Every model we deliver to a puck is cryptographically signed. The puck trusts exactly one thing: bytes that verify against our public key. Here is that key — and its hash — so you can check it yourself.
SHA-256 over the 32-byte raw Ed25519 public key. This is the single value that anchors our signed-delivery chain. Match any signed pack against it; nothing else is trusted.
Thirty seconds, no tools of ours required. Hash the published public key and confirm it equals the value above:
# Python — hash the raw public key python3 - <<'EOF' import hashlib raw = bytes.fromhex("06e86c3a3434ab20005657d89e09b833625a6c8b9ed70c1d1dc6d8010e880df6") print(hashlib.sha256(raw).hexdigest()) EOF # -> 811f732421c51647a97d58d2535eaeacd1b49402a0d755a7116758d8bbc1f7c7
Given any model pack and its .sig, verification uses only this public key — exactly what the puck does. A tampered or unsigned pack simply fails:
# Python — verify a pack against the published key
python3 - <<'EOF'
from cryptography.hazmat.primitives.serialization import load_pem_public_key
from cryptography.exceptions import InvalidSignature
pub = load_pem_public_key(open("model_signing.pub","rb").read())
data = open("pack.bin","rb").read(); sig = open("pack.bin.sig","rb").read()
try: pub.verify(sig, data); print("VALID - chains to AuraVision's root")
except InvalidSignature: print("INVALID - refused")
EOF
The hardware layer anchors secure boot in one-time-programmable fuses. We publish the expected values before the silicon is burned — so the commitment is public before any unit exists, and we cannot quietly swap it. When you hold a puck, read its fuses and match them against these.
RSA-3072 secure-boot root. NVIDIA tegra-fuse format (big-endian) — the exact value that gets burned into the PKC_PUBKEY_HASH fuse, computed from the published public key below with NVIDIA's own tooling.
PKC_PUBKEY_HASH — anchors the boot root of trust to the key above; the bootrom authenticates every boot-chain image against it.
JtagDisable — permanently closes the JTAG debug port so the fused root cannot be probed or overridden. One-way; never re-enabled.
SecurityMode — forces the bootrom to require a valid signature. Burned last; after it, nothing unsigned boots.
Recompute the expected fuse value from the published public key, using NVIDIA's own tool from the public Jetson BSP — you don't take our word for the hash:
# NVIDIA tegrasign (public Jetson BSP) — hash the published PKC public key python3 tegrasign_v3.py --pubkeyhash pub.key pub.sha --key pkc_rsa.pub # -> tegra-fuse format (big-endian): 0x2af81016ec9ab994a64194564d864b59daacd18c9c5a94d61c...3c88
A published number is only worth something if you can read the fuse yourself. You can — through the SoC's immutable on-die BootROM, in recovery, with Continuum OS never loaded, so no AuraVision code can sit in the path:
# 1. Put your puck into NVIDIA Force Recovery (RCM). # The unit-specific entry ships in your box (not posted publicly, for obvious reasons). # 2. From your OWN host, using the public NVIDIA Jetson BSP (Linux_for_Tegra): sudo ./odmfuseread.sh -i 0x23 <your-board> # 3. It reads the fuse values directly via the on-die BootROM — Continuum never runs. # Match the PKC_PUBKEY_HASH it prints to the value published above. # PKC_PUBKEY_HASH -> 0x2af81016ec9ab994a64194564d864b59daacd18c9c5a94d61c417f2719d7efa9...3c88
Nothing of ours runs during that read: it is NVIDIA's silicon BootROM and your host. And we published the expected value before the fuse was burned, so we cannot have back-fit it to whatever a unit happens to report. Read it, match it — trust nothing we say. That is the whole loop: a value committed in public ahead of the silicon, read back through immutable hardware you control.
This is what "privacy by construction" has to mean to be worth anything: not a promise, a value you can check. The cloud asks you to trust a policy. We publish the key.