EdgeProof

Codex-operated verification for private LLM infrastructure on Linux, k3s, and constrained NVIDIA GPUs.

EdgeProof is the OpenAI Build Week proof and verification repository for the Edge-Computing-LLM platform.

It provides:

• a read-only Go verification CLI,
• a repository-level Codex Agent Skill,
• sanitized evidence,
• Build Week documentation,
• and competition-ready demonstration material.

Why EdgeProof exists

A local LLM platform is more than a model server. A working deployment depends on Linux, Kubernetes, GPU runtime support, scheduling, model availability, networking, telemetry, and operational verification.

A command can succeed while the underlying platform is unhealthy. EdgeProof creates a clear, reproducible verification entry point and documents the boundaries of what is proven.

What is verified

The edge-openai CLI checks selected infrastructure contracts:

• required host commands,
• NVIDIA GPU visibility,
• Kubernetes API access,
• k3s node access,
• NVIDIA RuntimeClass,
• allocatable Kubernetes GPU resources,
• Helm releases,
• GPU infrastructure namespace access,
• observability namespace access,
• and selected deployment, StatefulSet, and DaemonSet listings.

The current version verifies that these commands and contracts are reachable. It does not yet semantically parse every Kubernetes object or automatically repair failures.

Verified reference result

The reference environment produced:

Passed:   13
Warnings: 0
Failed:   0
Verdict:  VERIFIED

Reference hardware and software:

• Lenovo ThinkPad T450s
• Intel Core i7-5600U
• 12 GiB RAM
• NVIDIA GeForce 940M
• 1 GiB VRAM
• Ubuntu 24.04
• k3s 1.36
• NVIDIA container runtime
• Ollama
• Open WebUI
• OpenTelemetry Collector
• Prometheus
• Grafana

This is a demonstrated environment, not a minimum requirement.

Quick start

Clone:

git clone https://github.com/Edge-Computing-LLM/edge-computing-openai.git
cd edge-computing-openai

Run local validation:

make check

Run live infrastructure verification:

make prove

Or build and run directly:

go build -o bin/edge-openai ./cmd/edge-openai

./bin/edge-openai version
./bin/edge-openai inspect
./bin/edge-openai prove
./bin/edge-openai prove --format json
./bin/edge-openai prove --verbose

make prove requires access to the expected local Linux, k3s, Helm, and NVIDIA environment. GitHub Actions runs unit tests, formatting, vetting, race tests, and build validation, but does not run the live cluster proof.

Commands

version

./bin/edge-openai version

Prints the CLI version.

inspect

./bin/edge-openai inspect

Checks required commands and selected host and Kubernetes contracts.

prove

./bin/edge-openai prove

Extends inspection with GPU and observability namespace checks.

Output formats:

./bin/edge-openai prove --format human
./bin/edge-openai prove --format json

Bounded command output:

./bin/edge-openai prove --verbose

Custom timeout:

./bin/edge-openai prove --timeout 30s

Platform relationship

EdgeProof does not replace the platform repositories.

Repository	Responsibility
edge-cli	Unified installation and operational CLI
k3s-nvidia-edge	Linux, k3s, and NVIDIA infrastructure
llm-observability-stack	Local inference and observability workloads
gguf-observability	Prompt-free GGUF runtime evidence
edge-llm-tests	Cross-repository validation
Edge-Computing-LLM-Documentation	End-user documentation
edge-computing-openai	Codex Skill, proof CLI, evidence, and Build Week material

OpenAI and local-model boundary

• GPT-5.6 through Codex CLI was used to design, generate, refactor, test, and document the platform.
• A local GGUF model performs local inference.
• The local model is not presented as GPT-5.6.
• The local runtime does not require an OpenAI API key.
• EdgeProof does not send local prompts or model responses to OpenAI.

Codex Agent Skill

The repository includes:

.agents/skills/edge-llm-operator/

The skill instructs Codex to:

• start with read-only inspection,
• use existing platform tools,
• treat successful Completed jobs correctly,
• avoid reading or displaying secrets,
• avoid storing prompts and model responses,
• distinguish historical warnings from current health,
• and request approval before a mutating repair.

Privacy and evidence policy

EdgeProof must not intentionally collect or publish:

• prompt text,
• model response text,
• API keys,
• GitHub tokens,
• Kubernetes Secrets,
• kubeconfig credentials,
• private keys,
• browser session data,
• model weights,
• or unrestricted raw logs.

Generated live reports are ignored by Git. A sanitized example is available at:

evidence/examples/verified-reference.json

Build Week work

Substantial Build Week work across the organization includes:

• automatic CPU and NVIDIA orchestration,
• Go-first validation,
• low-VRAM model profiles,
• prompt-free GGUF evidence,
• k3s networking checks,
• observability validation,
• repository documentation,
• the EdgeProof Go CLI,
• and the Codex Agent Skill.

The submission does not claim that every historical line of organization code was created during Build Week.

See Build Week Evidence.

Development

make check

Equivalent checks:

gofmt -w cmd internal
go test ./...
go test -race ./...
go vet ./...
go build ./...

Run live proof:

make prove

Security

Version 0.1.0 is read-only.

Future mutating functionality must use:

• explicit approval,
• allow-listed actions,
• bounded execution,
• clear rollback instructions,
• and post-change verification.

See Security Policy.

Documentation

• Architecture
• Build Week Evidence
• Demo Guide
• Known Limitations
• Privacy
• Devpost Description
• Video Script
• Submission Checklist
• Contributing
• Support
• Changelog

License

MIT License. See LICENSE.