Why I'm building a from-scratch, zero-trust neural network instead of relying on API wrappers.

Everywhere I look, founders are building AI agents by wrapping an OpenAI or Anthropic API, giving the agent root access, and hoping it doesn't hallucinate a destructive command.

Coming from a 14-year background in Security Operations, I couldn't accept that risk. So I architected a different approach: JB & MARI, a dual neural-symbolic architecture built for zero-trust security and local-only inference.

Here is the breakdown of how it works:

JB (The Neural Fortress / Actor):
A strictly local, passphrase-gated system. All memory is encrypted with AES-256-GCM, and it maintains a hash-chained HMAC audit log. It operates in a sandbox with zero network access. If it detects tampering, it triggers an automatic LOCKDOWN.

MARI (The Neural Core / Critic):
An independent verifier. MARI is forbidden from using pretrained weights or cloud APIs. It uses a pure symbolic reasoning kernel (bounded forward chaining, STRIPS planning) to verify JB's actions. Missing info is treated as unknown, not assumed to be true.

The Next Phase (V0.11):
We are currently building a custom Actor-Critic neural network from scratch in PyTorch on a local RTX 3060. The goal is a Seq2Seq transformer that translates English into symbolic commands, scored by a critic network to ensure deterministic execution.

This is what I mean when I talk about the Accountability Layer. You cannot secure AI agents if you give them the keys to the kingdom. You have to build the fortress around them.

I wrote a full 1-page technical case study on the architecture and the cryptographic audit trails. If you're building agents and care about security, I'd love to get your feedback.

How are the rest of you handling the "hallucination to execution" risk right now? Are you relying on human-in-the-loop, or are you trying to automate the verification layer?