendix A, §A.3. Example: in a delivery system, the robots, the dispatchers and the customers are each actors. autonomy— Actor attribute:low,medium, orhigh(defaultmedium). Fromcadl0.3.10 a value other than these three is an error; the value is not checked against β or anything else.cadl sim-gen -t gouses it to give the actor only the central algorithm (low), only the local one (high), or both (medium).- Stakeholder — Everyone with an interest in an institutional design: the actors that take part in it, the people it affects, and those who supervise it. Example: the users, delivery companies, municipality and nearby residents of a delivery system.
- Contract — Assume-guarantee constraint attached to a set of parties and an authority. The formal unit of institutional design. Defined in Section 5.2.3 and Appendix A, §A.4. Example: "As long as the robot operates normally (assume), delivery arrives within 1.2× the promised time (guarantee)."
- Assume-guarantee — The basic structure of a contract: if the assumptions
(
assume) are satisfied, the guarantees (guarantee) are promised; if an assumption fails, the guarantees no longer bind. Example: assume "road accessibility ≥ 80%", guarantee "delivery success rate ≥ 95%". - Obligation — Something an actor must do: a contractual responsibility whose violation may be penalised. Example: "Data providers must keep their data accuracy at 95% or above."
- SLA (Service Level Agreement) — An agreement, in numbers, on the level of quality a service provider guarantees to its users. Example: "data latency within 5 s", "uptime ≥ 99.9%", "delivery within 1.2× the promised time".
- Authority structure — Who has the authority to decide what, from centralised (one actor decides) to distributed (collective decision); its degree is β. Example: β = 0.8, the dispatcher decides almost everything; β = 0.2, each robot decides autonomously.
- Incentive — The reward and penalty mechanism that encourages the behaviour of actors; its strength is λ. Example: "on-time delivery: bonus", "late delivery: −50% of the base fee".
- Transition — A switch between regimes, declared with
fromandto(both required) and optionallycondition,protocol, andsafety_invariant. Defined in Appendix A, §A.7.
Governance parameters
The letters α, β, λ are used in two distinct parameterizations. They share symbols and the range [0, 1]. β points the same way in both (higher = more centralized), but α and λ have different meanings.
Core language (per contract; Chapter 5)
- α (alpha) — Information sharing degree. 0 = no sharing (each actor
has only local information); 1 = full sharing. Written in a contract's
informationblock. Example: α = 1.0, every family member can see the schedule; α = 0.4, IoT data shares metadata only. - β (beta) — Decision centralization degree. 0 = distributed (each
actor decides autonomously); 1 = centralized (a single actor
decides). Written in a contract's
authorityblock. Example: β = 0.8, the dispatcher decides routes centrally; β = 0.2, each data provider decides on its own what data to publish. - λ (lambda) — Incentive strength. 0 = directive-based; 1 = market
mechanism. Written in a contract's
incentivesblock. Example: λ = 0.6, pocket money affects behaviour moderately; λ = 0.9, strong control through market mechanisms.
CADL Explorer simulation configuration (top-level governance: block;
Appendix C)
- α (alpha) — Autonomy level of the agents. Higher α ⇒ more local judgment.
- β (beta) — Centralization level. Higher β ⇒ the central authority is more dominant.
- λ (lambda) — Exploration probability. λ = 0 means deterministic routing.
Motivation extension (Appendix C)
- ρ (rho) — Motivation sensitivity. Governs how strongly decisions respond to per-agent motivation values. ρ = 0 is motivation-blind; ρ = 1 fully couples budget / wait decisions to motivation.
- κ (kappa) — Scale factor for motivation-to-budget conversion.
- Budget base — Baseline resource quota assigned to an actor before motivation adjustments.
- Motivation profile — Distribution of motivation across actors:
uniform,linear,polarized, orcustom(an explicit list of values).
System of Systems (SoS)
ISO/IEC/IEEE 21841:2019 classifies SoS into the four types below by their governance structure. CADL can describe all four.
