This folder contains ordered, human-readable benchmark reports. Each category has 10 real snapshots under cases/, and every case links to its full raw input and full compacted output.
The current corpus is generated by scripts/benchmark/update-real-samples.sh. It uses an adjacent OpenHuman checkout, live OpenHuman Docker logs when available, public RSS/page snapshots when reachable, and checked-in OpenHuman artifacts as fallbacks.
Percentages are token reduction: higher is better (90% means the output shrank to a tenth of its size; 0% means it passed through untouched). Applied counts the cases where compression actually fired — the rest pass through because they are too small or a shape the compressor declines.
| Category | Cases | Applied | Token reduction (mean) | Avg latency | Report |
|---|---|---|---|---|---|
| Service And Docker Logs | 10 | 10 | 86.2% | 1.360 ms | report |
| Polyglot Source And XML | 6 | 6 | 77.6% | 0.447 ms | report |
| HTML, RSS, And Page Snapshots | 10 | 10 | 77.0% | 0.184 ms | report |
| Unified Diffs | 10 | 10 | 69.1% | 0.274 ms | report |
| GitHub Log Files | 33 | 26 | 60.9% | 4.805 ms | report |
| JSON SmartCrusher | 10 | 4 | 35.3% | 1.337 ms | report |
| Search Results | 10 | 10 | 31.7% | 0.905 ms | report |
| GitHub Source Files | 47 | 43 | 31.6% | 0.633 ms | report |
| Rust Source | 10 | 7 | 27.2% | 0.837 ms | report |
| Test Failure Logs | 10 | 10 | 15.3% | 0.086 ms | report |
| Plain Text | 10 | 0 | 0.0% | 0.000 ms | report |