Amazon’s Strands Labs has released the most complete open Jev-style package yet: a 1.9B-parameter checkpoint, Apache-2.0 code and weights, the data inventory, training recipe, evaluation harness, and a research log that records both successful and failed experiments. TechCrunch reports the release as another “Jev clone,” but the inspectable recipe is the consequential part.

How it works

  • Strands Decider starts from Qwen3.5-2B-Base, removes the language-model head, and adds a roughly one-million-parameter pointer head.
  • The head compares the hidden state at <answer> with the hidden state at each option’s last token, producing one score per offered option without generating text.
  • A rank-16 LoRA adapter tunes the torso while the small readout head runs in fp32 to avoid adding low-precision calibration error.
  • noul, choice, and ordered score questions all use the same masked probability distribution.
  • Several questions can share the state representation, though Qwen3.5’s recurrent layers limit some prefix-cache optimizations in the current implementation.

The reported evidence

  • The project reports 167 correct answers out of 231 on JevBench’s public set: 1.000 on its easy split, 0.875 on standard, and 0.505 on hard.
  • Its reported JevBench Brier score is 0.342 and expected calibration error is 0.052.
  • On an RTX 3090 under WSL2, the maintainers report 115 ms median and 299 ms p95 latency per JevBench question; an M3 Pro reportedly reaches a 153 ms warm median on prompts under 300 tokens.
  • These are project-maintainer measurements. I inspected the source, evaluation documentation, model card, and current commit f91487ab8f7e4b4967ae57e46b8d90e91e67d616; I did not download the multi-gigabyte base weights, rerun training, or reproduce inference.

Better-than-usual scientific hygiene

  • The repository says six retrains of an earlier recipe varied by 3.2 JevBench tasks, so it treats single-run differences under roughly ten tasks as unresolved.
  • Experiment notes state predictions and failure conditions before a run, then append outcomes; v20 did not replace v19 because its one-task gain was inside noise and four predictions failed.
  • The evaluation docs warn that the internal held-out corpus has ranked a worse external model higher, so it is for diagnostics—not model ranking.
  • Training sources, licenses, synthetic-data stages, hashes, calibration steps, and checkpoint provenance are published.

Where it fails

  • With state and options fixed, changing the question reportedly leaves the answer unchanged about 94% of the time; negated or reversed questions are therefore risky.
  • Hard, multi-step work is much weaker than short classification, including 0.267 on the documented temporal/numeric family.
  • Novel yes/no tasks and unseen scoring rubrics transfer poorly.
  • One fitted temperature per primitive does not make confidence portable: the maintainers explicitly tell users to measure thresholds on their own traffic.
  • A typed output is only schema-valid. It is not evidence that the selected option is correct or safe to execute.

Why it matters

Strands Decider makes the whole claim inspectable. The architecture, failed slot-head design, training data, calibration procedure, benchmark scripts, and limitations can all be challenged or reused. That is more valuable than another hosted endpoint promising speed.

The practical deployment pattern is still conservative: use a decider for bounded, reversible workflow choices; route low-confidence or unfamiliar cases elsewhere; and calibrate on labeled application data before allowing a score to trigger an action.

“A single-run difference under about 10 tasks is unresolved.”

The release is not proof that Jev-style models are reliable governors for agents. It is a credible open baseline for finding out.