- SoS — System of Systems. A collection of independently managed systems that cooperate toward a shared purpose while retaining operational autonomy. Example: delivery robots of several companies that cooperate to make city-wide delivery efficient: each robot operates independently, but together they form an efficient delivery network.
- Directed SoS (D-SoS) — A central authority dictates coordination among constituent systems. Example: a military chain of command, in which headquarters controls the actions of all units.
- Acknowledged SoS (A-SoS) — Constituent systems retain autonomy but recognise a shared governance arrangement. Example: robot delivery, in which the dispatcher manages the whole but each robot avoids local obstacles on its own.
- Collaborative SoS (C-SoS) — Cooperation emerges through mutual agreement without a dominant authority. Example: family rule-making in family meetings; the running of a condominium management association.
- Virtual SoS (V-SoS) — No formal coordination mechanism; cooperation is implicit and opportunistic. Example: the Internet as a whole: its services are run independently but together form a huge ecosystem.
- Five-layer framework — A classification of SoS design into five layers [Shimoyama & Matsubara, 2026]: (1) objective (policy), (2) institution and governance, (3) control and coordination, (4) execution and operation, (5) environment and disturbance. CADL mainly describes layers 2 and 3; the correspondence with CADL's own three layers is given in Section 2.2. Example: policy "safety first" → governance "collision-avoidance rules" → control "route planning" → execution "motor control".
Institutional dynamics
- Regime (also written operational mode) — The complete set of
institutional settings effective at a given time. As environmental
conditions change, the optimal regime also changes. In the v0.3 syntax
a regime is a name used in
transitions:; fromcadl0.3.32regimes:binds parameter values and the contracts in force to it (Appendix A, §A.7.1); see Section 5.1 and Appendix A, §A.7. Example: a normal regime in which the dispatcher manages centrally, and a failure regime in which each robot decides on its own. - Regime map — A map of which regime is optimal under which
environmental conditions: the environmental parameter space divided
into regions, each assigned a regime. The stochastic model extension
defines it precisely as a query on a parametric Markov model
(Appendix H, H.9), and
cadl model-check --regime-map ID --certifycomputes it as a set of certified cells. Thecadl regime-mapcommand reports something narrower: the graph of regime names and transitions. Example: low failure rate and low latency → centralised regime; high failure rate or high latency → distributed regime. - Regime transition (also: institutional transition) — Switching from one regime to another in response to a change in environmental conditions. Safety (collision avoidance, for instance) must be kept during the switch. Example: when communication latency exceeds 300 ms, switch from centralised to distributed operation, keeping "no collisions, no lost orders" during the switch.
- Safety invariant — A condition that must never be violated, also during a regime transition. Example: "robots do not collide with each other", "order data is not lost".
SoS-DSL extension
Terms of the SoS Contract DSL extension; see Appendix E.
- Lifecycle — The
lifecycle:block of a contract: itsstates, theinitialandterminalstates, and thetransitionsbetween them. - Contract instance — One execution of a contract, created when its initial trigger fires. It advances through the lifecycle states.
- Monitor — An entry of a contract's
monitors:block: what toobserve, thesampling(event or periodic), and arulepredicate. - Severity — Grade of a violation:
Minor,Major, orCritical. deadline/on_violation— A time bound on a lifecycle transition. When it expires, the instance moves to the state named byon_violation.transition, with the given severity.on_match— What a monitor does when its rule matches: report a violation, move the instance to a target state, or both.
Norms extension
Terms of the norms extension; see Appendix F.
- Norm — An obligation, permission, prohibition, or power of one
bearer, declared in the
norms:block of a contract. - Obligation / Prohibition / Permission — What the bearer must bring about (by a deadline, if one is given), must not bring about, or may bring about.
- Power — The ability of the bearer to perform an institutional act that activates, suspends, resumes, or discharges other norms.
- Counts-as rule — An entry of
counts_as:that makes a message or a condition create an institutional event of the contract. - Institutional event — A named event of a contract, created by a counts-as rule; it can trigger, end, or fulfil norms and is the content of powers.
