# Prepare one Kam State evaluation

This is a planning checklist for an Amazon Bedrock AgentCore application, not a
public SDK installer or access to a hosted service. Agree scope, price and the
non-production environment before an evaluation begins.

## 1. Name one problem

- Application and source: [brief description]
- Problem: [a supported retry, a correction, or changed recorded inputs]
- Integration owner: [person responsible for the application]
- Success check: [one observable result]
- Baseline: [what the application does today and what it costs to operate]

Do not put credentials, private records or raw customer data in an inquiry.

## 2. Inspect the public example

The recorded synthetic refund example has three stages:

1. `eligibility@1` references `request@1`: the recorded evidence is current.
2. The request changes to `request@2`: the old check is blocked because its input
   reference does not match.
3. `eligibility@2` references `request@2`: the recorded evidence is current again.

Current evidence does not approve a refund. The application decides which tools
to run and authorizes any external action.

- Recorded result: https://kam-ai.com/examples/readiness-result.json
- Source runner: https://kam-ai.com/examples/readiness-contract.py
- Asset checksums: https://kam-ai.com/examples/readiness-manifest.json
- Configuration example: https://kam-ai.com/examples/readiness-configuration.mjs
- Explanation and replay command: https://kam-ai.com/developers#readiness

## 3. Arrange the private prerequisites

To replay the contract, obtain the approved private Linux x64 worker matching the
checksum in the recorded result and runner, plus Python 3. The documented replay
was performed with networking disabled. A public worker binary is not included.

For SDK calls, separately agree the reviewed private SDK artifact, an authorized
endpoint and principal, server-side tenant binding, and an application-owned
producer adapter for retained request/eligibility Observations. Configuration
does not establish convergence. History `sourceRefs` are not trusted dependency
witnesses.

## 4. Check outputs and failures

- Compare the replay with the recorded `current → blocked → current` sequence.
- Verify that the runner accepts the genuine candidates and rejects its altered
  candidate; keep the output and worker identity together.
- On the authorized evaluation endpoint, use the documented History example to
  record a result twice with identical supported retry input. Confirm the original
  accepted receipt and no extra revision while evidence is retained.
- Correct the result and read both exact retained revisions. Treat an unavailable
  result as unavailable, never replace an exact requested revision with latest.
- Record any missing access, unsupported operation, failed assertion or incomplete
  source evidence as an unresolved requirement.

Retry example: https://kam-ai.com/docs/agent-result-retries
Changed-input guide: https://kam-ai.com/docs/changed-agent-inputs

## 5. Review the result

Keep the example, accepted receipts, exact inputs, and failures with the agreed
baseline. Measure investigation time, integration work, repeated tool calls and
total cost in the customer's environment before claiming savings. List deployment,
authorization, retention and exit requirements that remain open.

The public example proves a synthetic Rust contract only. It does not prove SDK
acceptance, persistence, agent execution, cloud qualification or a deployed
service. An inquiry does not create a subscription or grant access.

Evaluation scope: https://kam-ai.com/evaluation
