Luni/laya-jev-benchmark

Laya vs Jev

TypeSafe released Jev on 15 September 2026, a closed Model that returns typed Decisions instead of Text. Three Days later an open Reproduction appeared, Laya (convaiinnovations/laya, Apache 2.0, 421M).

Laya's Model Card claims 83.8% against Jev's 67.8% and calls it a "+16.0% Advantage". Those two Numbers are from two different Benchmarks, so the Comparison says nothing.

I ran both on Benchmarks where Jev has published Numbers. One RTX 5090. Everything below is measured except the Rows marked "published", which are quoted.

Phishing

AreLit/PhishNChips, 2000 Emails, 1000 Phishing and 1000 legitimate. Neither Model was trained on it.

Model	Accuracy	ECE	AUROC	Recall	p50
Laya, raw	0.505	0.441	0.678	0.012	9 ms
Laya, Platt-calibrated	0.611		0.679		9 ms
Jev (published)	0.626	0.154	0.689	0.432	239 ms
Claude Haiku 4.5 (published)	0.813	0.097	0.837	0.764	687 ms

Raw, Laya is at Chance. It says "not Phishing" to almost everything, Recall 1.2%.

Its AUROC is 0.678 against Jev's 0.689, so the Ranking is nearly the same. Only the Threshold is wrong. Fit a Bias Term on 1000 Emails, score the other 1000, and Accuracy goes to 0.611.

Temperature Scaling cannot do this. It has no Bias Term, so it moves Confidence toward 0.5 but never across it. Platt Scaling can.

Jev's 0.626 is raw. Our 0.611 uses a Calibration Half. Jev would probably also improve. I cannot test that.

typed-decisions

LocalLLaMA/typed-decisions, 400 Cases, 2000 Decisions.

Model	Accuracy	ECE	ms per Case
Laya, no Fine-Tuning	0.360	0.175	15.9
Laya fine-tuned on this Task	0.767	0.212	16.4
Jev 1.13.0 (published)	0.727	0.144	710
Teacher Self-Agreement	0.735

The fine-tuned Model beats Jev. It also beats the Teacher Agreement Ceiling of 0.735, which is the Limit of real Signal in the Labels. Above that Line it is memorising Noise. Do not read this as better Understanding.

Latency

RTX 5090, fp16, after Warmup.

Questions in one Pass	p50	per Question
1	10.7 ms	10.7 ms
10	42.6 ms	4.3 ms
50	246 ms	4.9 ms
100	496 ms	5.0 ms

Cold Load 14.9 s. The Model Card quotes 38.4 ms and 156 ms on weaker Hardware, so its Speed Claims hold.

Failures Accuracy does not catch

Eleven Assertions any Decision Model should satisfy. Script in bench/probe.py. Run on one Support Ticket and one Phishing Email.

The Ticket:

I've been trying to export my data for three days and the button just spins forever. I'm on the Pro plan and I have a compliance audit on Monday. This is the second time I've written in.

Question	Answer	Truth
Has this Customer contacted Support before?	0.20	stated in the Text
Is the Customer on a paid Plan?	0.50	stated in the Text
Is the Customer reporting a Bug?	0.52	yes
P(needs a Human) + P(a Bot can resolve it)	0.09	should be about 1.0
P(Phishing) + P(legitimate Sender)	1.73	should be about 1.0

The last two are Question and Negation, both answered "no" at Confidence 0.94 and 0.97.

Routing, same Ticket, only the Option Names change:

Options	Verdict	Confidence
engineering / support / account	account, 52%	0.069
technical / support / billing	technical, 67%	0.221
"which Team fixes broken Features?"	account, 43%	0.019

Laya fails 7 of 11.

Fine-Tuning on public Data

180k Items, mostly BoolQ, SQuAD v2, SNLI, MultiNLI, ANLI, SciTail. Three Epochs, 55 Minutes.

	Probe Failures	typed-decisions	held-out Accuracy	Macro ECE
Laya base	7	0.360
no Consistency Loss	1	0.636	0.838	0.166
with Consistency Loss	2	0.676	0.840	0.156
EuroBERT-610m Backbone	12	0.296	0.366	0.326

Question	base	fine-tuned
Has this Customer contacted Support before?	0.20	1.00
Is the Customer on a paid Plan?	0.50	1.00
P(needs a Human) + P(a Bot can resolve it)	0.09	0.96
P(Phishing) + P(legitimate Sender)	1.73	1.00

Grounding 2/5 to 5/5. Contradiction 0/3 to 3/3.

The Arm without the Consistency Loss scored best on the Probes. The Data fixed this, not the Loss. The Loss bought Calibration: the Control answers every Grounding Probe at exactly Confidence 1.00, the other answers 0.89, 0.98, 0.62.

Checkpoint: Luni/laya-grounded.

Limitations

Stability got worse. Renaming Options still moves the Verdict. Base 2 of 3, fine-tuned 0 of 3 and 1 of 3. Shuffling Option Order during Training did not fix it. If you build Schemas at Runtime, this is not ready.

The Fine-Tune made Phishing worse. 0.512 against the base Model's 0.611, AUROC 0.569 against 0.679. Use the base Model for Phishing.

EuroBERT-610m failed for a boring Reason. Its Head started random while the other Arms inherited a pretrained one. Three Epochs does not converge a Head from scratch. This is not a Verdict on EuroBERT.

ECE is measured wrong on soft Targets. It is computed against argmax Correctness, so a Model correctly reporting 0.65 on a Target of 0.65 counts as wrong 35% of the Time. The high ECE on unli, chaos_mnli and typed_decisions is mostly that Artefact.

The Probe Suite is eleven Assertions I wrote myself over two Examples. It catches real Failures but it is not a Benchmark. Different Assertions would give different Numbers.

Both Models are about 20 Points behind Claude Haiku 4.5 on the Phishing Set.

Reproducing

uv venv --python 3.13 --managed-python .venv
uv pip install --python .venv/bin/python torch --torch-backend=auto
uv pip install --python .venv/bin/python laya datasets

.venv/bin/python bench/bench_phish.py convaiinnovations/laya
.venv/bin/python bench/platt.py convaiinnovations/laya
.venv/bin/python bench/probe.py convaiinnovations/laya
.venv/bin/python bench/eval.py convaiinnovations/laya --by-workflow

--torch-backend=auto or you get a CPU Build.

All Numbers in results/RESULTS.md, raw Logs in results/.

Licence

Code and Results Apache 2.0. Parts of bench/ derive from github.com/NandhaKishorM/laya, Apache 2.0.

The Checkpoint is CC-BY-NC-4.0, because the Mixture includes facebook/anli and Tobi-Bueck/customer-support-tickets. Drop those two and retrain for a commercial Version. The Phishing Numbers above are measured on the Apache 2.0 base Model and are not affected.

Attribution in CREDITS.md.