The model.
The process.

What model do we use?

The BERT component uses BERT-base-uncased (google-bert/bert-base-uncased), a pretrained transformer encoder. We reuse an encoder adapted in our earlier local study and keep its weights frozen in this version. A new balanced logistic-regression classification head is trained on the revised provisional labels. We did not train a foundation model from scratch. The second choice uses the contrastively adapted MiniLM encoder and its saved logistic-regression head.

Sentence representations use the non-padding mean of the final hidden states, including special tokens, followed by L2 normalization. The head uses C=1, balanced class weights and a fixed random seed of 20261102.

How this version was prepared

1. Preserve the original dataset, labels and historical experiments.
2. Use Expert-03’s classifications provisionally for 101 reviewed cases; retain Expert-0’s classifications for other cases.
3. Exclude 14 insufficient-context cases from binary training. Fit the new head on 1,212 unchanged sentences.
4. Reuse the earlier outer-fold-1 adapted BERT encoder for implementation, not because it was selected as the best-performing fold.
5. For a submitted sentence, tokenize without truncation, compute its representation, and propose BIAS or NO_BIAS. Requests exceeding 256 tokens are declined.
6. Compare the selected classifier’s label with Gemma’s fresh label. Show a simple wording checklist and general guidance, then require human expert finalization.

Expert-02’s review is pending. A later update will create a separate version; it will not silently replace historical research findings.

Two live hybrid choices

Choose fine-tuned BERT + fine-tuned Gemma or SetFit–MiniLM + fine-tuned Gemma. Each choice makes fresh predictions on Runpod GPUs and works when the researcher’s Mac is off. This is a decision-level combination: the classifier and Gemma assess the same input separately. Agreement produces a provisional label; disagreement produces no combined label and requests priority expert review. Model weights are not merged between components.

Both choices reuse the fixed outer-fold-1 deployment snapshots. The revised Expert-03 classification head applies to BERT only. The SetFit and Gemma components retain their earlier training versions; they have not been retrained on Expert-03’s revisions.

How Gemma assesses the sentence

Gemma-2-9B-it uses the earlier local LoRA adapter, exported from its MLX affine four-bit base into fused half-precision weights for CUDA serving. Export checks verified the reconstructed weights and two fictional local label examples. These checks establish conversion consistency, not independent research accuracy.

With the fixed label-only prompt, the scorer compares mean completion log-likelihoods for BIAS and NO_BIAS. These are relative candidate scores, not calibrated probabilities. This deployment uses a different computing system from the earlier Apple Metal runs, so numerical differences remain possible.

What do the outputs mean?

A classification is a provisional suggestion. The bias score is uncalibrated and is not a validated probability. The wording checklist identifies surface terms; it is not SHAP, LIME or an explanation of BERT’s internal reasoning. Gendered terms can be appropriate in context. Suggested next steps are general guidance, not a trained or validated rewriting system.

Validation and limitations

This head was fitted on all eligible provisional development data. It has no independent accuracy estimate. Previous out-of-fold results belong to earlier model and label versions and cannot be claimed for this updated model. Diagnostic expert-review cases were selected rather than randomly sampled. The original corpus has few authentic positive cases, unresolved synthetic parentage and limited source coverage. Expert disagreement remains relevant.

Passing application tests demonstrates software operation, not classification accuracy or reliability. Do not use this prototype for automatic decisions about people.

Privacy and deployment

No research sentences, dataset files, saved model responses or expert annotations are published. Submitted text is sent to Runpod to produce a response and is not saved by this application. Runpod’s job and output retention may apply. Avoid submitting confidential personal information. There is no public review database or dataset download. Model weights are kept outside publicly served files.

A small DigitalOcean server hosts the website and private submission gateway. Runpod serves the models on demand, scaling to zero when idle. The seven-day test permits up to 30 predictions daily across both choices, with a shared spending guard. Cold starts can take several minutes. Hosting and software checks do not establish model validity.