WhichJudge / decision
ns compares Jev with 4o-mini only
Tied: Jev 73.2%, Mini 71.4%, the 2026 model 71.0% (Holm p=1). Neutral is the hard class for all three; Jev catches 53.9% of it. TF-IDF trails at 60.2%.
Replaces: Negative / neutral / positive on tweets
Gold: cardiffnlp/tweet_eval
- tweet_eval sentiment, n=500, seed=11
Labels: negative / neutral / positive
| Model | Accuracy | 95% CI | Latency (p50 unless stated) | Cost per 1M decisions |
|---|---|---|---|---|
| Jev jev-1.13.0 | 73.2% | [69.2% - 76.9%] | 728 ms | $15.47 |
| gpt-4o-mini-2024-07-18 | 71.4% | [67.3% - 75.2%] | 969 ms | $19.37 |
| gpt-5.4-mini-2026-03-17 | 71.0% | [66.9% - 74.8%] | 856 ms | unpriced (113.2 tok/call) |
| Laya (self-hosted, convaiinnovations/laya) | 56.6% | [52.2% - 60.9%] | 193 ms local cpu, no network | $0 API, compute not priced |
| gpt-6-luna (OpenAI Decisions API) | 69.8% | [65.6% - 73.7%] | 966 ms, one client, region unset, includes network | $16.56 (list $0.10 per 1M input tokens, 2026-10-09) |
| @cf/cloudflare/clef (Cloudflare Workers AI) | 63.2% | [58.9% - 67.3%] | 1074 ms, one client, region unset, includes network | $45.66 (list $0.24 per 1M input tokens, 2026-10-09) |
| @cf/cloudflare/clef-flash (Cloudflare Workers AI) | 59.6% | [55.2% - 63.8%] | 629 ms, one client, region unset, includes network | $17.12 (list $0.09 per 1M input tokens, 2026-10-09) |
| TF-IDF + LR (train n=45,615) | 60.2% | [55.8% - 64.4%] | 27 us mean per row, batched | $0 |
McNemar Jev vs Mini: ns, p=0.349 (41 Jev-only correct, 32 Mini-only correct). TF-IDF vs Jev: jev.
At >=0.9: 79.8% on 61.4% of messages, 6.6pt above ungated. ECE 0.157 is too high to quote the number; thresholding still works.
Auto slice: at Jev confidence ≥ 0.9, 79.8% accurate on 61.4% of traffic, which is +6.6pt against 73.2% ungated. That threshold is the strongest one still covering at least half the traffic, picked by rule from the measured gates.
Calibration: ECE 0.157 (max bin gap 0.181). 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 96a5cf86addc, schema 43230c9f5f38.
receipts/tweet_sentiment.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/tweet_sentiment.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