# An Evolving Semantic Representation Language **Status:** Open research intention and candidate mechanisms. **Edition:** English editorial reconstruction. Earlier conceptual ancestors are retained as ancestors, not retroactively described as a completed contemporary architecture. The formal notation and evaluation designs below are candidate models already distinguished from the motivating intention. ## 1. The target exceeds memory compression The central aim is a system that develops new concepts, relations, and operations from experience, applies them to later problems, and can change the representational organization it uses while operating. Remembering facts, merging repeated fragments, generating concepts, changing representations, and changing how representations are learned are related but different achievements. A new semantic unit need not abbreviate two particular tokens. It might identify a relation across examples, bind variable roles, encode conditions of applicability, or transform the way a problem is described. ## 2. Conceptual ancestry without retroactive completion Earlier formulations explored abstract language, multidimensional words, representations that unfold differently for different readers, and relations among images or conceptual matrices. Later directions concerned paths that alter a semantic space, actions becoming reusable memory, concepts not restricted to human words, and representations that can affect thought without appearing as output words. These are conceptual continuities. They do not show that an implemented latent-token or adapter architecture existed in the earliest formulations. A later correction makes the difference explicit: combining tokens is a subproblem of creating a new meta-representation language, not its complete definition. ## 3. What a language includes A candidate model writes: L_t = (V_t, T_t, C_t, interpretation_t) V denotes available units, T their types and roles, C their composition or application rules, and the final component their interpretation. Units could be symbols, vectors, programs, or combinations. A public concept record might retain an identifier, roles, conditions, meaning, and evidence. Those fields describe an inspection layer; they need not all be literal components inside a token. Merely adding a name does not demonstrate a new capability. The unit must change a prediction or computation in a relevant setting. ## 4. Composition versus a parameterized schema One limited target is: T_G(a,b)(s) ≈ T_b(T_a(s)) A combined code preserves the effect of reading or executing a then b. A broader target is a schema with variable roles ξ: F̂_μ(α(s), ξ) ≈ α(F^(π_μ,ξ)(s)) The left side predicts through a concept; the right side represents what actually happens when its associated policy is applied. Conditions, termination, cost, and uncertainty are part of the comparison. The concept must work beyond its originating objects and arrangement. Otherwise it may be a memorized macro. ## 5. Representations as things to operate on Some learned operations act on a situation. Others change its description: M_μ: R → R′ Still others change the language itself: U_μ: L_t → L_(t+1) The input and effect of a meta-operation must be specified. Prefixing an ordinary description with “meta” is insufficient. Possible operations include forming an object from a relation, reopening a distinction, changing a reference frame, and comparing alternative descriptions. These remain candidate operations, not a complete taxonomy of thought. ## 6. Birth, merging, splitting, and incomplete concepts A system could propose a new unit, merge redundant units, split a concept when counterexamples reveal different cases, or narrow its validity conditions. Merging does not automatically release physical memory in every architecture. Reusing capacity can also break older readers or erase distinctions needed later. An incomplete concept can remain useful if unknown roles and conditions are explicit. Competing partial schemas may be preferable to premature certainty. ## 7. A problem-space that carries previous computation A rich space of relations, reusable transformations, and learned routes may allow a modest agent to perform better than the same agent operating on a poor representation. The phrase “a space more intelligent than the agent” can be tested by holding the agent and its budget fixed while changing the supplied structure. Success would concern generalization, error, and search cost. A visually attractive embedding is not sufficient. ## 8. Which direction approaches the goal fastest? The motivating geometric question is goal-directed: from which direction can the system approach its target most efficiently? Different targets imply different costs: execution time, action count, planning work, learning loss, or communication. A candidate discrete objective is: V*(s) = min_a [c(s,a) + E(V*(s′) | s,a)] V*(goal) = 0 This is a standard optimal-cost formulation used as a modeling option, not a recovered new theorem. An instantaneous steepest-descent direction, a geodesic of a selected metric, and a globally cheapest route through obstacles are different objects. Changing coordinates while correctly carrying the costs does not improve the physical optimum. A learned representation may help a bounded agent find it. Encoding ten actions in one symbol can save search without making execution cost one action. ## 9. Accounting for the language itself A candidate total cost includes building the language, encoding a task, planning, execution, task error, and loss of previously useful behavior. The comparison horizon and units matter. An expensive representation may pay for itself over many tasks, while being wasteful for one. The meaningful question is not only whether a representation is short, but whether its construction and use improve performance under a stated resource budget. ## 10. Ways to persist change Possible implementations include a temporary latent state, an external versioned concept library, a trainable internal module, a changing vocabulary or composition system, and a learned update rule. A low-rank adapter could carry changed behavior: W_(ℓ,t) = W_(ℓ,0) + B_(ℓ,t) A_(ℓ,t) This is one candidate implementation. It does not by itself define concept formation, reliable memory writing, or a complete evolving language. A new text label is also not automatically a new trained vocabulary token. ## 11. Puzzle environments as representation probes Two kinds of puzzle motivate different questions. In one, actions alter which future resources or combinations become available. A representation must capture that causal dependence rather than only the current visible pieces. In another, a movable object can also provide an interior environment, including recursive references. A reference graph may preserve identity better than an unqualified containment tree. These are proposed task families. No reconstructed move history or benchmark is asserted here. Useful candidate distinctions include object/environment dual roles, identity versus copy, containment versus screen coordinates, and validity conditions for entering or leaving a nested structure. ## 12. Human assistance that changes a representation A human might supply a relational distinction rather than the next move. The key question is whether that distinction improves later unaided performance. A candidate evaluation records: 1. The observed state and known rules. 2. A prediction made before action. 3. The action and verified outcome. 4. The human contribution and its scope. 5. The concept change. 6. A transfer attempt on a different arrangement and a counterexample. This records observable interventions and concise explanations. It does not claim complete access to a model's internal computation. ## 13. Tests that discriminate among explanations Compare raw versus relational representations with the same agent; memorized macros versus variable-role schemas; a prose description versus an explicit concept record; move hints versus representation hints; fixed versus revisable concepts; and external storage versus learned internal modules. Separate familiar structures in new appearances, new depths, new compositions, and cases that defeat an overgeneralized concept. Keep final tests out of concept selection. Success in one interaction is not evidence of persistent parameter learning. A human supplying every needed concept is not evidence of autonomous discovery. Both can still be useful if labeled correctly. ## 14. Cross-model meaning A compact representation requires a reader. Two equal-sized vectors need not have equal meaning. Transfer may require a common interpreter, learned translator, or synchronized library. Evaluate the resulting behavior, the compatibility of versions, and the cost of maintaining shared context. A short message can hide a large infrastructure cost. See [Memory, Tokens, and Learning Geometry](memory-tokens-and-learning-geometry.md). ## 15. Branches that remain independent Physical realizations, cosmological compression metaphors, social selection, distributed biological analogies, and material-production hypotheses can be investigated alongside this program. They are not necessary consequences of the semantic-language proposal. Similar words do not establish identical mechanisms. ## 16. Open questions What differentiates a useful new concept from a label? Which counterexample should split it? What detail must remain recoverable? When does changing the representation improve actual problem solving rather than a proxy metric? Can the concept survive a new task, a new model, or a new version of its reader? How do updates interact with caches, optimizer state, and continuing tasks? How much human work and prior infrastructure does the claimed improvement require? The central test is whether a new representational distinction makes a previously difficult inference more directly and reliably available.