> ## Documentation Index
> Fetch the complete documentation index at: https://docs.useharmonia.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Evaluate Harmonia

> A practical path for testing contracts, agent quality, authority boundaries, and end-to-end evidence.

Evaluation has three layers: deterministic contract tests, offline agent-quality fixtures, and authenticated evidence runs. Passing an earlier layer does not prove a later one.

## 1. Contract and application checks

Run `npm test -- --run`, `npm run lint`, `npx tsc --noEmit`, and `npm run build`. These verify schemas, state transitions, APIs, UI behavior, documentation, and production compilation without claiming a cloud effect.

## 2. Agent evaluations

Run the Python suite under `agent/tests/`. Cases cover grounded output, missing evidence, invented references, authority overreach, incomplete contracts, revision boundaries, and eligible Memory Bank context. The evaluators in `agent/harmonia_agent/evaluation_contracts.py` fail closed on prohibited or ungrounded output.

## 3. Vertical-slice evidence

Follow [Evidence runbook](/evidence-runbook). Use an authorized real source, authenticated providers, exact approval digests, immutable receipts, and independent read-back verification. Never substitute fixtures or mock mode for authenticated proof.

## Interpreting results

| Result                    | Establishes                               | Does not establish     |
| ------------------------- | ----------------------------------------- | ---------------------- |
| Unit tests pass           | Local contract behavior                   | Provider availability  |
| Build passes              | Production compilation                    | Deployment success     |
| Authenticated invocation  | Provider accepted one request             | End-to-end correctness |
| Receipt plus verification | Applied effect was independently observed | Future reliability     |

Record known gaps explicitly and link evidence to immutable identifiers rather than screenshots alone.