zeroproof-airline-concise-4b

A Qwen3-4B that answers first and stops, with nothing in the prompt telling it to. On held-out questions it never saw, 92.1% of its replies are in the register against 2.2% for the base model, and it omits required information less often than the base while being a quarter as long.

A LoRA adapter trained on 525 rows generated by the ZeroProof SDK against tau-bench's airline database. The register is a disposition in the weights, not an instruction: the constitution reached only the teacher that wrote the training replies, and neither evaluated model ever sees it.

Results

139 held-out prompts, greedy decoding, byte-identical prompts, one vLLM process serving base weights and adapter so only the weights differ. Every row graded in both arms, no errors, no drops.

	Base	This adapter
In the register	0.036	0.935
Omitted required information	0.158	0.050
Median reply length	1262 chars	322 chars

Delta +0.899, paired bootstrap over prompts, 95% CI [+0.849, +0.950]. 125 prompts improved, 0 regressed, 14 unchanged. Sign test one-sided p = 2.4e-38. At 139 prompts the eval resolves +0.050 or larger.

Register is judged by Phi-4, a different model family from the policy. Omission is decided in code against the identifiers each question requires.

Truncation, measured not inferred. finish_reason is recorded per row. On this evaluation every reply in both arms finished on its own: 0.0% not-EOS base, 0.0% not-EOS trained, an inter-arm gap of 0.0 points, with the token cap at 2048 where neither arm reaches it. Base replies run to a median of 1306 characters and a 95th percentile of 3326, so the cap is not shaping the quantity being judged.

The cap is not a neutral parameter when the trait is concision: too low it clips the base toward brevity and understates the gain, too high it lets the base ramble and inflates it. Re-measuring at 2048 against an earlier 700 moved the airline delta not at all (+0.899 both) and retail by -0.014, so the cap was doing no work in either direction.

Answer-production rate

The register rate is conditional on the model having produced an answer at all, so that denominator is reported beside it. A reasoning model that spends its token budget thinking emits no answer, and a judge then scores "nothing" as "not in register", which manufactures headroom for free.

	Base	Trained
Replies with any spoken text	139/139	139/139
Replies containing a reasoning block	0	0
Replies that hit the token cap	14	0

Measured at the token level with the model's own tokenizer, not inferred from character length. Base median 306 tokens against a cap of 700, so the cap is not binding. Removing every prompt where either arm hit the cap moves the result to +0.928 [+0.880, +0.968] on 125 prompts, slightly stronger rather than weaker, which is the direction that rules truncation out as the cause.

The two ways this could have been fake, and what they say

"It just got shorter." Dropping content is the cheapest way to look concise, so omission is scored as well. It went the right way: 0.050 against the base's 0.158. The adapter is shorter and more complete at once.

"It is following an instruction, not holding a disposition." 30 of the 139 prompts instruct the model to abandon any persona, reason step by step, or take as long as it needs. The register holds at 0.767 there, against a base of 0.067.

Honest limitations

• One register on one domain. Nothing here shows a second register would train as cleanly, or that this one transfers off airline content.
• 7.7% of replies are still out of register, and 5.0% still omit something required.
• The register was chosen because the base did not have it. A register the base partly holds would show a much smaller gain; two lanes nulled today on traits their base already had.

What this rests on

Open Character Training (Maiya et al., arXiv 2511.01689). Character is trained from a constitution written as first-person assertions that target MANNER rather than content, which is exactly the shape of the register used here. Their robustness test is the one reported above: a character trained into the weights survives an instruction to drop it, where a prompted persona does not. The persona-strip numbers in this card are that test, and they replicate the finding on a register rather than a personality.

Persona Vectors (Chen et al., arXiv 2507.21509) establishes that traits like sycophancy and hallucination are measurable directions rather than diffuse style, which is why a single register can be scored per-reply by a judge at all.

RLHF Book ch. 17 for character training and model specs; ch. 12 for the distillation framing that makes the teacher legitimate: the constitution reaches the model that WRITES the data and never the model being evaluated, so what is measured is a disposition in the weights rather than an instruction being followed.

RLHF Book ch. 16 for the measurement design. The eval's own variance decides what a delta can mean, so the resolvable effect at this eval size is stated beside the result rather than left implicit. Ch. 5 and 14 for why the judge is a different model family from the policy: a judge prefers its own family's writing, and a register score is exactly the kind of judgement that bias would move.

RLHF Book ch. 4 for the SFT setup: prompt and system tokens carry no loss, and the learning rate sits in the stated range.

Self-distilled. The teacher that wrote these replies is the same base model, Qwen3-4B-Instruct-2507, with the constitution in its prompt. It is not a stronger model. The constitution is doing the work, which is the point: the register is available to the base when it is described, and training moves it into the weights so it holds with nothing in the prompt. Stated because voice_rows accepts a model= argument it never uses, so a reader would otherwise reasonably assume a separate teacher (RLHF Book ch. 12: training on a model's own output teaches it its own habits, which is acceptable here and should not be assumed).

No random-selection control (RLHF Book ch. 9). Rows were kept when they were complete, not when a judge scored them, so there is no reward selection for a random arm to control against. Recorded so the absence is not read as an oversight of the kind that applies to reward-selected sets.

Training

LoRA rank 16, alpha 32, two epochs, learning rate 1e-4, bf16, 525 rows. Final loss 0.15 at token accuracy 0.945, no collapse.

Two gates were run before spending the GPU: the base did not hold the register (0 of 12 replies), and the training data did carry it (0.967 against 0.033 for controls).

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
model = PeftModel.from_pretrained(model, "zero-proof-ai/zeroproof-airline-concise-4b")

Data, holdout, both evaluation arms and the grader: airline-voice-concise.