Suite 5 — Projection Quality
T-5.1
Function-Level Projection
Not yet tested
Hypothesis: A single function can be projected from the graph with correct signature, parameter types, and a semantically reasonable body.
Test: Select 10 functions from DreamGraph's codebase. For each, project from graph data alone. Compare against actual implementation: signature match, type match, behavioral equivalence (unit tests pass).
Pass: ≥ 8/10 functions have correct signatures AND ≥ 6/10 pass their unit tests.
Fail: < 8/10 signature match OR < 6/10 pass tests.
T-5.2
Concept-Level Projection
Not yet tested
Hypothesis: An entire subsystem (multiple files) can be projected from a high-level concept description, maintaining internal consistency.
Test: Describe "the scheduling subsystem" conceptually. Project all related files. Check: all imports resolve, no circular dependencies introduced, type checker passes.
Pass: Projected subsystem compiles AND all cross-file references resolve.
Fail: Compilation errors or unresolved references between projected files.
Suite 6 — Latency Budgets
T-6.1
Graph Read Latency < 500ms
Not yet tested
Hypothesis: A pure graph read (no LLM) for a single file projection completes within 500ms for a project with ≤ 1000 entities.
Test: Populate graph with 1000 entities and realistic edges. Time 100 single-file projections. Measure p50, p95, p99.
Pass: p95 < 500ms.
Fail: p95 ≥ 500ms.
T-6.2
LLM-Assisted Projection < 3s
Not yet tested
Hypothesis: An LLM-assisted single-file projection completes within 3 seconds, including LLM round-trip.
Test: Time 20 LLM-assisted projections across varied file complexities. Measure p50, p95.
Pass: p95 < 3s.
Fail: p95 ≥ 3s.
T-6.3
Full Project Rebuild < 30s
Not yet tested
Hypothesis: A complete project projection (all files from graph) completes within 30 seconds for a project with ≤ 1000 entities.
Test: Full project build from graph (parallelized). Measure wall time for 1000-entity project.
Pass: Wall time < 30s.
Fail: Wall time ≥ 30s.