- Norm instance — A norm within one contract instance; in version 0.1 each norm is judged at most once per contract instance.
- Contrary-to-duty norm — An obligation or prohibition that is
activated when another norm is violated (
on_violation.reparation). - Verdict — The value of a norm instance or a temporal guarantee on the
trace observed so far:
pending,fulfilled,violated, orexcused(the norm failed after an assumption of its contract had failed). - Amendment — The replacement of a contract by a new version while it
runs;
on_amendment:says whether an active or suspended norm is carried over, discharged, or transformed.
Agent governance extension
Terms of the agent governance extension; see Appendix G.
- Tool server / Tool call — An actor bound as
agent.kind: tool_serverthat offers namedtools:; a tool call is a message to it whose name is one of its tools. - Enforcement point — Where a message is decided before it takes
effect. The decision is
allow,deny, orescalate, computed from the prohibitions and permissions of the governing contracts. - Hold / Escalation — A message that is neither allowed nor denied
but held until the actor named by
on_denied.escalate_toapproves or rejects it, or itstimeout:passes. - Delegation / Provenance chain — A permission passed on by its bearer to another actor; the derived permission records the chain of holders back to the bearer and is valid only while every link is.
- Audit record — What an enforcement point records for a decision: the result, the time, the instances, and the outcome of a hold.
Stochastic model extension
Terms of the stochastic model extension; see Appendix H.
- Stochastic model — The
model:block: finite-domain variables, guarded probabilistic commands, and parameters, read as a parametric discrete-time Markov chain. - Design — A regime together with the parameter values it assigns; each design has its own Markov chain.
- Property — A statement about a design's Markov chain: the probability of reaching a state, or an expected reward, as a query or with a threshold.
- Regime map query — The question of which design is best for an
objective in which part of a parameter region (
model.regime_maps). Its result divides the region into cells that aredecided(with the best designs),infeasible, orundecided.
Strategic model extension
Terms of the strategic model extension; see
Appendix I. The reference implementation
provides it from cadl 0.3.23 (cadl game-check).
- Strategic model — A
model:block withkind: smg: the model of Appendix H in which some commands are owned by players, read as a turn-based stochastic game. - Player — A decision maker of a strategic model, standing for declared actors; it chooses among the commands it owns.
- Profile — One memoryless deterministic strategy per player,
named in
model.profilesby a list of rules. The compliance profile is the one the author declares as the behaviour the contracts ask for. - Payoff — What a player gains: a weighted sum of expected rewards accumulated until a target.
- Nash equilibrium / Deviation gain — A named profile is a Nash equilibrium when no player gains by changing its own strategy alone (a unilateral deviation); the deviation gain of a player is its best-response value minus its payoff under the profile.
- Equilibrium map — The answer to the question of where in a
parameter region a profile is an equilibrium: cells that are
equilibrium,not_equilibrium(with its deviators), orundecided.
Grounding extension
Terms of the grounding extension; see
Appendix J. The reference implementation
provides it from cadl 0.3.23 (cadl norms-indices).
- Derived index — α̂ or β̂: a ratio of counts computed from the norms of one contract. α̂ is the fraction of (declared item, party) pairs that a sharing entry or a sharing norm gives; β̂ is the largest share of the powers borne by one single actor.
- Incentive coverage — χ: the fraction of the obligations and prohibitions of a contract that carry a sanction or a reward. It is not an index of λ, which is a scale.
- Information item — A named piece of information of a contract, declared with the party it stems from.
- Magnitude — The size of the sanction or of the reward of a norm at full incentive intensity; a strategic model multiplies it by λ.
Verification
The checks of the reference verifier are described in
Section 6.3; the verification: block is specified in
Appendix A, §A.8.
- Formal verification — Checking mathematically, by computer, whether stated properties (such as freedom from contradiction) hold for an institutional description; it finds contradictions that human review tends to miss. Example: the consistency of two contracts is checked with an SMT solver, and freedom from deadlock of a protocol by structural analysis (Section 6.3).
