• 

  Luna Grounding Auditor — judge-ready deployment certification interface

• 

  Corrected surface: 100% grounded and certified safe to deploy

• 

  Broken surface: deployment blocked at the prompt-builder layer

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Luna Grounding Auditor

Devpost Draft

Project name

Luna Grounding Auditor

Tagline

The CI system for AI agents.

What it does

Luna Grounding Auditor certifies whether a specific AI surface can actually see the business facts it needs before that agent is deployed. It compares canonical knowledge with the exact context exposed to a voice agent, chatbot, copilot, or workflow. It then runs adversarial questions, checks expected evidence IDs, detects unsupported claims, identifies the layer where knowledge was lost, and returns a release decision: CERTIFIED or NOT SAFE.

The problem

Agent teams often validate ingestion but not runtime visibility. A document may exist in a vector database while the production voice prompt receives only a weak summary. The system appears trained, yet the live agent cannot answer basic questions or safely reject unsupported requests.

How we built it

The standalone Next.js application includes a deterministic audit engine, synthetic broken and corrected fixtures, an evidence explorer, a visual knowledge pipeline, downloadable certification reports, and an optional server-side GPT-5.6 workflow.

GPT-5.6 is used for adversarial question generation, surface-only answer simulation, semantic grounding evaluation, failure-layer diagnosis, and smallest-safe remediation. Codex accelerated the repository architecture, implementation, test coverage, UI iteration, CI workflow, and submission assets.

Why it is different

This is not another agent builder or observability dashboard. It is a pre-deployment safety gate focused on the gap between knowledge that exists and knowledge the active AI surface can actually access.

Judge demo

The deterministic demo needs no API key. Load the broken fixture, run the audit, inspect the missing evidence and knowledge-loss point, apply the suggested fix, and re-run the certification. The project moves from NOT SAFE to CERTIFIED in under 90 seconds.

Track

Work and Productivity

Built With

• codex
• digitalocean
• github-actions
• gpt-5.6
• next.js
• openai-api
• react
• typescript
• vitest
• zod