Factum — Object-Role Modeling (ORM 2)

factum — a thing done; a fact.

Draw a conceptual schema in ORM 2 notation, read it back as English sentences, check it for errors, and map it to a relational or property graph schema — without leaving the editor.

In the spirit of NORMA, but native to VS Code and storing models as plain JSON, so they diff and merge like the rest of your source.

Documentation and guide →

Why Object-Role Modeling

ORM describes a domain as elementary facts — Person works for Company — instead of tables or classes. Because those facts are attribute-free, every constraint is explicit and every model can be read back as plain sentences a domain expert can confirm or reject. You decide how it becomes tables or nodes afterwards, and the same model can become both.

Features

Draw the conceptual schema

Entity types and value types with reference modes (Person(.nr)), unary through n-ary fact types with multiple readings, subtyping, objectification (nesting) and derived fact types. Drag to move, marquee to multi-select, drag a role onto an object type to connect it, double-click to rename in place — typing Person(.nr) sets the name and reference mode at once.

Constraints are drawn in standard ORM 2 notation: internal and external uniqueness, preferred identifiers, simple and disjunctive mandatory, frequency, all ten ring types, subset, exclusion, equality, value and cardinality — each editable in the properties panel and each carrying an alethic/deontic modality.

Read the model back in English

A live FORML verbalization of the whole model. Click any sentence to select what it describes.

Each Person works for exactly one Company.
In each population of "Person has Skill", each Person, Skill combination occurs at most once.
It is necessary that the possible values of GenderCode are {'M', 'F'}.

Catch mistakes while you draw

Missing reference schemes, fact types without a uniqueness constraint, uniqueness constraints too narrow to keep a fact type elementary, unattached roles, reading/arity mismatches, subtype cycles and incompatible ring or set-comparison roles — reported in the Problems panel and highlighted on the diagram.

Map to a relational schema

Rmap-style mapping in the Relational tab, and SQL DDL for PostgreSQL, SQL Server, MySQL, SQLite or ANSI SQL. Functional fact types are absorbed as columns, compound-unique fact types get their own table, unaries become booleans, subtypes are absorbed into their supertype, and value constraints become CHECK constraints.

Map to a labeled property graph

The same model, mapped to a property graph for LadybugDB. ORM carries more than a hand-drawn graph model does, and the mapping uses it:

• entity types become node tables, keyed by their reference mode;
• lexical value types become properties, and are promoted to nodes only when played many-to-many, where a single-valued property could not hold them;
• binary fact types become relationship tables whose multiplicity is read off the uniqueness constraints — a constraint on one role makes that end the "one" end;
• an n-ary fact type cannot be an edge, so it is reified into a node linked to each role player: the Levi (bipartite) form of the hyperedge the fact type really is.

CREATENODETABLEPerson(nrINT64PRIMARYKEY, genderCodeSTRING);
CREATERELTABLEWORKS_FOR(FROMPersonTOCompany, MANY_ONE);
CREATERELTABLEHAS_STUDENT(FROMEnrolmentTOStudent, MANY_ONE);

LadybugDB enforces primary keys and multiplicities. Everything else ORM can state — mandatory roles, value ranges, ring and set-comparison constraints — is carried into the script as verbalized comments rather than quietly dropped:

// [mandatory] It is necessary that each Person works for some Company.

Exchange models with other fact-based tools

ORM: Import Model reads NORMA .orm XML, the FBM Exchange MetaModel .fbm, an Apache Ossie ontology or a Unified Modelling Schema document, and writes an .orm.json beside it — picking the reader from the file, since Ossie and UMS share .yaml. ORM: Export Model As writes any of the four back out.

Format	Import	Export	Fidelity
NORMA .orm	yes	yes	Conceptual both ways; diagram geometry is read but not written
FBM .fbm	yes	yes	Conceptual both ways; the closest to a lossless round trip
Apache Ossie ontology	yes	yes	Conceptual; objectification and the diagram have no counterpart
Unified Modelling Schema	yes	yes	Logical; export is faithful, import recovers the schema, not the model

Both directions report what they could not carry rather than dropping it silently, so an export to Ossie tells you it left your objectified fact type behind. See the format comparison for what each format holds and why Factum keeps one of its own.

Start from examples, not from a blank diagram

ORM: Derive Model from Example Data (CSV) reads a table and proposes a first-draft schema — the step both Halpin's design procedure and FCO-IM begin with. An identifying column becomes the entity's reference mode, types and enumerations are inferred from the values, every row is kept as a sample fact, and each assumption comes back as a note to confirm.

