# AI Frontier Institute

> A public tracker of AI policy, research and news, and a record of what independent testers found when they tried US AI models on business work.

## Assurance data

Scores are in each source's own units. Never average or rank scores across sources. A model marked not_tested was not tried by that source.

- [Tasks](https://aifrontierinstitute.org/assurance/data/tasks.json): the business tasks, with coverage and last-checked dates
- [Models](https://aifrontierinstitute.org/assurance/data/models.json): the models in the record
- [Pulling data from documents](https://aifrontierinstitute.org/assurance/data/documents.json)
- [Writing and fixing code](https://aifrontierinstitute.org/assurance/data/code.json)
- [Analyzing data and spreadsheets](https://aifrontierinstitute.org/assurance/data/data-analysis.json)
- [Finance and analyst work](https://aifrontierinstitute.org/assurance/data/finance.json)
- [Agent reliability](https://aifrontierinstitute.org/assurance/data/agent-reliability.json)
- [Mathematics](https://aifrontierinstitute.org/assurance/data/mathematics.json)
- [Reasoning](https://aifrontierinstitute.org/assurance/data/reasoning.json)
- [Scientific tasks](https://aifrontierinstitute.org/assurance/data/science.json)
- [What improves results, and what it costs](https://aifrontierinstitute.org/assurance/data/what-helps.json): setup findings from every category, with cost where the tester published it
- MCP server (Streamable HTTP, read-only, no login): https://aifrontierinstitute.org/api/mcp
- [Weekly update](https://aifrontierinstitute.org/assurance/updates/latest.json): what changed in the latest Friday refresh
- [Agent skill](https://aifrontierinstitute.org/assurance/skill.md): install in a coding agent to use this record with its rules