FINAL-Bench/ZTC-Judge-27B
<!-- ZTC-ECOSYSTEM:START -->
ZTC ecosystem
 
Zero-Token Confidence (ZTC) reads a model's own internal state to judge whether an answer is right — generating zero tokens. Darwin-397B-ZTC ships a self-readout probe; ZTC-Judge-27B judges other models' answers; the arcade shows the gate deciding execute or hold in real time. <!-- ZTC-ECOSYSTEM:END -->
ZTC-Judge-27B
Ask a model how confident it is and you get a coin flip (AUC 0.500). This reads the same forward pass and gets 0.7289 on domains it has never seen — above the surface baseline in all five.
ZTC-Judge-27B takes a question and an answer written by any model and scores whether that answer can be trusted — without generating a single token.
ZTC — Zero-Token Confidence Judge — it evaluates someone else's answer, not its own
How it works
[question + answer] → one forward pass through the 27B model
→ final-layer hidden state at the last position (5,120-d)
→ one dot product with the probe
→ scoreGenerated tokens: 0. No access to the answering model's weights, logits or log-probabilities is required — the text of the answer is the only input. Because nothing is decoded, latency is one forward pass and batching translates directly into throughput.
Usage
from ztc_judge import ZTCJudge
judge = ZTCJudge.from_pretrained("FINAL-Bench/ZTC-Judge-27B")
judge.score(
"Which defensive compound does the insect release when it meets a predator?",
"C. Allomone",
domain="scientific_reasoning",
)
# {'score': -0.61, 'verdict': 'review', 'domain_auc': 0.7410, 'generated_tokens': 0}Batched scoring (recommended — there is no decode step, so batch size is throughput):
judge.score_batch([(q1, a1), (q2, a2), ...], domain="scientific_reasoning")The score is an unbounded real number; higher means more likely correct. It is a ranking signal, not a calibrated probability. Pick a threshold from your own review budget (see below).
Evaluation
Two independent measurements are reported. They use different test sets and different protocols, and are not interchangeable — read the protocol line before quoting a number.
A. Independent leaderboard — 2,018 items, leave-one-domain-out
The Typed Decision Leaderboard scores answer verifiers from several vendors on one identical item set with identical labels: <https://huggingface.co/spaces/mayafree/typed-decision-leaderboard>
A second probe ships with this model, for answers from models we have not seen
The table above is measured on a set where 79% of the answers come from one model family. A probe fitted on it can end up specialised to that family rather than to correctness. We checked, using a model that appears in neither the fitting nor the scoring — Claude Haiku 4.5 — answering the same 2,018 questions. Those 1,939 answers were graded with the same code and the same domain weights.
3,000 paired bootstrap resamples, weighted by domain.
Which file to use. ztc_curve_probe.npz produces the reported figure on the benchmark above and is the reference for it. ztc_curve_probe_v2.npz is fitted on answers from several models and is the one to use on answers written by a model that is not in that benchmark — which is the usual case in deployment.
🔴 The two do not agree, and the direction reverses. Scored on the 2,018-question benchmark under one identical method, v2 reaches 0.7066 where v1 reaches 0.7236. Specialisation helps on the distribution it was specialised to and costs elsewhere. Both files are published so the trade can be made deliberately rather than discovered.
Revised 2026-09-21. Earlier revisions of this card reported 0.7364 for Darwin-397B-ZTC and 0.7282 for ZTC-Judge-27B. Those figures were produced by a run whose standardisation statistics were computed over all five domains, including the held-out one, which leaks a small amount of the evaluation domain into every figure. Re-run with the statistics fitted inside the training domains only, the figures are 0.7272 and 0.7289. Cite the current values.
| Patronus Lynx 8B | 0.5179 | | The answering model's own stated confidence | 0.5000 |
Per domain, and against the surface baseline in the same domain:
Above the surface baseline in all five domains. That baseline — a model reading nothing but answer length, digit count and formatting — is the bar that matters; a verifier below it is detecting shape, not correctness.
Protocol. Every figure comes from a domain the probe never saw. Hyper-parameters are selected inside the training domains only. Scores are computed per domain and then size-weighted; pooling all items into one AUC inflates the result, because score scales differ between domains.
Self-readout. Given only the question, this model answers on its own and the same forward pass tells whether it was right: 0.7322 (3 domains, 1,595 items). Tools that see only text from outside a model cannot do this at all.
B. Repository-internal measurement — 400 items, random folds
This number is higher than the leaderboard figure because the protocol is easier: one domain, random folds instead of held-out domains. Quote 0.7289 when comparing against other systems.
API — drop-in for an existing JEV integration
The endpoint takes the same request shape and returns the same response shape, so switching an existing integration is a URL change.
POST /v1/evaluate
Authorization: Bearer <token>
{"model": "vidraft/ztc",
"state": {"question": "...", "answer": "..."},
"questions": {"correct": {"type": "boolean",
"instructions": "Is the ANSWER factually correct?"}}}{"model": "vidraft/ztc-judge-27b",
"answers": {"correct": {
"probability": 0.1043,
"verdict": "review",
"score": -0.72,
"position": 0.268,
"band": "low",
"action": "hold_or_escalate",
"measured": {
"band_accuracy": 0.485,
"base_accuracy": 0.748,
"if_lowest_20pct_dropped": 0.814,
"escalate_gain_at_20pct_budget": 0.0134,
"do_not": "resample_same_model",
"why_not": "measured: fixes 6.7% of wrong answers, breaks 13.1% of right ones"}}},
"usage": {"generated_tokens": 0}}type accepts boolean and noul. Existing clients read answers.<key>.probability and ignore the rest; the additional fields are there for clients that want to act on the score rather than merely record it. 0.19 s per call, zero generated tokens.
