Before Forms Can Meet: Typed Interfaces for Creative Systems
08 November 2026
Technical design note

Abstract
A creative system can connect images, physical models, symbolic motifs and human judgments without pretending they are the same kind of object. CalcuArt v2.3 proposes ArtIR, a typed intermediate representation whose fields retain those differences. This note considers the interface as the architectural unit: a contract describing what may cross a boundary, what must be converted and what cannot meaningfully be combined. It concerns representation and dimensional consistency, not the previously examined realism slider or archive-recovery theorem. [1, pp. 1–4]

A connection is not yet a composition
CalcuArt replaces an ambiguous numerical product with a record containing SourceArchive, Context, WorldIR, SceneCommit, DepictionIR, MaterialProcessIR, PatternIR, TemporalIR, EvidenceIR and EvaluationIR. Its composition symbol denotes constrained assembly of compatible records, not scalar multiplication or a tensor-product theorem. This prevents a visually impressive equation from silently assigning arithmetic meaning to incompatible information. [1, p. 1, §1]

Consider a garment rendered inside an architectural scene. Its surface may have thickness in meters, a material roughness parameter, a symbolic pattern and a viewer's preference rating. Each can influence the design, but adding them directly is undefined. An operator needs a domain, output type, units, preconditions and limitations. Combining outputs requires an explicit adapter into compatible fields or a common vector space. The source states that giving anatomy, psychology and perspective symbols does not make their sum meaningful. [1, pp. 1–2]

The adapter carries an argument
Suppose a design objective uses a measured feature f_j. CalcuArt specifies normalization y_j=(f_j−f_j,ref)/s_j, with s_j positive and carrying the same units as f_j. The resulting coordinate is dimensionless. Choosing the reference and scale is nevertheless a modeling decision: changing them changes the optimization problem. A shared numerical range does not prove that two features represent the same construct. [1, p. 1, §0]

This is where interface documentation becomes scientific work. The adapter should identify its source quantity, transformation, valid range and uncertainty. A missing thickness measurement cannot be entered as zero merely to fill a table. Nor should a saliency model's prediction be labeled observed attention. Such substitutions may preserve a software schema while invalidating the claim that the schema was meant to support.

Two materials, two responsibilities
The framework separates an object's physical material from the medium used to depict it. A painted rendering can change its brush response without changing the building's concrete or the garment's fabric. SceneCommit holds the authoritative world state; DepictionIR controls the permitted visual realization. If a physical material, camera, pose or event changes, the source requires an explicit new scene or version. [1, p. 2]

MaterialProcessIR goes further by asking how a surface was made. Its proposed fields distinguish support, tools, ordered operations, deposition, drying and abrasion. A digital brush may imitate a visible effect without simulating fluid chemistry. The document labels process equations as model interfaces requiring constitutive laws, boundary data and calibration before they become executable physical models. A texture that looks weathered is not evidence of a verified weathering process. [1, pp. 3–4, §3.1]

A small implementation contract
A practical prototype could define three adapters: physical width in meters to raster width under a declared projection; motif identity to a permitted ornament placement; and an attention prediction to a clearly marked evaluation field. Each adapter would reject unavailable prerequisites rather than invent defaults. Unit tests would include a missing scale, a changed camera and an unsupported motif topology. No such prototype was executed for this note.

The evaluation should test both successful and rejected connections. A system that accepts every input is not necessarily more capable; it may simply have erased its type boundaries. Conversely, a rejected combination can be useful when it identifies the missing evidence or conversion needed to proceed. The source's verification ledger allows pass, fail, unknown and not applicable rather than collapsing every state into one score. [1, pp. 10–11, §8]

Conclusion
Typed interfaces make a broad creative architecture more inspectable by preserving differences at its joints. They do not automatically supply cultural knowledge, physical accuracy or artistic judgment. Their contribution is narrower and valuable: a reviewer can see which relationship was declared, which conversion was performed and where evidence stops. That is a stronger basis for integration than treating all creative variables as interchangeable numbers.

Reference
[1] CalcuArt v2.3 formal calculus, Calculus Only, 7 November 2026, §§0–1, §3.1 and §8, original PDF pp. 1–4 and 10–11. Source: CalcuArt_v2.3_Calculus_Only_2026-11-07.pdf. Interface examples in this note are proposed applications, not reported implementation results.
