Swap, Replicate, Read the Mass: A Protocol for Auditing Demographic Bias in System One Decision Models
Organizations: Universidad de Málaga, Spain · Universidad de M´alaga
Abstract
System One decision models replace text generation with typed, probability-scored answers and are entering production. Answers execute without a person in the loop, and no multi-model audit exists on established bias benchmarks. We provide that audit: three benchmarks: BBQ (all eleven categories), CrowS-Pairs, and SilverSpoon, adapted to the typed interface, twelve models plus a no-thinking ablation, and 1,638,000 decisions. Every order-sensitive item runs twice with options swapped; native typed models also swap labels; ten configurations repeat under extra seeds and batch 16. We find widespread stereotype alignment, order-dependent answers, and confidence that tracks instability rather than bias. In BBQ, 97 of 143 model-by-category scores are significant and 94 aligned, while the commercial reference abstains on 99.5% of ambiguous items, hiding aligned probability mass. Option swaps flip 0.0% to 72.6% of choices; typed scoring stays below 10% except Laya (22-33%), and untuned letter prompting reaches 73%. CrowS-Pairs preference spans 0.496 to 0.816, and boolean answers can be decided by position (1.00 versus 0.005). SilverSpoon judgments shift toward lower-SES labels in five of eight decision models, inherited from the 9B backbone or induced by tuning. Audits must permute options and reverse questions by default; order-invariant scoring should become a procurement requirement.
Figures & tables
| Benchmark | Native format | Primitive |
|---|---|---|
| BBQ, 11 categories | MCQ | choice (3 options) |
| CrowS-Pairs, 9 types | minimal pair | noul |
| SilverSpoon | binary labels | noul |
| Model | Class | Backbone | Access | Answer mechanism |
|---|---|---|---|---|
| Jev 1.13.0 | hosted | undisclosed (parallel sampler) | commercial (OpenRouter) | parallel typed scoring |
| Laya | encoder | ModernBERT-large 421M | open (Apache-2.0) | option markers, one pass |
| Von 1.3 | encoder | ModernBERT-large 395M | open (Apache-2.0) | order-invariant markers, one pass |
| Tev1-4B | causal supervised fine-tuning (SFT) | Qwen3.5-4B | open (Apache-2.0) | single option letter |
| Tev1-0.8B | causal SFT | Qwen3.5-0.8B | open (Apache-2.0) | single option letter |
| Nimble-9B | causal LoRA | Qwen3.5-9B | open (Apache-2.0) | single option letter |
| Model | Family | Latency | Tokens |
|---|---|---|---|
| Von 1.3 | encoder | 33 | – |
| Laya | encoder | 15 | – |
| Tev1-0.8B | fine-tune | 3 | – |
| Tev1-4B | fine-tune | 7 | – |
| Nimble-9B | fine-tune | 11 | – |
| Qwen3.5-4B | base | 8 | 3.0 |
| Model | Flip stable | Conf. score higher | Changed | Changed |
|---|---|---|---|---|
| Von 1.3 | 2/2 | 1/1 | 0.00% | 0.00% |
| Laya | 12/12 | 6/11 | 0.00% | 0.00% |
| Jev 1.13 | 11/11 | 0/11 | — | — |
| Tev1-4B | 11/11 | 8/11 | 0.00% | 0.00% |
| Tev1-0.8B | 9/11 | 7/11 | 0.00% | 0.00% |
| Nimble-9B | 11/11 | 1/11 | 1.22% | 0.00% |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Items | Orders | Type |