What probability means
The raw score is unbounded. The shipped calibration maps it to P(answer is correct), fitted leave-one-domain-out — the mapping never sees the domain it is applied to.
Measured on the same 2,018 items. ZTC and JEV are effectively tied on calibration; the difference of 0.0016 is not meaningful. Figures published elsewhere for these systems were measured on other test sets and do not reproduce here.
🔴 Calibration is uneven across domains: 0.0225 on biology & medicine, but 0.2941 on scientific reasoning and 0.2381 on general reasoning. Treat the probability as reliable in the first case and as a ranking signal only in the other two.
What to do when the score is low
The score alone is not actionable, so the response carries the measured consequence of each choice.
Three things that work, measured on 2,018 items:
Generate several candidates and let the verifier pick
When the score is low, the most effective next step is not to escalate — it is to produce more candidate answers and select between them. Measured on 346 questions with five candidates each (1,730 candidates), all scored through this endpoint:
The same five candidates swing by 5 points depending on how one is chosen. Majority voting is worse than not resampling at all: when a model prefers a wrong answer, more samples make that wrong consensus more certain. A verifier that ranks the candidates is what turns extra samples into accuracy.
Spend the budget only where it is needed. Generating extra candidates only for low-scoring first attempts captures most of the gain at a fraction of the cost:
At a 30% trigger rate you get three quarters of the benefit for 1.8× the tokens, where always generating costs 4.9× for 1.3× the benefit.
Scope: one generator (GPT-4o-mini), one item set, five candidates. The oracle row shows the headroom that remains — a correct candidate is present far more often than any policy recovers it.
One thing that does not work:
🔴 Do not take a majority vote over resamples. Measured: five resamples with majority voting score 46.53% where a single attempt scores 49.13%. More candidates make a wrong consensus more certain unless something picks between them — see the table above.
Escalation pays for itself through precision, not recall. Re-answering repairs about 38% of wrong answers and damages about 30% of right ones, so a gate is only worth its budget if it mostly calls answers that are actually wrong.
Two probes ship with this model
Both read the same input: the final-layer hidden state at the last position, from a single forward pass with zero generated tokens. The curved probe is the one to use when the number matters; the linear probe is there for deployments where 59 KB and a single dot product are worth the difference.
Review budget
If you can only re-check part of your traffic, sort by score ascending and review the lowest.
Intended use
- Triage of LLM output at scale — route low-scoring answers to a human, a larger model, or a retrieval step, and let the rest through.
- Regression monitoring — track the score distribution of a deployed system over time.
- Cheap pre-filter in front of an expensive verifier — costs one forward pass, no decoding.
Out of scope
- Not a grounding checker. It does not take a source document and decide whether the answer follows from it. Tools built for that task target a different axis.
- Not a safety, toxicity or policy classifier.
- Not a calibrated probability. Use it to rank and threshold, not as a likelihood.
- Not a substitute for human review in high-stakes decisions.
Limitations
- Domain coverage. Scores are only meaningful for the five domains above. When
domain_aucreturnsNone, the domain was never measured and the score carries no published guarantee. - Weakest domain is 0.6172 (general multi-step reasoning). General multi-step reasoning is the hardest case and is listed first in this section on purpose.
- Sensitive to which model wrote the answer. The probe is fitted on answers from four models. Adding 1,772 answers from a single additional model shifted the mixture and lowered the size-weighted score from 0.7278 to 0.7177 — professional exams rose to 0.8575 while every other domain fell. Treat "works on any model's output" as a design goal, not a measured guarantee: if your generator differs sharply from the training mixture, measure before relying on the number.
- Revision lock. The probe is fitted to one specific revision of the base model. Running it on a different Qwen3.5-27B revision produces no error and silently wrong scores. This repository ships the matching weights so that failure mode cannot occur.
- It reports the verifier's judgement, which is not the same thing as the answering model's own confidence — that quantity measures 0.500 here.
- Long answers are read within the model's context window; answers exceeding it are truncated before scoring.
Lineage
What this repository contains
The probe is the product; the material it was fitted from is not distributed.
Independent comparison
This model is listed on the Typed Decision Leaderboard, which measures answer verifiers from several vendors on one identical item set with identical labels, and publishes scores, labels and grading code:
<https://huggingface.co/spaces/mayafree/typed-decision-leaderboard>
License
The base model, Qwen3.5-27B, is Apache-2.0 and redistributable. The probe, the inference code and the evaluation tables are assets of FINAL-Bench / VIDRAFT.
Citation
@misc{ztc_judge_27b_2026,
title = {ZTC-Judge-27B: zero-token answer verification from a single forward pass},
author = {FINAL-Bench},
year = {2026},
url = {https://huggingface.co/FINAL-Bench/ZTC-Judge-27B}
}