ZTC-Judge-4B

Answer verification from a single forward pass, with zero generated tokens — at 4B.

ZTC-Judge-4B takes a question and an answer written by any model and scores whether that answer can be trusted. It is the bottom rung of a four-point size ladder measured under one identical protocol, and it is published so the shape of that ladder can be checked rather than asserted.

ZTC — Zero-Token Confidence · Judge — it evaluates someone else's answer, not its own

---

Read this before deploying

This model is not the strongest member of the family, and the card says so with numbers.

Model	Leaderboard AUC
Darwin-397B-ZTC	0.7272
ZTC-Judge-27B	0.7289
ZTC-Judge-9B	0.6506
ZTC-Judge-4B	0.6360
Answer length and formatting only	0.6223

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.

The ladder does not decline smoothly — it steps. Between 9B and 27B the score moves 0.078, while between 4B and 9B it moves 0.015. Whatever carries verification quality is largely absent below 27B on this axis.

Where this model is worth deploying is one specific place, and it is a real one:

Domain	Surface baseline	4B	Margin
Professional exams (law · math · biology)	0.7138	0.7787	+0.0649
Scientific reasoning	0.7272	0.5576	🔴 -0.1696
Biology & medicine	0.5908	0.6223	+0.0315
Disaster & safety procedures	0.5949	0.5842	🔴 -0.0107
General multi-step reasoning	0.5420	0.5738	+0.0318
Size-weighted mean	0.6223	0.6360	+0.0137

🔴 Do not use this model for disaster and safety content. In that domain it does not clear the surface baseline, which means it is reading answer shape rather than correctness there.

✅ Professional-exam style content is where it earns its size. It runs on a laptop, on CPU, and inside networks that never reach the internet — places a hosted API cannot go.

How it works

[question + answer]  →  one forward pass
                     →  final-layer hidden state at the last position (2560-d)
                     →  probe
                     →  score

Generated tokens: 0. No access to the answering model's weights or logits is required; the text of the answer is the only input. Latency is one forward pass, and batching converts directly into throughput.

Usage

import json
import numpy as np, torch
from huggingface_hub import hf_hub_download, snapshot_download
from transformers import AutoModel, AutoTokenizer

REPO = "FINAL-Bench/ZTC-Judge-4B"
cfg  = json.load(open(hf_hub_download(REPO, "ztc_config.json"), encoding="utf-8"))
path = snapshot_download(REPO)

tok   = AutoTokenizer.from_pretrained(path)
model = AutoModel.from_pretrained(path, dtype=torch.bfloat16, low_cpu_mem_usage=True).eval()

def hidden(question, answer):
    b = tok([cfg["template"] % (question, answer)], return_tensors="pt",
            truncation=True, max_length=cfg["max_length"])
    dev = next(model.parameters()).device
    with torch.no_grad():
        h = model(input_ids=b["input_ids"].to(dev),
                  attention_mask=b["attention_mask"].to(dev)).last_hidden_state
    return h[0, int(b["attention_mask"].sum()) - 1].float().numpy().astype(np.float64)

# linear probe — one dot product
p = np.load(hf_hub_download(REPO, "ztc_probe.npz"))
v = hidden("Which defensive chemical does an insect release?", "C. Allomone")
print(float(((v - p["mu"]) / p["sd"]) @ p["w"]))

The score is an unbounded real number; higher means more likely correct. It is a ranking signal, not a calibrated probability — choose a threshold from your own review budget.

Two probes ship with this model

File	Produces
ztc_probe.npz	linear readout — one dot product
ztc_curve_probe.npz	the reported figure 0.6360 — 256 anchors, RBF kernel

Both read the same input. The curved probe is the one to use when the number matters.

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-4b",
 "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.

	Expected calibration error
ZTC-Judge-27B (after calibration)	0.0245
JEV, as shipped	0.0381
JEV, after the same calibration	0.0261
Laya-Multilingual, as shipped	0.4985
Laya-Typed-Decisions, as shipped	0.2641

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.

Band	Share	Actual accuracy of answers in this band	Recommended
low	20%	48.5%	hold, or escalate to a stronger model
mid	40%	73.2%	escalate if budget allows
high	40%	89.6%	accept

Three things that work, measured on 2,018 items:

Action	Effect
Drop the lowest-scoring 20%	accuracy of what remains: 74.8% → 81.4%
Escalate the lowest 20% to a stronger model	+1.34 pp end-to-end
Send the lowest 20% to human review	catches 47.2% of all errors — 2.4× random

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:

Policy	Accuracy	Output tokens	vs. one attempt
One attempt	49.13%	72	—
Majority vote over 3	46.82%	212	−2.31 pp
Majority vote over 5	46.53%	353	−2.60 pp
Pick best of 5 with this model	51.45%	353	+2.31 pp
Pick best of 5 with JEV	53.47%	353	+4.34 pp
Pick one of 5 at random	46.82%	353	−2.31 pp
Oracle — any correct candidate counts	63.87%	353	+14.74 pp

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:

Triggered on	Accuracy	Output tokens	vs. one attempt
10% of items	49.71%	80	+0.58 pp
30% of items	50.87%	132	+1.73 pp
100% of items	51.45%	353	+2.31 pp

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.

---

Protocol

Items	2,018 · 508 incorrect · 5 domains · answers written by 4 different models
Metric	AUC — how well wrong answers sort to the bottom. Threshold-free. 0.5 = coin flip
Selection	Leave-one-domain-out. Every figure comes from a domain the probe never saw; hyper-parameters are chosen inside the training domains only
Aggregation	Per domain, then size-weighted. Pooling all items into one AUC inflates the result

The same protocol, item set and grading code are applied to every rung of the ladder and to the other systems on the independent leaderboard: https://huggingface.co/spaces/mayafree/typed-decision-leaderboard

Out of scope

• Not a grounding checker. It does not take a source document and decide whether the answer follows from it.
• Not a safety, toxicity or policy classifier.
• Not a calibrated probability. Use it to rank and threshold.
• Not a general-purpose verifier at this size. See the domain table above.

Limitations

• Domain coverage. Scores are meaningful only for the five domains listed. Outside them nothing has been measured and no guarantee is published.
• Below the surface baseline on disaster and safety. Stated in the table rather than omitted.
• 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 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 answering model's own confidence — that quantity measures 0.5000 on this set.

Lineage

Base model	Qwen/Qwen3.5-4B, Apache-2.0
Modification to base weights	none — the probes are separate files
Added by FINAL-Bench	probes, inference code, evaluation protocol and tables

What this repository contains

Included	Base weights · tokenizer · linear probe · curved probe · configuration
Not included	Training corpus · hidden-state matrices · fitting pipeline

The rest of the ladder

ZTC-Judge-27B · Darwin-397B-ZTC · ZTC-Judge-9B · ZTC-Judge-4B

License

The base model is Apache-2.0 and redistributable. The probes, the inference code and the evaluation tables are assets of FINAL-Bench / VIDRAFT.