Nothing is proposed that the data does not support: a column distinct across three rows is not called unique, because at that size it is coincidence.

Check the model against its own examples

Fact types carry a sample population — real tuples, stored in the same file. The verbalizer substitutes them back into the readings, so a domain expert reads "101 works for Acme" rather than a placeholder, and the validator checks the constraints you drew against the examples you gave. A uniqueness constraint your own data contradicts is reported, not believed.

Export

SVG and PNG export of the diagram, plus force-directed auto-layout.

The command line

The model is a text file, so a build can check it. factum runs the same core the editor does.

factum validate model.orm.json --format github # annotate a CI run
factum verbalize model.orm.json --population # read it back, examples and all
factum diff before.orm.json after.orm.json # what the model now says
factum drift model.orm.json schema.sql # where the database disagrees
factum derive people.csv -o people.orm.json # a first draft from examples
factum ddl model.orm.json --dialect postgres
factum convert model.orm.json --to fbm

In a pull request

The bundled GitHub Action validates the model and comments with the sentences that changed — not "roles": ["r1"] appearing in a diff, but:

- Each Person works for at most one Company.+ Each Person works for exactly one Company.

- uses: Volland/factum-orm@v0with:
model: model/domain.orm.jsonbase: /tmp/base.orm.json # the same file from the base branchstrict: 'true'

A complete workflow is in examples/model-check.workflow.yml; copy it into .github/workflows/.

Schema drift

factum drift model.orm.json schema.sql compares the schema the model maps to against one that already exists — a pg_dump --schema-only, a migration, anything with CREATE TABLE in it — and prints the differences with the ALTER statements that would reconcile them. No database driver is needed, because it reads SQL rather than connecting.

For coding agents

factum-mcp is an MCP server over a model, so Claude Code, Copilot or any MCP client can read the conceptual schema rather than guessing at it from the tables:

{ "mcpServers": { "factum": { "command": "factum-mcp" } } }

It exposes read_model, verbalize_model, validate_model, generate_schema, diff_models, detect_drift, read_population and apply_model. Everything is read-only except the last, which validates first and refuses to write a model with blocking errors.

The skill pack

Ten slash commands and two skills, distilled from Fact-Based Agents, that put the modelling procedure itself in front of your agent — /orm-model runs Halpin's seven steps with you, /orm-review reads a model back and argues with it, /orm-fix works through validator codes, /orm-memory designs an agent's memory as a conceptual schema.

factum skills install # asks which agent, and whether globally or here
factum skills list # what is in the pack

Agent	Skills install to	Slash commands
Claude Code	~/.claude/skills or ./.claude/skills	yes
Cursor	~/.cursor/skills or ./.cursor/skills	yes
Codex CLI	~/.codex/skills or ./.codex/skills	skills only
OpenCode	~/.config/opencode/skills or ./.opencode/skills	skills only

--target <agent>, --global/--local and --dir <path> skip the question. Nothing already installed is overwritten unless you pass --force. Codex and OpenCode read the skills but not the pack's slash commands, so those are left out rather than written where nothing reads them; the same procedures are in the skills either way. The pack is plain markdown in agent-skills/, so it also works as a Claude Code plugin directory as it stands. What each command does is on the skill pack page.

Quick start

1. Run ORM: New ORM Model from the Command Palette and choose where to save it.
2. The diagram opens with a small example schema. Press E and click to add an entity type, 2 and click to add a binary fact type, then C and drag from a role box to an object type to connect them.
3. Click role boxes to select them, then use the constraint buttons in the toolbar (U uniqueness, P preferred identifier, ● mandatory, and so on).
4. Watch the Verbalization tab to check the model says what you meant.

Keyboard and mouse

Action	Binding
Select tool / entity / value type	V / E / T
Unary / binary / ternary fact type	1 / 2 / 3
Subtype link / connect role	S / C
Zoom to fit	F
Auto-layout	Ctrl/Cmd+Alt+L
Delete selection	Delete
Undo / redo	Ctrl/Cmd+Z / Ctrl/Cmd+Shift+Z
Pan / zoom	Middle-drag or drag empty canvas / wheel
Rename in place	Double-click a shape
Add to selection	Shift-click

