# Decision models vs classifiers, embeddings and rules > Compare decision models with trained classifiers, embeddings and rules using sourced task results, calibration checks and cost and latency limits. Updated 2026-11-05 (Australia/Adelaide) The recorded tests do not show a general advantage for decision models over a classic classifier, embeddings with a trained head, or rules. Logistic regression matched or beat [Jev]() on banking intents; Voyage reports a win for its trained embedding baseline; Coroot's burn-rate rule won its synthetic paging test. Each result applies to the task tested. [Evidence: ankur287 · jev-open-calibration, code and data for a Towards Data Science article · Oct 2026]() · [Source]() · [Evidence: Voyage AI blog · How to create your own Jev using Voyage models · 27 Oct 2026]() · [Source]() · [Evidence: Coroot · Anomaly detection for alerting · 03 Nov 2026]() · [Source]() ## What each approach needs A supervised classifier learns task labels from examples. BERT's paper describes fine-tuning a pretrained encoder with an added output layer. [Jev]()'s Choice API instead takes the question and option descriptions with each request. Runtime labels also predate [Jev](): Yin, Hay and Roth studied classification of unseen labels through textual entailment. A runtime-label interface does not remove the need for labelled examples when you measure accuracy or fit calibration. [Source]() · [Source]() · [Source]() · [Evidence: ankur287 · jev-open-calibration, code and data for a Towards Data Science article · Oct 2026]() · [Source]() ## Classic classifiers: the Banking77 result In ankur287's independent test, logistic regression with temperature scaling tied [Jev]()'s Choice accuracy on seven banking intents at 0.975. [Jev]()'s calibration error was 0.014 against 0.022, with the difference within noise. On all 77 intents, the trained classifier scored 0.888 accuracy against [Jev]()'s best tested 0.859 with three-shot criteria. The classifier used labelled Banking77 training data. These results do not establish how either approach would perform on a different label set. [Evidence: ankur287 · jev-open-calibration, code and data for a Towards Data Science article · Oct 2026]() · [Source]() ## Embeddings plus a trained head Voyage embedded documents, combined the embeddings with raw fields, and trained ridge regression on labelled documents. Across 29 classification tasks, its experimental voyage-json baseline averaged 88.3 AUROC against [Jev]()'s 77.6. VENDOR-REPORTED: Voyage evaluated its own models. This comparison gives the embedding baseline task-specific training examples while [Jev]() labels documents without learning from that collection. It is evidence for using existing labels, not a matched zero-shot comparison. [Evidence: Voyage AI blog · How to create your own Jev using Voyage models · 27 Oct 2026]() · [Source]() ## Rules: a paging example Rules need an explicit condition to test. Coroot's burn-rate rule used a fixed one-second service-level objective and a three-minute hold, scoring 30 of 31 synthetic scenarios correctly. Tuned [Jev]() 1.13 scored 29, [Tev1-4B]() 28, and untuned [Jev]() 27. VENDOR-REPORTED: these are Coroot's measurements. Model thresholds were tuned on the same scenarios, which were designed to be obvious under that fixed objective. [Evidence: Coroot · Anomaly detection for alerting · 03 Nov 2026]() · [Source]() ## Reranking has its own evidence Dinzinger and colleagues compared six Qwen backbones on ten TREC DL and BEIR collections. Reranker fine-tuning beat decision fine-tuning by 5 to 29 nDCG@10 points up to 9B parameters; at 27B they tied. Decision fine-tuning never produced the better reranker in that study. Its calibration advantage also depended on the question wording. [Evidence: Dinzinger, Ghosh Dastidar, Granitzer · System One models Are Not Better Rerankers · arXiv 2610.32318 · 23 Oct 2026]() · [Source]() ## Latency, cost and calibration can change the choice In Lucca's 231-email Italian and English comparison, logistic regression trained on about 200 labels had the highest accuracy on every field, with some statistical ties. Yet its composed actions ignored 18% of emails needing action, against 4% for the best generative model. A cost-chosen threshold reduced the classifier's misses. Per-field accuracy alone did not choose the best action policy. [Evidence: Federico Lucca · Measure Before You Train · arXiv 2610.32955 (v2) · 23 Oct 2026]() · [Source]() ## Measure the operating point The banking test fitted temperature scaling on out-of-fold predictions. Its same 0.90 threshold let three open backends automate 0.2%, 21.1% and 78.9% of traffic. Vals AI reports $0.025 per 1,000 claims for [Jev]() and median latency near 0.1 seconds in its long-document timing test. Those figures cannot establish a cost or latency ranking against the banking classifier, Voyage's embedding baseline or Coroot's rules on another workload. There is no matched Jev-versus-XGBoost result in these evidence cards. [Published decision model prices]() · [Test on your own workload]() [Evidence: ankur287 · jev-open-calibration, code and data for a Towards Data Science article · Oct 2026]() · [Source]() · [Evidence: Vals AI · Jev evaluation 6 Oct 2026, Mercury Decide evaluation 8 Oct 2026]() · [Source]() [What is a decision model?]() · [Architecture and RLCD]() · [Compare decision models]() · [Jev alternatives]()