The thesis.
Separan starts from a source-level problem: when a block’s purpose and boundary are implicit, a reader or editing tool has to reconstruct them from indentation or surrounding syntax. Separan instead names the boundary. A logic unit opens with SEP:name and closes with END_SEP:name; labelled control-flow blocks use the same matching-label principle. The parser reports a structural mismatch rather than treating a differently named closer as an equivalent boundary.
Name the boundary, then check that the edit stayed inside it.
The primary move is syntactic, not a claim that naming alone proves a program correct. Names remain readable to a person, while matching rules make selected boundaries mechanically checkable. That design is distinct from Tacit’s canonical-AST approach: Separan keeps ordinary source text and meaningful identifiers rather than making text a projection of an AST.
What it looks like.
SEP:main if user.active :active_user print "active" endif:active_user END_SEP:main
The opener and closer name the logic unit, and the conditional must close with the label used at its opening.
Separan’s structural tooling extends the same idea to review. It assigns hierarchical identities such as SEP:main#1/if:active_user#1 to AST boundaries and computes versioned SHA-256 fingerprints that exclude source positions, indentation, blank lines, and comments. The separan-structure diff command compares parsed structures; separan-structure verify checks that changes fall within explicitly allowed scopes. Semantic tags can also select separately located logic units for a scoped change.
Distinctive moves.
- Named, checked boundaries.
SEP/END_SEPand labelled control-flow closers expose the structure in source and reject mismatched labels. - AST identities for review. Hierarchical identities keep repeated labels addressable, while fingerprints distinguish structural changes from movement or formatting.
- Explicit edit scopes. A reviewer or CI job can permit changes beneath a selected boundary and fail when changes occur outside it. Ambiguous short labels require a more specific path rather than fuzzy matching.
Maturity.
Separan has Python and native C implementations, a command-line runtime, and a published language specification. Structural diff and edit-scope verification are implemented as a preview tooling layer. These checks constrain where a change occurs; they do not prove that the changed code has the intended behaviour. Tests, type checks, and other semantic checks remain necessary.
Agent tooling.
The project ships a VS Code extension and a language server, along with structural inspection, diff, and scope-verification commands. The review commands also offer JSON output for CI and review integrations. The structural AI workflow specification documents identities, fingerprints, scope rules, and exit statuses; the README provides the language overview and runnable examples.
- TacitsyntacticBoth make program structure central to authoring, but take opposite positions on names. Tacit treats the AST as authoritative and removes semantic weight from display names; Separan keeps readable names in source and makes matching opener/closer labels part of the checked structure.
- reqlanverificationBoth give names a machine-checkable role. reqlan resolves requirement-graph edges to files, symbols, and tests beside host-language code; Separan parses an executable language and uses named AST boundaries to identify and constrain source edits.
- IntentverificationBoth expose machine-checkable structure to support agent-authored code, but check different properties. Intent verifies contracts and their references; Separan's structural review tooling checks edit boundaries and does not establish behavioural correctness.