Seven Deciders In A Week. The Model Was Never The Product.
Verdict: stop choosing a decision model. Choose what it sits inside. We measured a loop that lifts a coin-flip decider to 92.5% accuracy and a 72.5% one to 96.1%, on a seventh of the model calls — and it takes any of this week's seven deciders as a drop-in part.
claude mcp add awdecide -- uvx awdecide mcp
Jev launched on September 15: a model that returns a typed choice, score or yes/no with a probability, never text. It was closed, so people rebuilt it. Within about 48 hours there were six open reproductions, as catalogued by AINews:
| project | what it is |
|---|---|
| Laya | 421M parameters, a ModernBERT-large encoder with two added layers that score the options, trained with PPO |
| DiffusionGemmaJev | the same task attacked from a diffusion model |
| Bespoke Nimble | a LoRA fine-tune of Qwen3.5-9B on contrastively curated data |
| SemIf (formerly OpenJev) | 4B and 35B Qwen3.5 backbones with a three-class classifier on the last token |
| Jevlike | a 40K byte-embedding model where each option attends to a shared context |
| Kev-0.5B | a LoRA adapter and a small readout head on Qwen2.5-0.5B |
Six backbones, from 40 thousand parameters to 35 billion. That spread is the finding. When an encoder, a diffusion model and a half-billion-parameter adapter all land in the same neighbourhood on the same benchmarks, the backbone is not where the value is.
What all seven have in common
Every one of them is a function. State and options go in, a scored answer comes out, and nothing comes back. None has a way to be told how the decision turned out. The ten-thousandth identical question gets the same answer as the first, at the same price, with the same chance of being wrong. AINews notes that at least one reports a confidence derived from entropy rather than one fitted to outcomes — which is what you get when a model has never seen an outcome.
That is not a flaw in any of them. It is what a model is. The missing part was never a better scorer.
Put any of them behind the loop
The loop is three verbs: ask, act, report what happened. A decision whose outcome was reported is answered from that evidence next time, with no model call. An answer reported wrong is withheld for good. We ran one experiment five ways, swapping only the accuracy of the decider plugged in — 40 situations, 10 sightings each, five random seeds per row:
| the decider you plug in | called every time | behind the loop | model calls |
|---|---|---|---|
| a coin flip (50%) | 52.0% | 92.5% | 400 → 65 |
| 60% accurate | 62.5% | 93.9% | 400 → 62 |
| 72.5% — Jev's independently measured accuracy | 73.1% | 96.1% | 400 → 55 |
| 80% accurate | 80.7% | 97.5% | 400 → 50 |
| 90% accurate | 89.7% | 98.7% | 400 → 45 |
Read the first row again. A decider that is guessing ends at 92.5%, because every wrong guess is reported once and never repeated. The quality of the model sets where you start. The loop sets where you finish.
These rows simulate a decider of each accuracy; we have not run the six
reproductions themselves. That is the point of the design — the decider is a
plug. awdecide bench --brain-acc 0.725 reproduces any row on your machine.
One honest edge: in four of the five rows the loop's probabilities also beat the base rate on Brier score. At a 90% decider they tie it (0.0138 against 0.0133). Accuracy still rises; the confidence numbers stop adding information up there.
The rung nobody else has: a 0.6B embedder of your own
Exact repeats are the easy half. The harder half is a situation that is almost one you have seen. That takes an embedder, and a rule for when two situations are close enough to share an answer.
Get that rule wrong and you have built a semantic cache, which fails in a specific way: it serves an answer borrowed from a situation whose right answer was different. We measured it on 40 labelled situations:
| neighbour rule | accuracy | model calls | borrowed-wrong rate |
|---|---|---|---|
| off (exact repeats only) | 97.5% | 40 | — |
| loose similarity floor, 0.85 | 96.0% | 6 | 11.59% |
| floor measured for that embedder, 0.96 | 98.5% | 30 | 0% |
The loose floor saves the most calls and gives back accuracy to do it. The measured floor beats exact-match on both columns. So the floor belongs to the embedder, it is measured, and an embedder nobody measured gets the strictest floor known, never the loosest. (That table was measured with a general-purpose embedder, whose safe floor is 0.96; our distilled one carries its own measured floor of 0.85.)
The embedder itself is ours: a 0.6B student distilled from a larger teacher, small enough to run beside your agent. On a real document set and 28 questions it put the right passage first 89.3% of the time, against 85.7% for the general-purpose model of the same size, 82.1% for a popular open embedder and 75.0% for the small default most stacks ship with.
What to do with this
Pick whichever decider you like. Laya if you want open weights and an encoder. Kev if you want something that fits anywhere. Jev if you want an API. A local 8B if you already run one. Then put it behind the loop, report outcomes, and watch the call count fall while the accuracy climbs.
Seven deciders shipped in a week, and there will be seventy by the end of the year. They are parts. The loop is the machine.
pip install awdecide && awdecide bench
Earlier in this series: $40 Million For A Function. We Built The Loop On One Desktop.
Sources
- explainx.ai, 6 Jev Clones in 2 Days
- lilting channel, Comparing 6 Open-Source Jev Clones
- awesome-jev-gallery — papers, open reproductions and independent evaluations
- Laya on Hugging Face
- jev-frontier-bench — the independent 72.5% figure