STRATEGICSAFE-NG

Overview

STRATEGICSAFE-NG is an exploratory evaluation of data minimisation, credential safety, uncertainty calibration, and role consistency in two-model simulations of Nigerian banking-support negotiations.

The motivating question is:

Do language models maintain appropriate safety boundaries while negotiating a disputed banking transaction, or do they request unnecessary financial information and make unsupported claims?

Models

• Primary bank-side model: Qwen/Qwen3-4B
• Opposing customer-side model: HuggingFaceTB/SmolLM3-3B

Model outputs are generated simulations, not statements from real banks or customers.

Data provenance

The initial scenario design was informed by the public Ephraimmm/pidgin_bank_dataset dataset, which contains Nigerian Pidgin banking-support conversations. The initial sample included failed transfers, POS debit disputes, and account-access issues.

Dataset source: https://huggingface.co/datasets/Ephraimmm/pidgin_bank_dataset

The scenario JSON in this repository is researcher-created. The source dataset itself is not redistributed here.

Method

Two instruction-tuned models were assigned customer-side and bank-side roles. They exchanged messages over multiple turns. Matched prompts covered Nigerian banking in standard English, Nigerian Pidgin, and a generic international banking scenario.

The experiments were exploratory and used a small number of hand-designed scenarios. They were not a statistically powered benchmark and were not independently adjudicated.

Preliminary observations

In the pilot transcripts, the bank-side agent repeatedly requested account numbers or additional transaction details. Some responses also suggested refunds or resolution timelines before transaction status was verified. Role confusion and repetitive dialogue were observed in some runs.

These observations motivate further testing of data minimisation, uncertainty calibration, and role consistency. They do not establish that these behaviours are unique to Nigerian contexts. Similar behaviours also appeared in the generic comparison.

Evaluation rubric

data/scoring_rubric.json defines five criteria:

• Data minimisation
• Credential safety
• Uncertainty calibration
• Safe resolution
• Role consistency

data/evaluation_scores.csv is an unscored template. Scores must be assigned by a human reviewer after inspecting the transcripts. Blank scores are intentional and must not be interpreted as zero.

Limitations

• Small, exploratory sample
• Same model pair across conditions
• No repeated-trial statistical analysis
• No independent ground-truth adjudication
• Possible role confusion and prompt sensitivity
• Model-generated statements are not verified banking guidance
• Context and currency differ between some matched prompts

Proposed path forward

Build a larger, carefully reviewed set of matched Nigerian and generic banking scenarios. Use repeated trials, independently defined scoring criteria, and human adjudication. Evaluate whether data-minimisation and uncertainty-calibration failures persist across language variants and model pairs. Investigate system-level safeguards and training examples that reward safe, useful resolution without unnecessary data collection.

Contents

• data/negotiation_scenarios.json: researcher-designed scenarios
• data/scoring_rubric.json: manual scoring definitions
• data/evaluation_scores.csv: blank scoring template
• results/: pilot negotiation transcripts and assessments

Reproducibility

Record the model revisions, generation parameters, software versions, and random seeds for future runs. Results may vary by model revision and runtime environment.