---
type: "post"
id: "analysis-of-decision-model-implementations-and-benchmarks-analysis-e34faf"
slug: "analysis-of-decision-model-implementations-and-benchmarks-analysis-e34faf"
title: "Analysis of decision model implementations and benchmarks"
description: "Thesis: Decision models turn selected agent control steps into constrained, single-pass probability prediction rather than open-ended text generation."
retrieval_nugget: "Thesis: Decision models turn selected agent control steps into constrained, single-pass probability prediction rather than open-ended text generation."
published_at: "2026-10-08"
updated_at: "2026-10-08"
record_date: "2026-10-08"
date_kind: "published_at"
topics: []
source_urls: ["https://habr.com/ru/articles/1091188?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:analysis-of-decision-model-implementations-and-benchmarks-analysis-e34faf"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:0b13e82f8af08e02b533b046dc48e635a504e4150db64e39b85c574a5313451a","reviewed_at":"2026-10-08T14:04:00.568Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Decision models turn selected agent control steps into constrained, single-pass probability prediction rather than open-ended text generation.","observed_facts":[{"text":"The article says Jev accepts system state plus typed questions and returns probability distributions over allowed answers, with no generated output tokens.","source_urls":["https://habr.com/ru/articles/1091188?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"It describes six releases or implementations within a 16-day period: Jev, Clef, Clef-flash, pplx-decider-v1-27b, Strands Decider 2B, Laya and GLiDE.","source_urls":["https://habr.com/ru/articles/1091188?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"It reports Clef ahead of Jev on BANKING77 (94.2 versus 79.7) and behind on When2Call (72.4 versus 81.0).","source_urls":["https://habr.com/ru/articles/1091188?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]}],"mechanism":"A conventional language-model output head scores a vocabulary and autoregressive decoding emits tokens; the described decision interfaces instead score a bounded answer set after processing the input. That can avoid token generation and output parsing when an agent needs a classification, routing choice, or risk label, while leaving thresholds and actions to application code.","why_now":"The article documents several vendors converging on this interface in a short period, making implementation differences and portability relevant to teams evaluating agent control paths.","implications":["The reported benchmark variation means teams should test decision models against their own tasks rather than assume a single model wins broadly.","Teams evaluating the pattern should include model-specific context windows, serving availability, thresholds, and fallbacks in their tests."],"evidence_boundary":"These are article-reported comparisons and claims; the supplied material does not independently validate benchmark design, provider claims, or production outcomes.","watch_conditions":["Track whether independent evaluations on real agent workloads confirm both quality and latency gains after including thresholds, fallbacks, and provider-side context limits.","Assess model-specific context windows, serving availability, and whether the chosen interface preserves the question semantics an application needs."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied, so this creates a distinct analytical record about the implementation and evaluation pattern described in this article."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/analysis-of-decision-model-implementations-and-benchmarks-analysis-e34faf/","markdown":"https://newruntime.com/posts/analysis-of-decision-model-implementations-and-benchmarks-analysis-e34faf.md","json":"https://newruntime.com/posts/analysis-of-decision-model-implementations-and-benchmarks-analysis-e34faf.json"}
---

# Analysis of decision model implementations and benchmarks

## Retrieval answer

Thesis: Decision models turn selected agent control steps into constrained, single-pass probability prediction rather than open-ended text generation.

Thesis: Decision models turn selected agent control steps into constrained, single-pass probability prediction rather than open-ended text generation. The Habr article provides a useful cross-section of six implementations, but it does not establish a universal winner.

Observed fact: The article says Jev accepts system state plus typed questions and returns probability distributions over allowed answers, with no generated output tokens. Observed fact: It describes six releases or implementations within a 16-day period: Jev, Clef, Clef-flash, pplx-decider-v1-27b, Strands Decider 2B, Laya and GLiDE. These are the article's reported facts, not independently reproduced results here.

Mechanism: A conventional language-model output head scores a vocabulary and autoregressive decoding emits tokens; the described decision interfaces instead score a bounded answer set after processing the input. That can avoid token generation and output parsing when an agent needs a classification, routing choice, or risk label, while leaving thresholds and actions to application code. Why now: The article documents several vendors converging on this interface in a short period, making implementation differences and portability relevant to teams evaluating agent control paths.

Implication: The reported benchmark variation means teams should test decision models against their own tasks rather than assume a single model wins broadly. The article reports Clef ahead of Jev on BANKING77 (94.2 versus 79.7) and behind on When2Call (72.4 versus 81.0). Evidence boundary: These are article-reported comparisons and claims; the supplied material does not independently validate benchmark design, provider claims, or production outcomes.

Watch condition: Track whether independent evaluations on real agent workloads confirm both quality and latency gains after including thresholds, fallbacks, and provider-side context limits. Also assess model-specific context windows, serving availability, and whether the chosen interface preserves the question semantics an application needs.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)