The desired reasoning programme decides what to do by proving, in a formal language, that some strategy will achieve an assigned goal. That sentence packs in several requirements. The language must describe current and hypothetical circumstances, the effects of actions, which actions an agent can perform, what the agent knows and how knowledge may be acquired. It must also support a proof procedure capable of connecting those descriptions.

McCarthy and Hayes distinguish this aim from a purely behavioural test of intelligence. A black-box definition might call an entity intelligent when it solves hard classes of problems or survives an intellectually demanding environment. They instead take a calculated risk: organise the machine as a manipulator of represented facts, even though introspection could misdescribe human thought and other intelligent organisations may be possible.

Their working criterion demands an adequate model of the world, including mathematics, goals and aspects of the agent’s own mental processes. The entity must answer a variety of questions from that model, obtain more information when needed and carry out tasks permitted by its goals and physical capacities. Answers need not always be correct. Intelligence is therefore not logical omniscience, but coordinated representation, problem solving, information seeking and action.

This produces the paper’s major division. The epistemological part concerns a form of representation in which solutions follow from expressed facts. The heuristic part concerns mechanisms that find those solutions and decide what to do. Earlier AI work is characterised as concentrating on heuristic mechanisms. The paper deliberately concentrates on the epistemological side, even while acknowledging interaction between the two.

A representation is metaphysically adequate when the relevant aspect of reality could have that form without contradicting known facts. The paper lists particle systems, a universal quantum wave function and interacting discrete automata as examples for different purposes. Such representations support general theories, but obtaining observable consequences from them is an additional task.

Epistemological adequacy is relative to a person or machine. It asks whether the representation can practically express the facts actually possessed. A complete particle description may be compatible with physics yet unusable for saying that John is at home, dogs chase cats or a telephone has a particular number. Ordinary language expresses many communicable facts, though even it may not capture everything involved in recognising a familiar face.

Heuristic adequacy asks whether the reasoning processes used to solve a problem can be expressed in the language. McCarthy and Hayes introduce this third category tentatively and do not develop it fully. The distinction nevertheless prevents a common collapse. A theory may contain enough information in principle, but if every useful consequence requires reconstruction at an inaccessible level, it does not support the agent’s actual reasoning.

The interacting-automata picture demonstrates the separation. It is metaphysically attractive for causation: component outputs affect other components over time. Yet it becomes epistemologically inadequate for people because nobody knows the enormous internal-state table of another person. Geometry and changing interconnections are awkward too. The authors therefore use automata to explicate “can” and cause, then turn to a fluent-based formalism better suited to facts an agent can state.

The reasoning programme interacts through sensory, motor and communication devices. Internally it may use images, regions, bodies, trajectories, speech signals, phonemes and parse trees. One representation nevertheless dominates: sets of sentences in a logical language rich enough for functions, descriptions, conditionals and sets, with modal operators left as an open question.

Dominance does not mean every pixel becomes a theorem. Other data structures receive linguistic descriptions stating their relation to the world. Subroutines receive descriptions of what they do. Beliefs about world dynamics, consequences of strategies, goals, subgoals and progress are all expressed in sentences. Information is adequate for a problem if some strategy’s success is a logical consequence of this combined body.

The reasoning programme then searches for a strategy whose goal achievement it can prove and executes the strategy once found. Strategies may include subgoals, further proof search or information acquisition. An intellectual action can therefore be part of a practical plan. “Find a telephone number” is not merely preprocessing; it can be a strategy step needed before dialling.

The Missouri programme removes the heuristic achievement requirement. An experimenter supplies proof steps; the programme checks them and acts when persuaded that it ought to. This isolates whether the proposed language and axioms are sufficient. If even a human-guided proof cannot be stated and checked, better automatic search will not repair the representational gap.

In an interacting discrete-automata system, each component’s next state depends on its current state and incoming signals. The system can be deterministic: fixed initial states and external inputs determine its entire future. Yet an internal reasoner has not computed its own future choice in advance. For deliberation, it needs a notion of available alternatives that does not collapse into the single action it will actually take.

McCarthy and Hayes conceptually cut the outgoing signal lines of a subautomaton and replace them with external inputs. The component can bring about a condition if some sequence on those lines eventually makes the wider system satisfy it. This counterfactual intervention holds the surrounding mechanism fixed while varying the candidate agent’s outputs. “Can” is therefore evaluated against a model and a chosen boundary.

