-74.0%

Measurement methodology (version 2): **ashlr__read**: For each sampled source file, we measure raw file bytes and token count (chars/4 heuristic). We then apply the same snipCompact transformation used at runtime — wrapping the content in a tool_result message and calling snipCompact() — and measure the resulting byte/token count. The ratio is ashlrTokens / rawTokens. Files below 2 KB are excluded because snipCompact only fires on tool results > 2 000 chars; savings are zero by design for small files. Files are selected deterministically: the repo HEAD commit SHA is folded into a 32-bit seed (mulberry32 PRNG), then up to 4 files are sampled from each of four size buckets (2–5 KB, 5–15 KB, 15–50 KB, 50+ KB). Re-running on the same commit always picks the same files. **ashlr__grep**: Five common patterns (function, import, TODO, class, interface) are run via rg --json against the repo root. Raw output bytes are measured directly. The ashlr__grep fallback path (no genome) truncates output to 4 000 chars (head 2 000 + tail 1 000). The ratio is truncated/raw. Note: when a .ashlrcode/genome/ index is present, real-world grep savings are substantially higher. This benchmark measures only the conservative no-genome baseline. **ashlr__edit**: Three synthetic edits (small ~15 chars, medium ~300 chars, large ~3 000 chars) compare the naive "ship before+after as text" approach against ashlr__edit's diff-summary format (one header line + removed/added first-lines). The ratio is summary tokens / naive tokens. **Aggregation**: per-tool mean/p50/p90 are computed over each tool's ratio values (lower ratio = more savings). The `overall.mean` is pooled across every individual sample regardless of tool — so tools with more samples (read has 15, grep has 5, edit has 3) weight the overall figure proportionally. That makes the headline number reflect the workload mix, not a uniform per-tool average. The unweighted mean of per-tool means is intentionally NOT published because it gives equal weight to a 3-sample tool and a 15-sample tool, which over-weights the synthetic edit overhead. Token counts use the chars/4 heuristic, the same estimator the plugin uses at runtime for savings accounting.