Commands

Command	Description
ORM: New ORM Model	Create a starter .orm.json and open the diagram
ORM: Import NORMA (.orm) File	Convert a NORMA ORM 2 XML file
ORM: Import Model (NORMA, FBM, Ossie, UMS)	Convert any supported interchange document
ORM: Export Model As (NORMA, FBM, Ossie, UMS)	Write the model out in an interchange format
ORM: Derive Model from Example Data (CSV)	Propose a first-draft model from a table of examples
ORM: Verbalize Model	Full FORML verbalization as Markdown
ORM: Show Relational Mapping	Mapped tables as a Markdown table
ORM: Generate Relational Schema (SQL DDL)	SQL DDL in a new editor
ORM: Generate Property Graph Schema (LadybugDB)	LadybugDB Cypher DDL in a new editor
ORM: Export Diagram as SVG / as PNG	Save the diagram as an image
ORM: Auto-Layout Diagram	Re-arrange the diagram
ORM: Open Model Source (JSON)	Open the model as text

Settings

Setting	Default	Description
orm.ddl.dialect	postgres	SQL dialect for generated DDL
orm.ddl.quoteIdentifiers	false	Always quote generated identifiers
orm.graph.subtypeStrategy	nodeTable	nodeTable (own label + IS_A) or absorb
orm.graph.ifNotExists	false	Emit IF NOT EXISTS in the generated graph DDL
orm.validation.enabled	true	Report model problems in the Problems panel
orm.verbalization.mode	forml	forml or plainer English
orm.diagram.snapToGrid	true	Snap shapes to the grid
orm.diagram.gridSize	10	Grid size in pixels
orm.diagram.showGrid	true	Show the grid

File format

*.orm.json holds the conceptual schema and its diagram layout in one id-addressed structure, described by a published JSON Schema that VS Code uses to validate and complete the file in the plain text editor:

{
"$schema": "https://www.factum-orm.com/schema/orm-model-2.schema.json",
"version": 2,
"name": "HR",
"objectTypes": [
{ "id": "ot_person", "name": "Person", "kind": "entity", "refMode": "nr", "dataType": "integer" }
],
"factTypes": [
{
"id": "ft_works",
"roles": [
{ "id": "r1", "objectTypeId": "ot_person" },
{ "id": "r2", "objectTypeId": "ot_company" }
],
"readings": [
{ "id": "rd1", "roleOrder": ["r1", "r2"], "text": "{0} works for {1}", "isPrimary": true }
]
}
],
"subtypeRelations": [],
"constraints": [
{ "id": "uc1", "kind": "uniqueness", "roles": ["r1"] },
{ "id": "mc1", "kind": "mandatory", "roles": ["r1"] }
],
"diagram": { "shapes": { "ot_person": { "x": 120, "y": 200 } } }
}

Readings use {0}, {1}, … placeholders indexing into roleOrder, the same convention NORMA uses. The file is safe to hand-edit; the editor repairs missing collections on load. Because the diagram is a custom editor over a text document, VS Code's dirty state, undo stack and file watching all work normally — and ORM: Open Model Source (JSON) shows the text behind the picture at any time.

Metadata

Every element may carry a meta object — guid, uri, title, shortDescription, description, synonyms, tags, aiContext and source. None of it changes what the conceptual schema means. It is there so a model survives a round trip through another fact-based modeling tool, and so generators have something to say beyond an element's name: meta.description becomes the comment on the generated table and node table.

The fields have counterparts in the formats Factum has to interoperate with rather than being invented — guid and the two descriptions are the FBM Exchange MetaModel's GUID, ShortDescription and LongDescription; synonyms is its Synonyms and Apache Ossie's ai_context.synonyms; aiContext is Ossie's ai_context.

Schema generation hints

A hints object steers generation per target. A hint never changes the conceptual schema, only how that schema is rendered — strip every hint from a file and the model still says the same thing. A name given in a hint is a physical name and is used verbatim.

{
"id": "ot_person", "name": "Person", "kind": "entity", "refMode": "nr",
"meta": { "description": "A human being known to the business." },
"hints": {
"relational": { "tableName": "HR_PERSON", "columnName": "PERSON_NR" },
"graph": { "label": "Employee", "labels": ["Party"] }
}
}

Hint	Applies to	Effect
relational.schemaName	model	Qualifies every generated table
relational.tableName	object type, fact type	Physical table name
relational.columnName	value type, role	Physical column name
relational.sqlType	value type	SQL type, instead of the one derived from dataType
relational.mapping	fact type	absorb or separateTable
graph.label	object type, fact type	Node label or relationship type
graph.labels	object type	Additional node labels
graph.propertyName	value type, role	Name of an absorbed property
graph.mapping	value type	node or property

A hint that would make the generated schema lose facts is refused rather than obeyed, and the refusal appears in the mapping notes: asking for graph.mapping: "property" on a value type played many-to-many leaves it a node, because a single-valued property cannot hold many values.

