0007 — Authoring is a schema-layer concern, not a model-layer one

• Status: Accepted
• Date: 2026-08-26
• Deciders: openbimrs contributors
• Supersedes: —

Context

Model::push(Entity) is the only construction primitive. It takes a type name and a positional Vec<Value>, so an application that generates IFC must know that IfcAnnotation has seven slots, that slot 0 is a 22-character GlobalId and slot 2 is the name. Nothing checks any of it. The resulting file parses back happily in this library and is rejected by other tools.

The obvious fix — "add typed setters to Model" — is not available. ADR 0001 makes the entity graph free of domain semantics, and the dependency tiers place ifc-model at L0 with ifc-schema above it at L1. A schema-checked builder inside ifc-model would invert that edge and make the schema tables a mandatory dependency of every consumer, including ones that only stream bytes.

The second obvious fix — generate a builder struct per entity type — is what ADR 0002 and the ifc-schema design already rejected for readers: 776 types in IFC4, 876 in IFC4x3 with renames, and a recompile to support a new schema release.

Decision

Authoring lives in a new L2 crate, ifc-author, which depends on ifc-model and ifc-schema and is reached through the facade's author feature.

It is schema-driven, not generated. A builder names attributes; the crate resolves each name to its STEP slot via Schema::attributes, which already returns inherited-first positional order. Arity, unknown attribute names, required-but-missing slots, and declared-type mismatches are checked against the same tables the readers use.

ifc-model gains no schema dependency and no typed setters. Its mutation module remains the owner of schema-agnostic edits.

Consequences

• A new schema release is a new .exp file, not a code generation run.
• Authoring is optional: openbim-ifc without author compiles no builder code.
• Validation is construction-time, so an invalid entity is never inserted — distinct from ifc-validate, which audits a model that already exists.
• The builder can only be as good as the schema tables. A type the schema does not declare cannot be checked; the crate reports that explicitly rather than silently accepting the entity.
• Applications that genuinely need to write an entity the schema does not know keep using Model::push, which stays public and unchecked.

Alternatives considered

Option	Why not
Typed setters on Model	Inverts the L0/L1 dependency; forces schema tables on every consumer
Generated struct per entity	776–876 types per schema version, recompile per release; the problem ifc-schema exists to avoid
Builder inside ifc-validate	Conflates auditing an existing model with constructing a new one; different failure timing
No checking, document the slots	The status quo; every application reinvents an untested private version