# Evidence status

CoreStory does not publish a universal model leaderboard on this site. Product and model choices should be evaluated against the intended work.

## What this site currently supports

- **Architecture claims:** descriptions of persistent intelligence, typed evidence links, provenance, and MCP retrieval are documented as product design and implementation boundaries.

- **Integration claims:** playbooks describe the expected connection and retrieval workflow; users should verify the tools exposed by their own account.

- **Comparisons:** category-level differences are explanatory, not scored rankings.

## What is not published here

No current public dataset on this site substantiates a universal percentage improvement, accuracy ranking, latency curve, or cost advantage. Do not infer one from the architecture. Quantitative results belong here only with the task set, exact system configuration, scoring method, raw or inspectable outputs, uncertainty, test date, and limitations.

## Recommended evaluation

1. Select representative maintenance or modernization tasks from the target codebase.

2. Compare the same agent and model with and without CoreStory context.

3. Hold tools, prompts, budgets, retries, and scoring constant.

4. Score task completion and factual correctness; also record cost, latency, and failure modes.

5. Review a sample manually and preserve receipts so results can be audited.

A successful connection or plausible answer proves neither productivity improvement nor correctness. Validate claims against source evidence and real task outcomes.