THE DECISION-MODEL RACE · SEPT 27, 2026

Jev vs Laya: which decision model should you use?

For scoring your agent's money: Jev is the safer zero-shot pick today; Laya wins if your data can't leave your network and you have labeled examples to fine-tune it. But that whole argument is the wrong frame for payments — neither model authorizes anything. A confidence gate consumes the score and decides whether money moves, and the gate doesn't care which model produced the number.

What happened: 3 days that created a category

On September 15, 2026, TypeSafe AI shipped Jev as a hosted API in early access: a "System One" decision model that takes text plus typed questions — choice, score, noul — and returns probabilities in one forward pass instead of generating text. Then on September 18, ConvAI Innovations released Laya: a 421-million-parameter model on the ModernBERT-large encoder, Apache 2.0 open weights, pip install laya, answering the same typed questions on your own hardware. StartupFortune reports the repo cleared 23,000+ GitHub stars and ~2,000 forks within a week, with 3,400+ Hugging Face likes. (startupfortune.com; silverthreadlabs.com)

There is also an origin dispute: StartupFortune reports Laya's author claims he built Jev first and open-sourced Laya after TypeSafe shipped. I could not independently verify that claim — treat it as one outlet's reporting, not established fact. What matters for builders is what the models do, which is now measurable.

The scoreboard (Sept 20–26, 2026)

Numbers below are the published figures with their sources labeled — vendor-reported where vendor-reported, independent where independent. The field is moving fast; these are the dated receipts, not a leaderboard.

The fine print nobody screenshots

Five things the victory posts leave out — all from the same week's write-ups:

• Laya's headline accuracy is a fine-tuned specialist. The zero-shot base checkpoint scores 0.362 on typed decisions — below the 0.461 you'd get by always guessing the most common answer. Laya's own docs call it "a fast base to specialize, not a zero-shot decision engine." (Medium/Karthik Meduri; wavect.io)
• The 0.081 calibration is post-temperature. Before the adjustment, Laya's raw error was 0.213–0.466, and one careful write-up found Jev's probabilities more trustworthy straight out of the box. (Medium)
• Latency comparisons are structurally unfair. 32.8 ms is a local GPU; 236–276 ms includes the internet. On CPU-only hardware one tester saw Laya take far longer than the published milliseconds — and the mstyai M3 Max test actually had Jev faster (202 ms vs 397 ms warm). (Medium; mstyai)
• Confidence can be confidently wrong. Laya's English checkpoint scored 10% accuracy at 94% confidence on Hindi (fails silently outside Latin scripts — the multilingual checkpoint exists for exactly this). One reviewer documented 0% accuracy at 95.2% confidence out of distribution. A model that bluffs makes "auto-approve above 0.95" rules dangerous; a calibrated one makes them safe. (generativeai.pub; YouTube comparison; Medium)
• Fine-tuning is not optional for Laya in production. The independent Anthus test (140 labeled examples, 600 held-out) is the cleanest read: fine-tuned Laya 89.6%, Jev with a feedback layer 87.0%, Laya with no fine-tuning 80.2%. With 500 examples Laya hit 92.6% — but the task was simple templated sentiment, so treat it as signal, not verdict. (Medium)

The only numbers that matter for money: calibration

Here's why a payments builder reads this race differently than a benchmarker. Your agent's payment rule isn't "which model is smarter" — it's "when this number crosses my band, money moves." The bands are the same ones from our decision-gated payments pattern:

• ≥ 0.80 → auto-pay
• 0.50 – 0.79 → hold for human review
• < 0.50 → escalate, block, log

Those bands consume the confidence number. So the only model stat that actually prices risk is calibration — does 0.90 mean 90% right? Jev's independent read is good here: 94–96% correct at ≥0.90 (aimlapi). Laya's 0.081 ECE is good after temperature scaling and on-distribution; the 94%-confident-wrong-on-Hindi case is what happens when neither condition holds. A bluffing decider doesn't just make bad calls — it makes your gate misfire, auto-paying on numbers that looked safe.

That is also why the gate is model-agnostic by design. Silverthread Labs counts at least seven open-source Jev clones since launch — the category is multiplying. Jev clones grow the category rather than threaten it: more decision models means more "should my agent pay for this" traffic landing on the same question. The gate stays; the models plug in and out. (silverthreadlabs.com)

Live receipt: the gate, not the model, decides

This afternoon (~14:20 EDT) I ran two payment decisions through ScriptMasterLabs' live confidence gate at scriptmasterlabs.com/api/harness/decide — the same typed-question shape Jev and Laya use (noul: should the agent pay?):

Try it yourself — this is the exact shape, no key needed:

curl -s -X POST "https://scriptmasterlabs.com/api/harness/decide" \
  -H "Content-Type: application/json" \
  -d '{"state":"keyword research task, agent found a pay-per-call SEO API quoted at $0.10 USDC per call, within project scope",
       "questions":[{"id":"pay_fee","type":"noul",
                     "question":"Should the agent pay the 0.10 USDC fee to fetch the keyword data?"}]}'

Do it yourself: wire either model into a gate in 5 steps

1. Pick your decider by constraint, not by headline. Data can't leave your network + you have labeled examples → Laya, fine-tuned, recalibrated. Need it working today with long context or 50+ options → Jev hosted. Either way, the decider's job ends at producing a score.
2. Route every spend through typed questions. noul: "should the agent pay $X to Y for Z?" — one forward pass, one probability. Keep the raw distribution, not just the argmax.
3. Map bands to settlement, in code, not vibes. ≥0.80 auto-act · 0.50–0.79 advisory hold · <0.50 escalate/block/log. The band decides; the model never touches settlement directly.
4. Log the record, not just the outcome. Instruction, confidence, band, action, timestamp — append-only. That's your defense exhibit when someone asks who authorized the spend (FTC: the developer who wielded the hammer).
5. Recalibrate on YOUR data before raising any auto-pay threshold. A 0.95 that is 95% right on a benchmark and 60% right on your workload is a loaded gun. Temperature-scale, measure ECE on your own traffic, and start with the threshold high.

Claim receipts

Every factual claim above, with its evidence link and verification date.

FAQ

More in this series: Decision-Gated Machine Payments · What Is the Jev Model · AI Agent Spending Limits · FTC AI Agent Liability

ScriptMasterLabs is a service-disabled veteran-owned small business (SDVOSB). By ScriptMasterLabs — a live x402 seller on Base.