NativeLink Agent Flight Recorder | Project Memo

video slotDemo #1 — trust an agent's work (record → verify → tamper goes red)

A ~3-5 minute walkthrough is scripted and ready to record: record an agent change into a signed, content-addressed attestation, run the independent verifier green, then flip one byte and watch the verdict go red (fail-closed). No public capture is recorded in this repo yet; pending capture, not fabricated.

pending

Demo #1 — trust an agent's work (record → verify → tamper goes red)

Bounded claim

Strongest applied-AI / platform-trust proof: a working local-first trust spine where signed, content-addressed attestations are independently verifiable with no producer access and fail closed on tamper.

RoleFounder / builder

StatusProof-grade demo; attestation core + independent verifier shipped

Available proof1 / 3

Sourceoperator-recorded

Stack

Launch surface

Repo

Project

Agent-validation trust spine (attestations + independent verifier)

When can you trust an AI agent's work? NLFR records an agent's build as signed, content-addressed evidence and ships an independent verifier — trust is re-derived from the artifact instead of taken on the agent's word, and verification fails closed the moment a byte is tampered.

Problem

Operating context

Agentic code work gets hard to trust because the truth is scattered: a chat thread says one thing, the logs say another, a cache hit hides what really ran, and the only 'proof' an agent did the work is the agent's own word. Teams adopting AI agents need a way to know — independently, after the fact, with nothing but the artifact in hand — whether an agent's change is what it claims to be.

Ownership

What I owned

Designed and built the trust spine: the content-addressed evidence model, the signed attestation envelope, and the independent verifier — sequenced behind hard gates so each layer was provably real before the next was built.