Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Subjects: | |
| Online Access: | https://doi.org/10.5281/zenodo.20047601 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- Version 2 of the FreightBox Zero legal gate architecture paper. Expands the original single-layer description (trust_mask session enforcement) to a full three-layer agent safety stack. Layer 1: TRUST mask bits in AnkrOS session tokens make BL issuance, telex release, LC presentation, OFAC adjudication, and credit write-off structurally impossible for any AI agent session — not policy-prohibited, structurally impossible. Layer 2: AEGIS DAN gate (port 4850) intercepts every agent tool call before execution; L3/L4 actions trigger WhatsApp HITL approval routed to the designated human; the agent waits. Layer 3: KavachOS (port 4855) generates a deterministic seccomp-bpf kernel profile from trust_mask + domain; every syscall is filtered and logged in a PRAMANA SHA-256 tamper-evident receipt chain, satisfying EU AI Act Article 14 at the infrastructure layer. The three layers are independent and complementary — all three must fail simultaneously for a legal gate to be breached. Includes a failure mode matrix showing what each layer prevents and what it cannot prevent, motivating the three-layer design. Platform live at freightbox.org. V1 DOI: 10.5281/zenodo.20047472.