graphify measured at 37.7x token savings, short of 71.5x

The test ran graphify against the author's own ai-100days repository, which held 22 days of experiment code across 90 code files. Building the graph with --code-only produced 914 nodes, 1,657 edges and 56 communities, and skipping 242 non-code files kept the process entirely offline. That confirms the tool's core promise of deterministic, local parsing without paid API keys or a GPU.

The gap between the benchmark and everyday use comes down to the baseline. The official benchmark compares against reading a 45,700-word corpus naively at roughly 60,933 tokens, yielding the 37.7x figure. When the author instead measured a specific question—where KronosPredictor is defined and used—the realistic approach of grepping and reading only the relevant lines cost about 1,799 tokens, versus about 242 for graphify explain, a roughly 7x difference. The author notes the case was relatively simple and that more tangled cross-references could widen the gap, but no measurement approached 71.5x.

Two caveats emerged during testing. The query command, marketed as natural-language search, returned irrelevant nodes for Japanese and failed to trace callers in English. The author also observed node counts shifting nondeterministically after running graphify update, recorded in the repository logs, and advises against treating exact node counts as a strict metric.