- Contradiction detection — Finding rules in an institutional description that contradict each other, such as one contract giving an authority to A and another giving the same authority to B. Example: "two chores assigned to the same time slot"; "privacy constraints contradict a data-publication requirement".
- Model checking — Examining every state a system can reach to confirm that no undesirable state is reached. Example: confirming that no state in which robots collide can be reached during a regime transition.
- Reactive synthesis — Synthesising automatically a controller that satisfies a specification of how a system should behave. Example: from an AI safety contract, a controller that always prevents harmful actions and escalates to a human when uncertain.
- Runtime monitoring — Checking, while an institution is in operation, that the conditions of its contracts are kept, and warning or acting when one is violated. Example: checking "delivery_time ≤ promised_time × 1.2" throughout a delivery and reassigning when it is likely to be exceeded.
- Safety property — "Nothing bad happens." Expressed as an invariant over the system state.
- Liveness property — "Something good eventually happens." Typically a ◇φ (eventually φ) temporal formula.
- Fairness property — A scheduling or resource-allocation constraint: no participant is indefinitely starved.
- Invariant — A predicate that must hold in every reachable state.
- Verification methods —
smt(SMT solving, e.g. Z3 [de Moura & Bjørner, 2008]),model_check,simulation,proof. The reference implementation implementssmtonly. - Deadlock — A global state where no protocol step is enabled.
- Regime transition safety — No invariant is violated during a transition between regimes.
Pipeline and toolchain
- IR (Intermediate Representation) — The three-layer data structure the toolchain builds from a CADL source file. Consumed by the simulator config generators.
- Code generation — Generating working program code automatically from a CADL institutional description: the implementation from the blueprint. Example: from the delivery SLA, a skeleton of the dispatch algorithm and code that monitors SLA violations.
- Codegen target — An output format named in the catalog of
Appendix D:
unityandgo(simulation config),ros2(runtime nodes),python,solidity(smart contracts),opa(Rego policies),unity-csharp(contract runtime for Unity), or a user-defined target. - Simulator config — Environment + actor + governance settings passed to the downstream simulator (e.g. Unity).
- Pipeline —
CADL → IR → Simulator config → Experiment → Evaluation. - CADL Explorer — Interactive web app that walks the pipeline end-to-end. See cadl-explorer.streamlit.app and the source repository.
- Saved comparisons — Session-local record of the comparisons saved in CADL Explorer (Save this comparison), exportable as CSV / JSON. Called "Run history" before CADL Explorer v0.4.0.
- CADL id / IR id / config id — The first 12 hex digits of the SHA-256 of each stage (the CADL config, the IR, the simulator config), shown by CADL Explorer. Equal CADL ids with equal seed sets produce identical results.
Evaluation metrics
- Throughput — Completed tasks per unit time across the fleet. CADL Explorer reports it as the total deliveries of one 300 s run.
- Autonomy — Share of decisions made locally (vs dictated centrally).
- Fairness — Distribution of workload or reward across actors.
- Region (performance–autonomy plane) — Convex area covered by a set of experiment runs in the 2-D plane of throughput × autonomy.
Related standards and acronyms
- IEC 62853 — Open Systems Dependability standard [IEC 62853:2018]. CADL's framing of the system life cycle, consensus building, and accountability draws on open systems dependability; CADL does not claim conformance to the standard. Example: describing in CADL the consensus-building process of an institutional change, or the accountability mechanism of a fault response.
- ISO/IEC/IEEE 21841 — Taxonomy of Systems of Systems (Directed / Acknowledged / Collaborative / Virtual) [ISO/IEC/IEEE 21841:2019].
- EBNF — Extended Backus–Naur Form [ISO/IEC 14977:1996]. Used in Appendix A to define CADL's concrete syntax.
- SMT — Satisfiability Modulo Theories. The backbone of CADL's
method: smtverification. - OPA / Rego — Open Policy Agent [Open Policy Agent] and its policy language; a codegen target for institutional constraints.