Dvader targets short report-style replies: status, review findings, proof lines, and next steps.
Dvader targets up to 80% fewer output tokens on short report-style replies.
That number must come from fixtures, not vibes. Some extreme replies may compress more. We do not market unrepeatable extremes.
The package includes rough and tokenizer-backed benchmark modes:
npm run bench:tokens
npm run bench:gpt
It compares normal report prose against Dvader-style reports in benchmarks/output-token-benchmark.json.
Current result:
| Metric | Normal | Dvader | Saved |
|---|---|---|---|
| Rough total | 685 | 130 | 81% |
| GPT tokenizer total | 775 | 212 | 73% |
The test is intentionally conservative:
--tokenizer=gpt through gpt-tokenizer.The public number should be repeatable. A one-word answer can beat 80%, but a useful answer often needs paths, commands, error lines, and proof. Those facts stay.
Shorter is good. Correct is mandatory.
Dvader adds words back for:
Next: FAQ.