WhichJudge / decision
ns compares Jev with 4o-mini only
Jev 72.6%, Mini 72.2%, the 2026 model 73.6%, TF-IDF 69.4%: nothing separates them. Every model misses about a third of the offensive posts (Jev 35.6%, Mini 36.0%).
Replaces: Offensive-content classifier
Gold: cardiffnlp/tweet_eval
- tweet_eval offensive, train split (test exhausted by v1), n=500, seed=11
Labels: not_offensive / offensive
| Model | Accuracy | 95% CI | Latency (p50 unless stated) | Cost per 1M decisions |
|---|---|---|---|---|
| Jev jev-1.13.0 | 72.6% | [68.5% - 76.3%] | 710 ms | $15.08 |
| gpt-4o-mini-2024-07-18 | 72.2% | [68.1% - 75.9%] | 943 ms | $18.78 |
| gpt-5.4-mini-2026-03-17 | 73.6% | [69.6% - 77.3%] | 812 ms | unpriced (106.7 tok/call) |
| Laya (self-hosted, convaiinnovations/laya) | 60.2% | [55.8% - 64.4%] | 189 ms local cpu, no network | $0 API, compute not priced |
| gpt-6-luna (OpenAI Decisions API) | 68.4% | [64.2% - 72.3%] | 1191 ms, one client, region unset, includes network | $15.68 (list $0.10 per 1M input tokens, 2026-10-09) |
| @cf/cloudflare/clef (Cloudflare Workers AI) | 77.0% | [73.1% - 80.5%] | 850 ms, one client, region unset, includes network | $42.09 (list $0.24 per 1M input tokens, 2026-10-09) |
| @cf/cloudflare/clef-flash (Cloudflare Workers AI) | 75.2% | [71.2% - 78.8%] | 741 ms, one client, region unset, includes network | $15.78 (list $0.09 per 1M input tokens, 2026-10-09) |
| TF-IDF + LR (train n=11,414) | 69.4% | [65.2% - 73.3%] | 45 us mean per row, batched | $0 |
McNemar Jev vs Mini: ns, p=0.901 (33 Jev-only correct, 31 Mini-only correct). TF-IDF vs Jev: ns.
At >=0.9: 81.4% on 59.0% of posts, 8.8pt above ungated. ECE 0.151: threshold, do not quote.
Auto slice: at Jev confidence ≥ 0.9, 81.4% accurate on 59.0% of traffic, which is +8.8pt against 72.6% ungated. That threshold is the strongest one still covering at least half the traffic, picked by rule from the measured gates.
Calibration: ECE 0.151 (max bin gap 0.302). ECE says whether the confidence can be read as a probability. It does not say whether gating helps: thresholding only needs the model to rank its own answers, so a badly calibrated task can still gate well. High ECE costs you the right to quote the number, not the right to threshold on it.
Frozen samples 646f37b1e64f, schema c28eed9baf6a.
receipts/content_offensive.json here is the slim copy: per call,
the id, gold, prediction, confidence, latency and the SHA-256 of the request and response.
The full request and response bodies are in the repository at
results/receipts/content_offensive.json
(text replaced by its hash; scripts/rehydrate.py restores it).
git clone https://github.com/RadRebelSam/whichjudge.dev
python3 scripts/verify_run.py # recounts this table from the receipts, no API key
python3 scripts/run_eval.py # hits the APIs again with your own keysIndependent bench, not affiliated with TypeSafe AI. Public gold labels, n=500, not a procurement study. All 11 decisions · source and data · contact@radrebeldeveloper.com
build e6d1f683a4 · this row from run 1 of 1, 2026-09-21T06:23Z