Different cuts yield different abilities. Intervening directly at joints may make an impossible endurance sequence look available because muscles are bypassed. Cutting at spinal inputs retains muscular limits but may count unlearned unicycle movements. Cutting at an imagined organ of will includes more knowledge and control constraints. The paper concludes that no single notion of “can” serves every context; several progressively constrained meanings are useful.

External inputs add another distinction. One definition requires one fixed output sequence to achieve the goal no matter what external signals occur. Another permits outputs to depend on inputs, equivalent to replacing the component with a controller that reacts appropriately. The second is strategy-like ability under observation; the first is open-loop robustness. The authors label both possibilities rather than selecting one.

The set Sit contains actual and hypothetical situations. The representation never completely describes one; it records facts about it. Those facts may support further claims about the same situation, future situations or situations an agent could bring about. A hypothetical branch need not settle every detail. One can reason that selling a car for a low offer would remove ownership while leaving the seller’s exact later mood undefined.

A fluent is a function on situations. A propositional fluent returns true or false; a situational fluent returns another situation. “Raining in Oxford” becomes a function whose truth can vary across situations. Writing the situation argument explicitly prevents accidental timelessness. Suppressing it remains convenient when context determines which situation is meant, much as ordinary speech leaves its temporal setting understood.

Time itself is a fluent. Two alternative situations may have the same clock time, so identifying situation with time would collapse distinct counterfactual branches. A situation records a complete possible state; time orders or labels it but does not exhaust its identity. This is what lets one compare “after opening the safe” with “after walking away” at corresponding times.

The future operator associates a present situation with later situations satisfying a condition. The authors use temporal operators to express eventual and universal future and past claims, and consider a situational fluent selecting the next future situation in which a condition holds. These devices allow causal laws and strategy results to be stated without enumerating a complete universe-state.

An action law can say that if person p has a fitting key k for safe sf in situation s, then sf is open in result(p, opens(sf,k), s). This is a sufficient-condition schema: it relates a precondition to a fluent in a hypothetical successor. It does not assert that p will perform the action, nor that every manner of manipulating a safe succeeds.

Strategies combine actions as ALGOL-like programmes. A finite sequence is simplest, but assignments, conditional jumps and loops are allowed. External actions are procedure calls; internal variables and labels have “mental” significance. An occurrence of an external action can be rewritten as an assignment to the distinguished situation variable: replace s by the result of the agent executing that action in s.

This rewriting connects action representation to programme verification. Drawing on Manna’s mathematical theory of computation, the paper explains how an initial condition, a final goal and a strategy can be transformed into a first-order sentence. A proof must establish termination and goal satisfaction. Loops require induction-like facts; a street-walking strategy also needs geography, action effects and observations.

The situation variable differs from an ordinary stored variable. A machine assigns an integer internally by changing memory. It “assigns” a new situation only by acting in the world. Likewise, reading a fluent whose value is not already known may require an observation. The formal programme thereby mixes internal computation, epistemic action and physical action without treating them as identical operations.

For a key safe, possession is straightforwardly represented: the person has an object whose shape fits the lock. A combination is different. Writing that the person “has 34125” makes little sense, even though 34125 may be the safe’s combination and entering it would open the safe. The physical effect of entering the right number is independent of anyone’s knowledge; feasibility is not.

The authors introduce a term for the person’s idea of the combination in a situation. Manipulating the safe according to that accessible term is feasible. If the person’s idea equals the actual combination, the effect law proves success. The device preserves referential transparency in the object language: equals may replace equals. But it pushes feasibility into a metalanguage distinction, because a strategy containing the person’s accessible idea can be feasible while the extensionally equal hidden number is not.

The telephone example sharpens the point. A strategy looks up a person’s number and then dials it. Looking up is an epistemic action whose result supplies knowledge required by the next step. To prove connection, the theory must also maintain that the caller still has a telephone and that relevant facts persist. This is where knowledge representation and the frame problem collide.

Natural-language knowledge is referentially opaque: from knowing that someone’s telephone number is a particular number and an equality between that number and another description, substitution can create conclusions the agent need not recognise. The George IV and Walter Scott puzzle illustrates the general issue. The paper expects a richer formalism with restrictions on substituting equals, rather than pretending transparent first-order terms fully solve knowledge.

The immediate example is almost comic: after looking up a number, the caller still has the telephone. Yet a strictly local effect axiom saying only that lookup produces knowledge does not entail continued possession. Logic’s refusal to invent persistence is correct. The modelling burden lies in expressing a general reason why unrelated facts remain unchanged.