Unknown target keys are legal and preserved, so another tool can carry hints.ossie or hints.typedb without a change to this format.

Extensions and versioning

Any key beginning with x- is an extension: legal anywhere, ignored by the editor and written back unchanged, following the OpenAPI convention. The loader is deliberately more permissive than the schema — it preserves every unrecognised top-level key, so a misspelled key is a warning in the editor rather than data lost on the next save.

version is the format's major version, currently 2. Version 2 only adds optional meta, hints and lang keys, so a version 1 file is a valid version 2 file: it is upgraded on load and written back as version 2.

Mapping rules

The relational mapping follows Rmap:

• a fact type whose only uniqueness constraint spans every role becomes its own table with a composite key;
• an n:1 or 1:1 fact type is absorbed as a column (or columns) into the table of the object type playing the uniquely-constrained role, on the mandatory side for 1:1;
• a unary fact type becomes a boolean column;
• mandatory roles produce NOT NULL, optional ones NULL;
• an entity type's reference mode expands into its identifying column; compound identifiers expand recursively, and an object type with no reference scheme gets a surrogate key plus a mapping note;
• objectified fact types map to their own table keyed by the objectified roles;
• subtypes are absorbed into their supertype's table with optional columns, and an exclusive subtype partition adds a discriminator column with a CHECK.

The property graph mapping starts from the same schema but answers a different question:

• an entity type becomes a node table, keyed by its reference mode, by a single-role preferred identifier over a value type, or — failing both — by a generated SERIAL key;
• a value type stays a property unless it is played many-to-many or in an n-ary fact type;
• a binary fact type becomes a relationship table with MANY_ONE, ONE_MANY, ONE_ONE, or MANY_MANY when only a spanning uniqueness constraint applies;
• a unary fact type becomes a BOOLEAN property;
• an n-ary or objectified fact type is reified into a node with one MANY_ONE relationship per role; role links to the same player share one relationship table with several FROM ... TO ... pairs. Objectify a fact type to name its node yourself;
• subtypes get their own label joined by IS_A, or are absorbed when orm.graph.subtypeStrategy is absorb.

Mapping notes explaining each choice appear beside the schema and as comments in the generated script.

Known limitations

• The NORMA exporter does not write diagram geometry: NORMA's diagram section carries shape state well beyond position, and a partial one is worse than none.
• Sample populations, multi-page diagrams and join paths on set-comparison constraints are read from FBM but not modelled, so they do not survive a round trip through Factum.
• Apache Ossie's ontology_mappings — the binding from concepts down to dataset fields — is neither read nor written. It is the natural home for a hints.ossie target.
• Derivation rules are stored and verbalized but not evaluated.
• Drift detection reads SQL text rather than connecting to a database, and understands CREATE TABLE only — indexes, views and triggers are ignored.
• Model derivation makes every column a binary fact type about one entity type. Splitting the columns that are really about something else is still the modeller's job.
• The graph mapping targets LadybugDB's Cypher DDL. Other property graph databases will need small syntax adjustments.

Documentation

The site includes a mini book on ORM 2 — ten short chapters covering elementary facts, the constraint family, subtyping and objectification, and Halpin's seven-step design procedure worked end to end. Every figure in it is rendered by the extension's own renderer and links to the model file behind it, so any example can be opened and taken apart in the editor. Regenerate the figures with npm run figures.

The full documentation site lives in docs/ and is published with GitHub Pages: book, getting started, mapping rules, reference and the file format.

Development

npm install
npm run compile # bundle the extension and webview into out/
npm run watch # rebuild on change
npm test# model, verbalization, validation, mapping and rendering tests
npm run typecheck
npm run vsix # build a .vsix package

Press F5 to launch an Extension Development Host on the examples/ folder.

The extension host code lives in src/extension.ts and src/editor/; the diagram runs in a webview (src/webview/). Both share the model, verbalizer, validator and mappers in src/model/ and src/core/, which are plain TypeScript with no VS Code or DOM dependencies — which is why they are straightforward to test and to reuse outside the editor.

Credits

Object-Role Modeling is the work of Terry Halpin and the fact-oriented modeling community. Factum is an independent implementation and is not affiliated with the ORM Foundation or the NORMA project.

License

MIT.