This is my entry for the Sanity Challenge, Path Two: Vibe-Code Something Strange. Code: Nuera Quicksilver (MIT). Sanity project ID: f87t11g1. Judge guide · public synthetic demo dataset.
What the build became
I set out to see whether a company’s operating rules could be treated as structured content—not buried in documents—and used by an agent without handing the agent the authority to act. Nuera Quicksilver grew into an operating console: state an objective, inspect a proposed plan, see the deterministic authorization result and its evidence, route the decision to a human when required, then inspect the recorded history.
The surprising part is not that a model can write a plan. It is that the rules around the plan are explicit and inspectable. Sanity holds typed company data: entities and capabilities, policies with scope and supersession, evidence with confidence and contradiction references, metrics, decisions and process definitions. The NQC Kernel is ordinary TypeScript with no model inside. It checks capability and authority, computes risk, and fails closed if a process definition is invalid. The agent can suggest; it cannot approve its own proposal.
The app is deliberately a narrow vertical slice, not a real-world ERP or autonomous factory controller. Approve and reject are human actions. “Execute” is simulated and writes an auditable demo decision/metric state; no real equipment, email, webhook or external business account is controlled by this demo.
The build process—and the parts that stayed rough
The work started by locking the company schema and the boundary between model and kernel. From there the project added the deterministic kernel, a seed scenario with conflicting policies and contradictory evidence, a Next.js console, and Sanity Context reads. A later pass wired in a real Sanity Knowledge Base and an independent advisory reviewer, then extended the decision lifecycle so the process and its guards are content in Sanity rather than hard-coded workflow prose.
The repository’s build log records the actual handoffs: an early architecture-through-hardening phase with MiniMax Agent, a later integration/deployment phase with Claude Code via Cowork, and developer-directed manual work in VS Code for terminal operations and secret entry. This is an AI-assisted build, not a claim that one coding session produced every line. The log also records bugs found and fixes made rather than presenting the process as a perfect first pass.
There are tradeoffs and unfinished proof. The current npm run verify path is credential-free and the README reports 1,282 tests. Live Context endpoints were verified by hand. Tests in CI use fakes for model and Context services; the prior build’s live stress/e2e results were not rerun against this dedicated project. There have been no real-user usability sessions. Model-evaluation calibration and real integrations remain unproven. The parity matrix and build log spell out the gaps.
Try the project
- Live app: project-quicksilver.vercel.app
- Source and build log: github.com/nueraresearch/project-quicksilver
- Judge walkthrough: FOR-JUDGES.md
-
Sanity project ID:
f87t11g1 -
Public dataset: inspect policies in the synthetic
demodataset. The app itself uses the privateproductiondataset; the public dataset is synthetic and contains no customer data.
The live app needs sign-in. Judges can choose Use an access token instead on /sign-in and use the challenge judge token:
qs_mh3a1ExwuX9QYAbsgjAFve6QdXgReucf6bU3Quug8Pk
It grants supervisor access to the demo app, and actions are recorded in Sanity. Please use it only to evaluate the entry. Execution is simulated; it does not operate external systems.
Top comments (0)