.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Tablekeeper: A BAND-Built Booking Service
Tablekeeper is a restaurant reservation service built through a three-agent software development workflow in BAND Desktop. The Coordinator assigns requirements and collects evidence, the Implementer builds and commits the service, and the Reviewer independently checks the implementation and reports actionable findings. Our submission covers Tablekeeper Stages 1 and 2. It includes a reservation API and a browser interface for restaurant availability, account signup and login, booking confirmation, reservation lookup, and cancellation. It also supports approved combinations of two tables and recovery when a booking succeeds but its response is lost. The team used Git commits, Docker containers, and the official evaluation harness. Stage 1 passed all 120 shipped checks in isolated mode. The BAND team reported independent acceptance of Stage 2 after review identified an incomplete-response handling issue and the Implementer corrected it. The run experienced usage-limit interruptions and runtime configuration recovery. Human resume messages and a later scope reduction to Stages 1–2 are disclosed rather than presented as an uninterrupted autonomous run. Stages 3 and 4 are not claimed. Passing the shipped checks does not guarantee passing the organizers’ additional tests.
6 Oct 2026
ProofPatch: Evidence Before Merge
AI coding assistants can generate plausible fixes quickly, but reviewing those fixes still requires developers to answer important questions: Did the patch satisfy the requirements? Were the tests changed to hide failures? Can another reviewer inspect and reproduce the evidence? ProofPatch is a working prototype that supports this review process. Built using IBM Bob IDE, it evaluates candidate patches against a fixed Git reference and protected verification files, then presents the results in an evidence dashboard. The demonstration uses a sample checkout implementation with two intentional bugs: a discount eligibility boundary error and incorrect inclusion of gift cards. Twelve requirement-derived tests evaluate three constructed candidate patches: • An incomplete fix passes nine tests but fails three. • A test-tampering patch is rejected before execution because it changes a protected test file. • A correct fix passes all twelve recorded checks. Developers run verification locally and inspect each candidate’s outcome, failed assertions, changed paths, commands, exit codes, and logs. The downloadable JSON report records the reference commit and SHA-256 hashes to identify the inputs used. Protected reference files are checked before and after patch application. ProofPatch brings patch screening, test execution, and review evidence into one workflow. It distinguishes failed checks, protected-file rejection, inconclusive execution, and success against the recorded checks. Rejected candidates retain an explanation without being presented as tested. The prototype targets developers and reviewers working with AI-assisted code changes. It automates repetitive verification and evidence collection while keeping the acceptance decision with the reviewer. The demonstration shows functional outcomes; it does not claim measured productivity gains. IBM Bob supported implementation and refinement of the verifier, automated tests, dashboard, and static export. Relevant Bob task screenshots are included in the public repository. The deployed website is clearly labelled as a recorded evidence demonstration. Live verification runs locally and is shown in the demo video. Temporary working directories provide working-copy isolation, not a security sandbox for untrusted code. Passing the recorded checks does not certify universal correctness or security.
27 Sep 2026
.png&w=828&q=75)
DuetShift
DuetShift is an adaptive bimanual Vision-Language-Action system designed for the Intel Physical AI challenge. It controls two simulated SO-101 robotic arms in MuJoCo to execute multi-step tabletop manipulation tasks from natural-language instructions and visual observations. Rather than executing a fixed sequence, DuetShift maintains a structured world state and task graph, assigns work between both arms, verifies important action postconditions, and replans when execution does not match expectations. The system is designed for coordinated reaching, grasping, transport, placement, shared-workspace manipulation, and bimanual handoff. Its signature capability is live goal revision. For example, while one arm is transporting an object, a user can change the requested destination. DuetShift updates the active goal, invalidates only the dependent future actions, dynamically reallocates work between the arms, and can use a safe handoff when the other arm is better positioned to complete the revised task. The same closed-loop architecture supports recovery when objects are displaced or the environment changes unexpectedly. The system combines MuJoCo simulation, camera-based perception, natural-language goal parsing, predicate-based world modeling, task planning, bimanual scheduling, manipulation policies, outcome verification, and recovery logic. The deployment pipeline is being designed around OpenVINO for efficient execution on Intel Core Ultra hardware. DuetShift focuses on robustness, interpretability, reproducibility, and responsive human-robot collaboration rather than one-shot scripted behavior.
16 Sep 2026
