T-4.1
Same Query → Same Output
Not yet tested
Hypothesis: Given a frozen graph state, requesting the same file twice produces bit-identical output. Pure graph traversal is deterministic.
Test: Snapshot graph. Project
src/cognitive/engine.ts twice via pure graph traversal (no LLM). Compare outputs byte-by-byte.
Pass: Outputs are byte-identical across 10 consecutive runs.
Fail: Any variance in output between runs (indicates hidden non-determinism in traversal or template).
T-4.2
LLM-Assisted Projection — Bounded Variance
Not yet tested
Hypothesis: When an LLM assists in projection, the output varies but stays within a measurable semantic bound. The structural skeleton (imports, exports, type signatures) is stable; only implementation bodies vary.
Test: Project the same file 5 times with LLM assistance. Parse each output's AST. Compare structural signatures (function names, parameter types, return types, import paths).
Pass: Structural signatures identical across all 5 runs. Body variance is cosmetic (variable names, formatting, comment wording).
Fail: Structural elements change between runs (different function signatures, missing exports, changed types).