# Evidence methodology

CoreStory treats understanding as a set of typed claims that remain connected to evidence. The system should be able to say what it knows, why it believes it, and where uncertainty remains.

## Evidence chain

1. **Observe.** Parse source artifacts and record stable anchors.

2. **Derive.** Produce typed claims about structure, behavior, architecture, and business meaning.

3. **Link.** Connect derived claims to supporting evidence and lower-order claims.

4. **Classify.** Record epistemic status, scope, provenance, and uncertainty.

5. **Retrieve.** Return only context relevant to the current task, preserving citations and qualifications.

6. **Replay.** Retain enough provenance and configuration to inspect or reproduce a result when the underlying inputs are available.

## Epistemic status

Source observations, human-authored assertions, model-generated summaries, and multi-source syntheses are different kinds of knowledge. Higher-order claims should point downward to evidence; conflicting claims should remain visible until resolved.

## Evaluation principles

- Define the task and intended use before selecting a metric.

- Disclose the tested model, version, system prompt, tools, context, retry policy, and budget.

- Report task outcomes with cost, latency, and uncertainty where available.

- Keep domains separate when aggregation would hide important differences.

- Publish limitations and update dates beside results.

- Prefer reproducible artifacts and raw results to an unexplained overall rank.

## Boundary

This page describes CoreStory's evidence discipline and evaluation standard. It does not claim that every product surface exposes every field or that every planned evaluation is complete.