# iVisionQA AI Disclaimer

**Effective Date:** March 6, 2026

## Nature of AI Analysis

iVisionQA uses Anthropic's Claude Vision, a large language model with vision capabilities, to analyze screenshots of web applications. The following disclosures apply to all AI-generated outputs from iVisionQA.

## Key Disclosures

### 1. Probabilistic Outputs

AI analysis scores and assessments are **probabilistic**. The same screenshot analyzed multiple times may produce slightly different scores and issue lists. This is inherent to how large language models operate and does not indicate a defect.

### 2. Not Deterministic

Unlike pixel-based visual regression tools, iVisionQA's AI analysis is not deterministic. Scores may vary by 5-10 points between identical runs. Use score thresholds (strictness levels) rather than exact score matching.

### 3. Not a Substitute for Human Review

AI scores and issue reports are intended to **augment**, not replace, human QA review. iVisionQA should be used as one tool in your quality assurance process. Do not use AI scores as the sole basis for:

- Production deployment decisions
- Accessibility compliance certification
- Legal compliance determinations

### 4. Accessibility Scanning

The `a11yScan` fixture uses axe-core (not AI) for accessibility testing. Axe-core results are deterministic and based on WCAG 2.1 AA standards. However, automated accessibility testing can only detect approximately 30-40% of accessibility issues. Manual accessibility testing remains necessary for full compliance.

### 5. Visual Regression Detection

AI-powered visual comparison (`aiCompare`) identifies semantic visual differences, not pixel differences. It may:

- **Miss** subtle changes (single-pixel shifts, minor color variations)
- **Flag** acceptable changes (dynamic content, timestamps, animations)
- **Vary** in sensitivity between runs

Use the `allowedChanges` parameter to reduce false positives from known dynamic content.

### 6. Model Updates

iVisionQA uses Claude's vision capabilities, which may be updated by Anthropic. Model updates can affect scoring consistency. When upgrading iVisionQA versions, re-establish baseline expectations by running your full test suite.

### 7. Cost Awareness

Each AI analysis call consumes API tokens billed by Anthropic. Costs depend on:

- Screenshot resolution and size
- Number of analyses per test run
- Model selected (configurable via `IVISIONQA_MODEL`)

Estimated cost: ~$0.01-0.03 per analysis at default settings. Use the `--a11y` flag to run accessibility-only tests without AI costs.

## Strictness Levels

| Level    | Pass Threshold | Use Case                            |
| -------- | -------------- | ----------------------------------- |
| `low`    | Score >= 50    | Early development, rapid iteration  |
| `medium` | Score >= 70    | Standard QA, CI/CD gates            |
| `high`   | Score >= 85    | Pre-release, production deployments |

## Recommendations

1. **Use CI gating with appropriate strictness** — `medium` for development branches, `high` for release branches.
2. **Combine AI and deterministic testing** — Use `a11yScan` (deterministic) alongside `aiAnalyze` (probabilistic) for comprehensive coverage.
3. **Review AI reports** — Treat AI findings as suggestions for human review, not automatic pass/fail decisions.
4. **Monitor costs** — Use `ivisionqa info` and check `~/.ivisionqa/usage.json` to track consumption.

## Contact

Questions about AI analysis: support@ivisionqa.dev