Contract Window for Bob
Preserve intent. Authorize the action. Inspect the evidence.
- # Contract Window — Bob First Mile
- ## Canonical user intent
- > Preserve user authority. Prevent
- > unauthorized scope expansion.
- > Produce verified evidence.
- Intent SHA-256: `b2c4f2d8…4db9`
- ### DONE
- - None. A card moves here only
- with an evidence reference.
What it does
Contract Window keeps the agreed task, the boundaries on an AI agent's actions, and the evidence for its work together. It gives the person directing the work a way to inspect what the agent is trying to do, what it is allowed to change, and what still needs proof.
The Bob First Mile implementation is a rehearsal kit for using IBM Bob to diagnose and repair an intentionally misconfigured Python application, execute it through the official watsonx.ai SDK, and check the result. The work follows a visible sequence: intent → diagnose → repair → execute → verify.
The intent stays in view
The repository records a canonical user intent, a version and a SHA-256 digest. The mini-Kanban defines the work. A card may move to Done only with an evidence reference. The ledger records the evidence rather than treating a confident answer as proof.
Boundaries distinguish permitted local repairs from actions that require human authority. Credentials, IAM roles, billable resources and changes to canonical intent do not become ordinary repair steps simply because an agent can suggest them.
A design with a wider purpose
The broader Contract Window design uses six sections in a fixed order: Intent Statement, Mini Kanban Chart, Kanban Notes, Verification and Truth, Next Steps, and Intent Score. A separate guard and a separate review workflow are requirements for the broader design; the interface alone does not establish their independence.
Corey's contribution connects instructional structure, governance design, research, engineering and human-readable evidence. The practical question is whether a person can follow the work and retain authority over it.
What the record supports
The inspected local status record describes the kit as offline rehearsal-ready. Live account certification, consecutive live inference successes, solo cold-run measurements and broader effectiveness are separate verification questions. The status shown in the picture is a historical artifact, not a claim that a live account is currently connected.
The team's submission was recognized among IBM's 50 top-scoring qualified submissions in its 2026 pre-conference hackathon. That recognition is distinct from validation of every research hypothesis or every proposed governance capability.