| # HWAI AI-detection ensemble — reproducibility benchmark |
|
|
| > **Thesis.** Closed commercial benchmarks (Originality.ai, Winston AI, GPTZero |
| > Premium) publish AUROC numbers on proprietary corpora you cannot audit. |
| > HWAI publishes its **exact smoke corpus**, **exact evaluation script**, and |
| > **exact calibration files** so anyone can reproduce our numbers in ≤10 min. |
|
|
| ## What you need to reproduce |
|
|
| 1. A running instance of `ml-services-hwai` (Hetzner CX43 ~$25/mo, or local Docker). |
| 2. Our calibration file (`/opt/ml-services/calibration.json`, shipped in v1.10). |
| 3. The eval script: `services/ml-services-hwai/scripts/eval_ensemble_corpus.py`. |
| 4. The hand-curated corpus: **embedded in the eval script as `CORPUS`** (44 texts, EN + RU, human + AI). |
|
|
| All four above are open in this repo. |
|
|
| ## Reference numbers (v1.10, 2026-04-24, 44-text hand-curated OOD) |
|
|
| | Lang | Ensemble AUROC | Brier | Detectors | |
| |---|---|---|---| |
| | EN | **0.770** (target: ≥0.80 post-weight-tuning) | ~0.18 | ai_detect + radar + binoculars + desklib | |
| | RU | **0.837** | ~0.12 | ai_detect + radar + binoculars | |
|
|
| Per-detector numbers: see `notes/contentos_ensemble_weights_tuning_2026-04-24.md`. |
|
|
| ## How to reproduce |
|
|
| ```bash |
| # On any host with Python 3.11+ and access to ml-services: |
| export ML_SERVICES_URL="http://<your-host>:3300" |
| export ML_SERVICES_API_KEY="<your-key>" |
| |
| python3 services/ml-services-hwai/scripts/eval_ensemble_corpus.py |
| # Output: /tmp/eval_ensemble_<timestamp>.json + .md |
| ``` |
|
|
| The output includes: |
| - Per-detector raw + calibrated scores, per sample. |
| - Per-lang AUROC + Brier (ensemble + individual detectors). |
| - Confusion-matrix style pass/fail per sample. |
|
|
| ## Why 44 texts? |
|
|
| Smoke battery sized for fast iteration (< 2 min wall-time), **hand-picked to |
| cover the failure modes** we've seen in production: |
|
|
| - **EN human formal** (press release, court filing, product manual) — biggest |
| false-positive risk for models that learned "AI = formal"). |
| - **RU human news/journalism** — known FP mode for RADAR-Vicuna. |
| - **EN + RU AI 2026-era** (Claude-4, Gemini 2.5, GPT-4o style) — the |
| distribution shift that breaks 2022-era-trained detectors. |
| - **Edge cases** (interview transcripts, casual parent notes, product reviews) |
| where detectors commonly overreach. |
|
|
| Full statistical-grade numbers on `n=750` OOD calibration split: |
| `services/ml-services-hwai/corpus/cal_test.jsonl`. Run: |
|
|
| ```bash |
| python3 services/ml-services-hwai/scripts/eval_ensemble_corpus.py \ |
| --from-cal-test services/ml-services-hwai/corpus/cal_test.jsonl --parallel 4 |
| ``` |
|
|
| ## Compare with external detectors |
|
|
| Free-tier cross-check (Sapling AI, 50 req/day): |
|
|
| ```bash |
| export SAPLING_API_KEY="..." |
| python3 services/ml-services-hwai/scripts/bench_competitors.py \ |
| --out /tmp/bench_competitors.json |
| ``` |
|
|
| Produces side-by-side AUROC on identical corpus. |
|
|
| ## What this benchmark does NOT claim |
|
|
| - It is **not** a universal ranking of AI detectors. Different corpora + |
| different objectives (academic integrity vs SEO vs marketing QA) produce |
| different rankings. |
| - It is **not** statistically powered for `p<0.05` differences of `<0.02` AUROC. |
| For that, use `cal_test.jsonl` (n=750). |
| - It is **not** stable under adversarial paraphrase attack (no detector is). |
|
|
| ## What it DOES claim |
|
|
| - Full reproducibility: bit-identical inputs → bit-identical scores. |
| - Calibration stability: pinned baseline via `test_calibration_regression.py` |
| auto-runs on every cal swap; rollback on drop >0.05 AUROC. |
| - Honest distribution reporting: the corpus is public so you can audit |
| whether it matches your use-case. |
|
|
| ## Cost to run full reproducibility |
|
|
| | Step | Cost | |
| |---|---| |
| | Spin up ml-services (Hetzner CX43 monthly) | $25 | |
| | Run smoke eval | $0 | |
| | Run `cal_test.jsonl` n=750 statistical eval | $0 (self-hosted) | |
| | Sapling cross-check (free tier) | $0 | |
| | **Total** | **$25 one-time hosting** | |
|
|
| Commercial equivalents: |
| - Originality.ai: $15 trial + per-call costs, no AUROC audit, closed corpus. |
| - GPTZero Premium: $15/mo, closed corpus. |
| - Winston AI: $29/mo, closed corpus. |
|
|
| ## Cross-references |
|
|
| - Corpus balance analysis: `notes/contentos_corpus_balance_analysis_2026-04-24.md` |
| - Ensemble weights tuning: `notes/contentos_ensemble_weights_tuning_2026-04-24.md` |
| - Fork #2 v1 → v2 progression: `notes/contentos_fork2_en_selfgen_spec.md` |
| - Regression test pins: `services/ml-services-hwai/tests/test_calibration_regression.py` |
| - 7-blocks gap assessment: `notes/contentos_7blocks_gap_assessment_2026-04-24.md` |
|
|