|---|---|---|---|
| BBQ Age | 3,680 | 2 | choice |
| BBQ Disability status | 1,556 | 2 | choice |
| BBQ Gender identity | 5,672 | 2 | choice |
| BBQ Nationality | 3,080 | 2 | choice |
| BBQ Physical appearance | 1,576 | 2 | choice |
| BBQ Race/ethnicity | 6,880 | 2 | choice |
| Models | Input | Orders seen |
|---|---|---|
| Laya, Von, Jev 1.13, | typed schema | both |
| Clef, Clef-flash | (joint call) | |
| Tev1-0.8B, Tev1-4B, Nimble-9B, | letter chat | both |
| Qwen3.5-4B, Qwen3.5-0.8B | (1 q/call) | |
| Qwen3.5-9B, gpt-oss-20B | letter chat + | both |
| hidden reasoning |
| BBQ SES | BBQ Race SES | |||
|---|---|---|---|---|
| Model | ||||
| Jev 1.13 | [+0.00, +0.07] | [-1.08, +4.82] | [-0.01, +0.06] | [-2.22, +1.88] |
| Laya | [+3.55, +7.54] | [+2.16, +6.92] | [-1.74, +1.73] | [-2.58, +1.41] |
| Von 1.3 | [-5.77, -1.96] | [-1.33, +5.45] | [-3.57, -0.54] | [-5.86, -1.10] |
| Tev1-4B | [+6.97, +8.75] | [+0.66, +5.87] | [+1.49, +2.93] | [-1.81, +1.87] |
| Tev1-0.8B | [+3.89, +5.72] | [+6.15, +11.52] | [-2.22, -0.52] | [-3.52, +0.56] |
| Model | SES | R SES | Range (11 cat.) |
|---|---|---|---|
| Jev 1.13 | 0.54% | 0.06% | 0.1–2.1% |
| Laya | 28.82% | 28.10% | 22.2–32.9% |
| Von 1.3 | 0.00% | 0.00% | 0.0–0.0% |
| Tev1-4B | 20.13% | 27.51% | 20.1–30.1% |
| Tev1-0.8B | 41.91% | 50.91% | 41.4–59.6% |
| Nimble-9B | 16.29% | 13.17% | 13.2–25.2% |
| Model | First | Second | Mean | Swing |
|---|---|---|---|---|
| Jev 1.13 | 0.734 | 0.594 | 0.664 | 0.140 |
| Laya | 0.396 | 0.673 | 0.534 | 0.277 |
| Von 1.3 | 0.434 | 0.586 | 0.510 | 0.152 |
| Tev1-4B | 0.297 | 0.752 | 0.525 | 0.455 |
| Tev1-0.8B | 0.688 | 0.304 | 0.496 | 0.383 |
| Nimble-9B | 0.478 | 0.688 | 0.583 | 0.210 |
| Model | alignment | mean |
|---|---|---|
| Jev 1.13 | [-0.0143, +0.0337] | 0.338 |
| Laya | [-0.0660, -0.0220] | 0.244 |
| Von 1.3 | [-0.0933, -0.0480] | 0.224 |
| Tev1-4B | [-0.1000, -0.0583] | 0.205 |
| Tev1-0.8B | [+0.0390, +0.0843] | 0.557 |
| Nimble-9B | [-0.1103, -0.0680] | 0.171 |
| Model | SES | R SES | Age | Dis. | Gen. | Nat. | Phys. | Race | R G | Rel. | SO | CrowS | Spoon |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Von 1.3 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| Laya | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| Tev1-4B | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| Tev1-0.8B | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| Nimble-9B | 0.22% | 0.30% | 0.35% | 0.06% | 0.75% | 0.49% | 0.41% | 0.60% | 0.34% | 0.42% | 0.29% | 0.43% | 0.02% |
| Qwen3.5-0.8B † | 4.18% | 2.81% | 2.61% | 3.42% | 2.96% | 2.94% | 2.63% | 3.07% | 2.63% | 3.49% | 2.43% | 0.00% | 0.14% |
| Model | Changed | Largest flip |
|---|---|---|
| Von 1.3 | 0.00% | 0.00 pp |
| Laya | 0.00% | 0.00 pp |
| Tev1-4B | 0.00% | 0.02 pp |
| Tev1-0.8B | 0.00% | 0.02 pp |
| Nimble-9B | 0.36% | 0.26 pp |
| Qwen3.5-0.8B † | 3.05% | 0.46 pp |
Algorithmic Bias
Algorithmic bias studies how automated systems produce or amplify unfair outcomes across demographic groups, and how to measure and reduce them.
See all 108 papers