Machinery
Reading path
Field notes · November 2026

Decision models arrive, and the guide goes live

A new kind of model that answers instead of writing, why its benchmarks already disagree, and the first frontier radar.

This is the first edition of the field notes. Each month they answer one question: what changed in the field that should change how you understand or use these systems? Most months, the honest answer will be “less than the headlines suggest”.

The big one: decision models

Since September, a new category of model has appeared that doesn’t write text at all. TypeSafe AI released Jev in mid-September; AWS’s Strands Labs followed on 1 October with Strands Decider 2B, an open-weight model under the Apache 2.0 licence. Both take a situation plus a narrow question with fixed answers (“allow, ask a human, or block?”) and return a probability for every option in a single pass, in roughly a tenth of a second.

Under the hood this is less exotic than it sounds. Strands Decider starts from an ordinary pretrained language model (Qwen3.5-2B) and replaces only the very last layer, the one that normally scores every possible next word, with a small layer that scores just the options you pass in. Everything the model learned about language stays; it simply can’t ramble any more.

Why it matters: agent systems are full of small yes/no and pick-one decisions: is this input suspicious, may this tool run, is the task finished, which team should get this ticket? Until now, each one meant a full, slow call to a big model. Decision models make those checkpoints cheap enough to put everywhere. The guide’s new chapter Around the model has a live demo of one acting as a safety gate.

Why to stay sceptical: the benchmarks already disagree. Strands’ own table puts Decider 2B at 0.78 on JevBench; an independent run on the 231 public tasks found 72.3%. One evaluation found a decision model matched a frontier model’s verdicts more often than the true answers. Another found Jev answering “cheap” for 99.5% of model-routing requests, which is no better than always giving the same answer. A fast probability is only useful if it’s calibrated. Measure on your own decisions before trusting one.

The first frontier radar

The radar now tracks 19 directions across architecture, training, systems and security. This month’s placements are the baseline. From next month on, you’ll be able to replay how entries move.

Two placements worth explaining:

  • Diffusion language models: emerging. A 2026 survey reports several-fold speed-ups at comparable quality, but reasoning, long texts and hallucination behaviour are still largely unexplored.
  • Prompt-injection defences: emerging, not shipping. Real mitigations exist and ship in products, but none solves the underlying problem that instructions and data share one channel. Calling them “shipping” would overstate the safety you get.

What we’re watching

  • Test-time training (models that learn while they read): two strong papers this year, one an ICLR 2026 oral. It’s the direction most likely to change the “two memories” picture in this guide.
  • Whether decision models hold up under independent evaluation, and whether open rebuilds close the gap to Jev.