Decision model
Adjudex 9B v5
1 question per request
- Accuracy, all 2,198 questions
- 53.1%51.0% to 55.2%, 95% interval; 1,168 right
- Answers kept at 99% confidence
- 132 of 2,1982 wrong; share right 94.6% to 99.6%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.026$26.30Basis and price under Cost
- Latency
- 0.038 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- Perplexity Decider server, self-hosted
- Paid as
- Our own hardware
- Model ID
jev-latest- Variant
- local
- Parameters
- 9B dense, as the maker states it (counted from the released files: 8.39B).language model + 255-way decision head (readout.safetensors); source: counted from the checkpoint's safetensors, newsbang/adjudex-9b-v5@ffdcb82, huggingface.co/newsbang/adjudex-9b-v5; read 15 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- newsbang's released Adjudex 9B v5 (newsbang/adjudex-9b-v5 at ffdcb82), a fine-tune of Qwen3.5-9B with a 255-way decision head, run on our hardware through its card's autojev server (Perplexity Decider v1.1's source, unchanged), which answers only to the model name 'jev-latest', text only.
- Confidence from
- Option score
- Run
- 1183, 15 November 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
815031979442- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 53.1% | 51.0% to 55.2% | 1,168 of 2,198 |
| Reading | 80.8% | 77.6% to 83.6% | 530 of 656 |
| Grammar | 41.4% | 38.9% to 43.9% | 638 of 1,542 |
Comparison group: YZαβγ. Compare with another setup.
Accuracy by topic
English
The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.
- Right answers
- 469 of 547, first English run (1186)
- Accuracy
- 85.7%
- Non-hallucination with “I don't know”
- 0.0%: 0 “I don't know” of 77 not right (run 1188)
- Answered with “I don't know”
- 100.0%: 547 of 547
- Error rate with “I don't know”
- 14.1%: 77 of 547 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.02524.93from GPU time
- Latency, median request, first English run
- 0.034 s; 1 request at once
- English run
- 1186
- Machine, English runs
- RTX PRO 6000
- Date
- 15 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 132, of which 2 wrong (share right: 94.6% to 99.6%, 95% interval); 2,066 skipped, the model less than 99% sure
- Calibration gap
- 9.2 points
- AUROC, all questions
- 0.79 0.77 to 0.80
- AUROC, reading
- 0.85 0.81 to 0.88
- AUROC, grammar
- 0.69 0.66 to 0.72
- Brier score
- 0.195
- Accuracy of its surest answers
- most confident fifth: 94.3% right; most confident half: 74.1% right; most confident four fifths: 60.1% right
- Typical score of a right answer
- 80.7%
- Typical score of a wrong answer
- 44.3%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.038 s (this run)
- Latency, 90th percentile
- 0.045 s
- Requests at once
- 1
- Throughput
- 1577.7 questions per minute of request time
- Request time
- 1.4 minutes, over the requests whose answers were kept
Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.
Cost
- Cost per 1,000 questionsper 1M questions
- $0.026$26.30
- Whole run
- $0.058
- Basis
- Estimated from GPU time: the run's request time x the GPUs its server held x Runpod's on-demand price for an RTX PRO 6000, $2.49 per GPU-hour (read 5 November 2026). A cloud-equivalent estimate, not measured spending.
- Price
- $2.49 per GPU-hour, 1 GPU, for the run's request time (source, 2026-11-05)
- Input tokens per question
- 388
- Output tokens per question
- 0
Going first in front of an LLM
Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.
| LLM behind it | Share taken | Accuracy change | Saving |
|---|---|---|---|
| Gemini 3.8 Flash high · chat | 5.4% | −0.08 points | 4% |
| GPT-6 Astra high · chat | 5.4% | −0.08 points | 5% |
| GPT-6.1 Sol high · chat | 5.4% | −0.08 points | 4% |
| GPT-6 Astra low · chat | 5.4% | −0.08 points | 5% |
| GPT-6 Astra low · chat · 1 per request | 5.4% | −0.08 points | 5% |
| Gemini 3.8 Flash low · chat · 1 per request | 5.4% | −0.08 points | 3% |
| GPT-6.1 Sol low · chat · 1 per request | 5.5% | −0.08 points | 5% |
| GPT-6.1 Sol low · chat | 5.4% | −0.08 points | 3% |
| Claude Opus 5.5 high · chat | 5.4% | −0.08 points | 5% |
| GPT-6.1 Terra high · chat | 5.5% | −0.08 points | 0% |
| Claude Opus 5.5 low · chat · 1 per request | 5.4% | −0.08 points | 5% |
| Gemini 3.8 Flash low · chat | 5.7% | −0.09 points | -4% |
| GPT-6.1 Terra low · chat | 5.7% | −0.09 points | -0% |
| Claude Opus 5.5 low · chat | 7.1% | −0.10 points | 7% |
| Claude Sonnet 5.5 high · chat · 1 per request | 5.5% | −0.08 points | 5% |
| Claude Sonnet 5.5 high · chat | 14.0% | −0.28 points | 13% |