# Model Card Transparency Audit

Audit model card releases against the Foundation Model Transparency Index, checking documentation completeness across pre-training data sources, compute hours, hardware types, carbon disclosures, and safety limitations.

## Workflow

1. Ingest model documentation audit checklist across 5 core transparency dimensions:
   - Data provenance (data sources disclosed, filtering method described, copyright handling).
   - Compute infrastructure (hardware type, cluster size, total FLOPs or GPU-hours disclosed).
   - Environmental impact (energy consumption in MWh, carbon emissions disclosed).
   - Model capabilities & limits (intended use cases, out-of-scope uses, known failure modes).
   - Evaluation methodology (benchmark harnesses named, prompts published, contamination checked).
2. Calculate transparency disclosure score:
   $$\text{Transparency Score} = \frac{\sum \text{Disclosed Indicators}}{\text{Total Indicators}} \times 100\%$$
3. Flag critical opacity gaps (e.g. models hiding training cut-off dates or omitting safety evaluations).

## Synthetic Input Format

JSON object with:
- `schemaVersion`: 1
- `model`: string
- `publisher`: string
- `disclosures`:
  - `dataSourcesNamed`: boolean
  - `dataFilteringDescribed`: boolean
  - `hardwareArchitectureDisclosed`: boolean
  - `computeBudgetDisclosed`: boolean
  - `carbonEmissionsReported`: boolean
  - `knownLimitationsDocumented`: boolean
  - `evaluationHarnessCodePublic`: boolean
  - `safetyRedTeamingReported`: boolean
- `assertions`:
  - `minTransparencyScorePercent`: optional decimal string (e.g. "75.0")
  - `requireLimitationsDocumented`: optional boolean (default: true)

## Exit Codes

- `0`: Valid audit meeting all transparency thresholds.
- `1`: Transparency score below requirement or mandatory disclosure missing.
- `2`: Malformed input or safety violation.

## Verification

```sh
python3 scripts/check.py examples/synthetic.json
```
