What is zero-token confidence judging?

A generation-free way to judge an answer: read the model's own internal representation and score every candidate in one forward pass, writing zero output tokens, so it cannot name an option that does not exist.

How do the three gates compare on the board?

Over 1,000 paired matches the reported gaps in lines cleared are ZTC minus JEV -0.07, a tie; ZTC minus Laya +1.03; and ZTC minus no gate +1.37.

Why is Laya's gate described as a null arm?

On the corrected set the root checkpoint scores AUC 0.4904, so the page says its gate reduces to always re-rolling; Laya-Multilingual's calibration error is 0.4985.

What does one judging call cost?

The page contrasts a 0.033 s readout against 3.206 s for a generative judge on the same 4,444-token, 200-candidate screen, and JEV is billed at 20 output tokens per decision.

Does re-rolling always help?

No. The single-retry measurement says re-rolling a bad move helps 38% of the time, while re-rolling a good move breaks it 30% of the time, so a gate is worth what it leaves alone.

Can ZTC run in an air-gapped network?

Yes, by construction: reading internal state requires local weights, so it cannot be offered as a hosted API against someone else's model, unlike the JEV API path.

Which models does the ZTC ecosystem include?

Darwin-397B-ZTC ships a self-readout probe, and ZTC-Judge-27B, ZTC-Judge-9B, and ZTC-Judge-4B judge other models' answers; this page races the judging gates on Tetris.

CommunityJev Demos & Experiments

Tetris JEV Laya ZTC

A static page that races three LLM judging gates — Laya, JEV, and zero-token ZTC — on the same Tetris board with identical pieces, proposals, and noise, publishing paired-match results and the measured constants behind them.

See highlights

Open huggingface.co

Format: Playground
Published by: Community
Author: FINAL-Bench
Last verified: Oct 2026

Tagscommunity playground demo calibration evaluation

Works forJev for Evaluation Jev for Decision gating Jev for Scoring

Used inJev for Software & DevTools Jev for Gaming

Platforms: Browser
Capabilities: Open sourceRuns locally

Highlights

Races Laya, JEV, and zero-token ZTC gates on one Tetris board with identical piece order, proposal, and noise across 1,000 paired matches.
Reported lines-cleared gaps (95% interval): ZTC-JEV -0.07 [-0.14, +0.01], a tie; ZTC-Laya +1.03 [0.79, 1.26]; ZTC-no gate +1.37 [1.16, 1.59].
On 1,939 Claude Haiku 4.5 answers the card reports ZTC-Judge-27B v2 at 0.7752 against JEV's 0.7521.
Measured constants: JEV AUC 0.7350 vs ZTC 0.7289 on 2,018 questions; gate latency JEV 0.591 s vs ZTC 0.636 s; calibration error ZTC 0.0245.
Ships in Korean, English, and Chinese, follows your browser language, and reads local model state, so it can run in an air-gapped network.

Watch out

A static, self-published comparison: the 2026-09-21 corrected set, the leave-one-domain-out protocol, and the wired constants come from the page itself rather than an independent rerun. JEV is billed at 20 output tokens per decision, and Laya's root checkpoint scores AUC 0.4904 on the corrected set, where the page says its gate reduces to always re-rolling.

Questions about this entry

Short answers drawn from the linked project or page.

Related terms

Glossary definitions related to this entry.

More Jev resources like this

JEV-9B Decision Demo

An interactive ZeroGPU demo of AutoTrust's JEV-9B: System 1 answers noul, choice, and score questions in one prefill pass with calibrated probabilities, while an Escalate tab hands low-confidence cases to the untouched Qwen3.5-9B base.