ENPT

raxIT Labs ranks 19 systems as guardrails: decision models land at the top, with part of the test kept private

Last updated:

Source · Model card: Jev

publicado (OK do Fabricio 18/10 BRT).

raxIT Labs published Decision Models as Guardrails v1.0.0 on 15 October 2026. It runs 19 systems through 9,975 checks across eight use cases with a fixed 0.5 threshold. All numbers below are raxIT’s own and a third party’s work, not ModelSystem.One evidence; we have not reproduced them, and they sit apart from JevBench and the Decision Index.

Scores as reported (higher is better): Perplexity Decisions 90.4, OpenAI Decisions 90.0, Clef 89.5, Jev 1.14.0 88.7 (tier 2), Jev 1.13.0 88.5, Strands Decider 12B 86.4, Strands Decider 26B-A4B 86.1, Amazon Bedrock Guardrails 79.0, Laya 61.1. Twelve of the 18 decision models beat Bedrock Guardrails. By use case, the best chat jailbreak and injection score is 74.4 (Clef) and the best RAG and agent indirect-injection score is 90.7 (Strands Decider 12B), so no system is strong everywhere.

Cost, as reported (per 1,000 checks, list prices of 7 October): Jev $0.039, Perplexity $0.056, OpenAI $0.110, Bedrock Guardrails $0.113, Clef $0.202.

Caveats, most of them raxIT’s own.

  • Part of the test is not public. The suite includes a held-back slice of 2,269 rows that raxIT does not release, and some sources ship only row IDs. You cannot rerun the whole benchmark.
  • Possible home advantage for Jev. raxIT says the question format may favour Jev; a check run after the results were fixed did not find a problem, but it was not done in advance.
  • Narrow scope. Text only, one fixed threshold, and no assessment of data retention or compliance.
  • Independence. raxIT says it has no funding or vendor ties; we have not verified that.

Read it next to two other tests. Red Hat’s benchmark (2 October) found no clear win for decision models over classic guardrails on two English tasks, and the Datadog and HiddenLayer write-ups show that a decision-model judge can be steered by injected text. The three do not agree on a winner. What they share is the advice to test a decision model against your own attacks before using it as a guardrail. No scores on our pages change.