dvader

The Math

Dvader targets short report-style replies: status, review findings, proof lines, and next steps.

Public Claim

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.

Current Fixture Check

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:

Why 80%, Not More

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.

When Compression Stops

Dvader adds words back for:

Next: FAQ.