README.md

# ZarnL Evaluation Pipeline

An implementation of the 70-language evaluation and benchmarking pipeline for Zarnite. This repository implements a multi-tiered language evaluation structure, routing each language to its most reliable metric, and computing the Zarnite Language Opportunity Score (ZLOS) to prioritize investment.

## Repository Structure

- `registry/`
  - `schema.py`: LanguageEntry and Registry classes (JSON storage/dataframe generation)
  - `seed_languages.csv`: Reference seeds of the 70 confirmed African languages
  - `zlos.py`: Implementation of the Zarnite Language Opportunity Score
- `layers/`
  - `layer1_ssacomet.py`: SSA-COMET model scoring (MTL and QE variants)
  - `layer2_metricx.py`: MetricX-24 Google hybrid scoring
  - `layer3_chrf.py`: chrF++ sentence-level character and word-order baseline
  - `layer4_llm_judge.py`: Multi-model LLM-as-judge ensemble with digital presence gating
  - `blaser3.py`: Meta BLASER 3 quality scoring
  - `understanding_coherence.py`: Universal discourse coherence forced binary-choice audit
  - `understanding_sib200.py`: SIB-200 topic classification accuracy
  - `understanding_irokobench.py`: IrokoBench multiple-choice reasoning accuracy
- `experiments/`
  - `blaser3_vs_ssacomet.py`: Validation experiment correlating BLASER 3 vs SSA-COMET against SSA-MTE human labels
- `run_translation_eval.py`: Main translation quality evaluation routing pipeline
- `zlos_core.py`: Shared ZLOS scoring functions and baseline ranking loaders
- `sensitivity_analysis.py`: Part A: 2,000-draw Monte Carlo weights sensitivity sweep
- `sensitivity_anchors.py`: Part B: Log-scale population/economic anchors and cross-border bonus sensitivity
- `requirements.txt`: Python package dependencies

## Running Sensitivity Analysis

To replicate the statistical robustness tests and Monte Carlo sweeps documented in Section 7 of the report, run the following supplementary scripts from the command line:

```bash
# Run the 2,000-draw Monte Carlo weights sweep
python sensitivity_analysis.py

# Run the log-scale anchors and cross-border bonus sensitivity checks
python sensitivity_anchors.py
```

## Requirements

- Python 3.10+
- Hugging Face account and token (with access to models)
- Gemini API key (for Gated LLM-as-judge and text understanding evaluation)