The first proposed remedy uses a frame resembling a state vector. Attach a set of fluents to it; describe an action by assigning the fluents it changes; presume the others unchanged through the general semantics of assignment. Walking changes location, for example, while the programme structure preserves other components. Feasibility conditions constrain what may appear on the right side of an assignment.

This solution is not presented as final. Components independent in one context can be coupled in another, suggesting block structure. A move may normally leave possession unchanged but not if an object is dropped. The paper points toward other approaches in its formal-literature discussion and literature review. Its major achievement here is to expose and name the scaling problem precisely.

The frame problem later accumulated several meanings, including relevance in general reasoning. In the 1969 section, the concrete issue is representational: how to avoid separately asserting the non-change of every unaffected fluent for every action. This companion keeps that technical core distinct from broader philosophical descendants.

Each state is a partial description: a map from fluent names to values. An action declares preconditions and effects. In frame mode, the successor begins as a copy of the current map and effects overwrite named entries. In sparse explicit mode, the successor begins empty; effects are added, followed only by persistence facts explicitly declared for that action. An absent entry is unknown, not false.

The key-safe strategy needs the possession and fit facts initially and then applies an open action. Both update modes can establish that the safe is open, because the goal is a direct effect. The combination strategy first observes the combination and then enters the agent’s known value. Sparse update does not preserve the safe’s combination fact across observation, so the opening precondition cannot be shown.

The telephone strategy recreates the paper’s persistence problem. Looking up supplies a number, but the next action also needs continued possession of a telephone. Frame mode preserves that fluent and reaches conversation. Sparse update declares that the directory remains available but deliberately omits telephone persistence, so dialling is not proved feasible.

Three invariant tests audit every configuration. Declared effects must appear in every successfully produced successor. Frame mode must preserve all previously known non-effects. Finally, a strategy may be reported successful only when every action’s preconditions were known and the explicit goal holds in its final situation. Hypothetical execution never mutates the initial object, ensuring branches remain distinguishable.

The result function is acknowledged as approximate. A counterfactual action other than the one actually performed need not determine a unique complete mental state. “What would follow if I challenged someone?” depends on how the unlikely action is produced. Result belongs to a useful representation, not as a primitive object transparently written into the world itself.

Formal literatures add history to formal language: what may be said next can depend on what has already been said. The authors sketch modal notions such as normally and probably to address indefinitely many possible exceptions. Looking up a telephone number normally yields knowledge; dialling normally establishes conversation. One cannot list every torn page, faulty entry, outage or sudden incapacity. They expressly warn that the proposal is tentative.

Probability is required eventually, but attaching a number to every sentence is rejected. Quantified statements resist a simple interpretation in degrees of conviction, and the numerical information needed is often unavailable. A formalism that mandates absent statistics would fail epistemological adequacy. The point is not that probability is useless; it is that representation must match the information actually obtainable.

Parallel processes pose another mismatch. ALGOL-like strategies serialize activity, while real situations change through simultaneous processes and interactions. Contemporary simulation languages were judged too restricted, and correctness theory for parallel programmes underdeveloped. The open list therefore spans logic, control, uncertainty and concurrency rather than pretending one calculus closes the world.

The primary source defines situations as complete instantaneous states and fluents as functions on them; introduces temporal, action, strategy, result, knowledge and ability devices; connects strategy correctness with programme verification; and gives the frame problem its classic AI formulation. It also supplies unusually explicit criteria for judging representations and explains why general AI makes metaphysics and epistemology operationally testable.

The paper does not present Reiter’s successor-state axioms, a modern basic action theory, regression algorithms, GOLOG, event calculus, answer-set action languages or a complete non-monotonic solution. Its proposed frames and formal literatures are sketches. Its automaton analysis of ability and causation is model-relative. Its knowledge formalism is knowingly strained by referential opacity.

Its situation is not merely a recorded action history in the later common presentation. It is a complete world-state at an instant, including possible alternatives. Its result function takes a person and strategy as well as a starting situation. Later formalisations often sharpen or reorganise these ideas. This companion uses the primary paper’s vocabulary when reconstructing its claims and labels later comparisons as later.

The evidence is conceptual, formal and illustrative. The authors claim an advance over earlier formalisms by accommodating loops and strategies that acquire knowledge. They outline how a first-order proof obligation may establish termination and a postcondition. They do not report running planners, comparative experiments or scalability results. Publication fairness requires preserving that distinction.