What is a decision model?

Open readouts: Kev

Kev's README describes a pointer head that compares each option's closing-token hidden state with the question's decision-token hidden state. Softmax converts the scores to probabilities. Kev documents cross-entropy training and stored calibration temperatures. This is Kev's disclosed implementation, not a specification of Jev's internals.

Kev model facts

Source: github.com

Open readouts: SemIf and Tev1

SemIf reads declared option logits in a forward pass without sampling an answer token. Its authors explicitly say it reproduces the interface pattern, not Jev's undisclosed model or training. Together's Tev1 retains a standard next-token head and generates an option letter. A decision-model interface therefore does not identify one shared architecture.

SemIf model facts · Tev1 model facts

Source: github.com · Source: huggingface.co

What TypeSafe means by RLCD

TypeSafe's AI primer expands RLCD as reinforcement learning for calibrated decisions. It describes training Jev to return decisions and probabilities, with higher probabilities intended to correspond to more frequent correct outcomes. Calibration concerns groups of predictions and does not guarantee an individual answer. VENDOR-REPORTED: this is TypeSafe's account of its training goal, not an independently reproduced training method.

Source: docs.typesafe.ai

Not public

The TypeSafe introduction, Choice documentation, AI primer and launch post linked here do not disclose the following details (checked 5 November 2026). Open implementations cannot fill these gaps.

• Jev's base checkpoint and parameter count.
• The training corpus and its selection or labelling procedure.
• The RLCD reward formula, loss and full training recipe.
• The internal probability head and parallel serving implementation.

Jev model facts · Decision models versus classifiers

Source: docs.typesafe.ai · Source: docs.typesafe.ai · Source: docs.typesafe.ai · Source: typesafe.ai

Choose the labels and error rule first

medgemma-decide-4b-v2 oversamples minority answers to improve one skewed question. VENDOR-REPORTED: its balanced accuracy on that question rises from 0.625 to 0.825, while the macro across sources slips from 0.689 to 0.686. Evaluate every question you need, with the same decision rule, rather than treating one corrected failure as an overall gain. The model is research only.

Medical-text fine-tune and limits · Compare with a classifier baseline

Evidence: Amaljith · medgemma-decide-4b-v2 card and results · 7 Nov 2026 · Source ↗

Keep calibration and evaluation separate

Laya Minecraft freezes its encoder and trains from verified game ticks, then fits temperatures on 284 calibration items. VENDOR-REPORTED: its 415-item evaluation contains four overlapping training or calibration items. Its accuracy and ECE changes therefore cannot establish an untouched-test gain. Preserve independent final examples when selecting weights and fitting thresholds.

Minecraft checkpoint and overlap · Held-out evaluation checklist

Evidence: Thomas Tesche · Laya Minecraft card and training report · 7 Nov 2026 · Source ↗ · Source: huggingface.co

Retain the actual training data and inference path

medical-jev-4b publishes a reconstructed training pool because its original file was lost. Its later 1,557-item evaluation may include sources seen by the released adapter. The default runtime accepts 52 options; the repository’s experimental tournament changes the benchmark result. Pin the data, checkpoint and deployed scoring path together. A candidate-count change can change what you are measuring.

Medical choice model facts · Source, code and GPU record

Source: github.com · Source: github.com

Check what the release actually includes

MoDe publishes inference code and three weight packages, but not its training code or data. Its weights use the FAIR licence for noncommercial research; the inference code is Apache-2.0. Laya Minecraft’s model card uses Apache-2.0, while medgemma-decide uses HAI-DEF terms. Check the specific model and data licences before reusing a recipe. These records do not supply a verified end-to-end fine-tuning tutorial or a universal compute budget.

MoDe access and licence · Documented local runtimes

Source: github.com · Source: huggingface.co · Evidence: Thomas Tesche · Laya Minecraft card and training report · 7 Nov 2026 · Source ↗ · Evidence: Amaljith · medgemma-decide-4b-v2 card and results · 7 Nov 2026 · Source ↗ · Source: github.com

When the next question depends on the last answer

TypeSafe’s cart cookbook starts with four menu items and one hard-coded cafe order. Code asks which remaining item comes first, saves that selection and removes it before the next request. Each question includes nothing_else. The loop stops on that option, an empty menu or a six-request cap. Removing a selected option prevents picking the same menu item twice. The recorded run finds a latte and a muffin, then stops. It is a worked example, not a general extraction benchmark.

Source: docs.typesafe.ai

What the loop leaves out

Each menu item can appear only once. The cookbook explicitly says two flat whites still produce one latte because the schema has no quantity field. It suggests adding quantity choices. Known fields such as size and milk are asked together after item discovery, with unspecified available when the order leaves a field out. Whether to ask independent yes/no questions or use a loop depends on whether the candidate questions are already known. Neither pattern supplies a universal accuracy or latency guarantee.

Source: docs.typesafe.ai · Source: docs.typesafe.ai

Check what the model actually sees

The workbench also offers experimental Jev image paths. Its local text Laya path receives measurements in words, so it cannot independently inspect pixels that OpenCV missed. Choose between a measurement check and a direct image judgment based on the information the question needs.

Laya model facts · Inspection workbench · Jev image inputs

Source: github.com

Training and image preparation change the result

The Glimmer inspection project reports 90.2% defect accuracy on 866 held-out images from known products after training its head on half of each product's test images. Its MVTec data is restricted to research. Separately, clef-playa's author reports lower vehicle counts on dense aerial tiles than two LLM annotators gave, and splitting tiles into quadrants narrowed that gap; true counting accuracy is not verified. Neither record compares OpenCV and a decision model on the same images, so these examples do not establish which is more accurate.

Glimmer inspection · Aerial counting limits · Evaluate on your own task

Source: github.com · Source: github.com

A question you can reuse

This is the benchmark's question object, to place under questions in a System One request. Supply the text in state and select the model separately. No response or accuracy on a new article is implied by this example.

{
  "sentiment": {
    "type": "choice",
    "instructions": "What is the overall sentiment expressed by the author of text ?",
    "criteria": {
      "positive": "The text expresses a positive opinion, emotion or attitude.",
      "neutral": "The text is factual or expresses neither a clearly positive nor a clearly negative sentiment.",
      "negative": "The text expresses a negative opinion, emotion or attitude."
    }
  }
}

What the saved test answers

The author reports Jev macro-F1 61.7 averaged equally across eight languages and 12,970 items; English is 68.0. These three-label benchmark scores do not give an accuracy rate for a new application's articles. Five LLM columns reuse paper results with different input formats. The report scores Jev's returned choice, including 53 items tied at the highest printed probability to two decimals.

Evidence: Uran Lajci · zero-shot conference report · run 6 Nov 2026, discovered 9 Nov · Source ↗