Internal lab
Paste an AI-flavored draft. Each button runs a different humanise strategy
against the local API, in increasing cost order. Detector scores come from
desklib/ai-text-detector-v1.01 (DeBERTa-v3-large, MIT). The
adversarial loop also uses sentence-transformers/all-MiniLM-L6-v2
for the similarity-floor invariant.
Cost ladder — click to expand
| # | Method | Per-call cost | Latency | Notes |
|---|---|---|---|---|
| 0 | Detect (baseline) | $0 | ~150ms | desklib DeBERTa CPU |
| 1 | Postprocess only | $0 | ~50ms warm | 8 deterministic rules |
| 2 | Editorial + Postprocess | ~$0.001 | 1–10s | LLM cascade single-shot |
| 3 | Adversarial + Feedback + PP | ~$0.001×iters | 5–30s | Closed loop, detector-aware |
| 4 | Local paraphraser (T5) | $0 | 5–60s | humarin T5-base ~880MB or DIPPER 11B ~22GB |
| 5 | Compare all | ~$0.003 | 10–40s | Runs 1–3 in parallel via /benchmark |