rabbit-rewrites

Benchmarks and drives the model-backed rewrite path over an OpenAI-compatible endpoint. The rewrite engine itself (rwlib/endpoint.py, rwlib/rewrite.py) lives in skills/rabbit-writes/scripts/rwlib/ and is documented there. This skill's own scripts exercise that engine rather than owning it.

Commands

# Run the skill test suite (the model battery, and the bench that scores against it)
python3 skills/rabbit-rewrites/tests/run.py

# Benchmark a running endpoint against the battery
python3 skills/rabbit-rewrites/scripts/bench.py --model-endpoint http://127.0.0.1:8080/v1 --model-name qwen2.5-3b

# Repeat for a stabler pass rate, and save JSON
python3 skills/rabbit-rewrites/scripts/bench.py --repeat 3 --json > qwen2.5-3b.json

Structure

• scripts/bench.py: The model-benchmarking runner.
• scripts/battery.json: The passages scored in the benchmark.
• references/models.md: Recommended local models, download sources, and quantization notes.

Gotchas

• A battery case that raises no findings is invisible from reading the file. Three of the original twelve in battery.json did: one used technical-blog vocabulary the register correctly exempts, one used a hedge not in the lexicon, and one was 24 words under the 120-word floor that gates every stylometric finding (wc >= 120 in scan.py). The bench happily printed a pass rate over nine cases while reporting twelve. test_every_battery_case_actually_raises_a_finding is the guard, and test_every_battery_case_carries_a_fact_to_lose is the other half: a case with no number, path, or quotation in it can only ever measure fluency, never preservation.