Joe Barrow explains decision models through Perplexity Decider v1.2, a model built to classify inputs against questions and possible answers. Give it a document and ask whether a summary is faithful, for example, and it returns scored choices rather than a generated explanation.

Jev’s implementation is closed, so Barrow uses Perplexity’s model to explain one design. Its details should not be assumed to describe Jev.

One pass, several answers

A conventional language model generates output tokens in sequence. Decider instead places a special marker after each question and reads the answer scores at those positions in a single pass. That can avoid generating a separate response for every question.

Barrow describes adapting a Qwen language model: its full-attention layers are changed so tokens can attend across the request, while its linear-attention layers remain causal. A small readout layer scores answer options. The model’s scores are adjusted after training to improve calibration: how closely confidence matches observed accuracy.

The tradeoff is that a single-pass answer does not have the same opportunity for explicit step-by-step reasoning. Speed and lower generation costs can come with reduced performance on harder decisions.

Where it can help

  • Evaluate outputs: score a summary for faithfulness or completeness.
  • Route work: send simpler requests to a cheaper model and harder ones to a stronger one.
  • Sort and rank: classify documents or reorder search results by relevance.

Scores are not proof. An 80% confidence is useful only if it behaves appropriately on the tasks and data where it will be used. Check against labeled examples, including failures and changes in the input distribution.

Grouping questions changes the result

Perplexity’s model card says related questions can benefit from sharing a request, but independent candidates can interfere with each other. Its own tests found some retrieval tasks performed worse when competing candidates were grouped. Separate scoring is a safer baseline to test.

Caveat: Barrow’s explanation describes one model architecture, and the model-card comparisons are Perplexity’s own measurements. Neither establishes accuracy for a new workflow.