Theorist Mode

Core Vocabulary

• Semantic projection — DFG, CFG, Scope etc. Each is an abstract interpretation of full program semantics
• Semantic role — cross-language operation class: Callable, Invocation, Declaration, Import, Assignment, Access, Control
• Flow rule — operational semantics for one AST node type in one projection: ConditionalExpression.DFG → consequent|alternate flows to parent
• Soundness — no false negatives. If there's a real dependency, the graph shows it
• Completeness — every relevant AST node type has a rule for every applicable projection
• Functor — mapping between abstraction levels (AST→Graph, Graph→DFG, Graph→Haskell types)
• Cognitive load — intrinsic (task complexity) + extraneous (tool friction) + germane (building mental model). Grafema reduces extraneous and pre-builds germane.

Grafema's Theoretical Identity

"Haskell for untyped code" — Grafema builds what Haskell's type system provides natively, but for languages where types don't exist.

The Key Number

Developers spend 58% of time on code comprehension. A tool that speeds this up by 30% saves 17% of total developer time. For 50 developers = 8.5 FTE.

Multi-Language Strategy

Best-in-class parser per language (NOT tree-sitter). AST = human understanding, CST = "code of code".

Order: Java first (simplest, reveals JS-coupling), then Kotlin → Swift → Obj-C.

Completeness Chain

@babel/types spec
  → generate semantic rules matrix (180 nodes × 7 projections)
    → generate visitors/edges from rules
      → graph is provably complete
        → LLM benchmark shows improvement
          → human study at ICPC/PPIG confirms

Discussion Guidelines

When in theorist mode:

1. Use formal vocabulary — "semantic projection" not "analysis type", "soundness" not "completeness-ish"
2. Reference the levels — "this is an L3 concern (projection design)" or "this is L5 (cognitive impact)"
3. Connect to evidence — cite the 58% comprehension number, NASA-TLX, Cousot & Cousot
4. Think in functors — "this transformation preserves/loses what properties?"
5. Check prior art — before proposing, check if Spoofax/CodeQL/Joern already solved it
6. Measure — every claim should have a measurable metric attached