0 / 4 steps done
Vault Proof Kit
Run the proof on your own machine.
Download the sealed eval image, run prove.sh, and arrive at your walkthrough call holding PROOF_REPORT.md. Ten minutes.
What the proof establishes
| What the proof establishes | How |
|---|---|
| Runs entirely inside your perimeter | Container starts and passes all checks with network isolated |
| Zero outbound egress | DNS lookup and TCP connections from inside the container both fail |
| All 16 LLM routes sealed | Every LLM-dependent route returns 503 — not stubbed, actually blocked |
| Approval history is intact | Hash chain on each wisdom entry returns PASS on clean data |
| Tamper detection works | Flipping one byte in the corpus causes the hash check to FAIL |
| Real quality scoring happens offline | POST /analyze returns scored output with no network |
Requirements
🐳 Docker Desktop 4.x · 4 GB RAM⌨ curl and bash (macOS/Linux) or Git Bash (Windows)⏱ 10 minutes📡 Internet optional after download
Walkthrough
Download these four files from https://vault.syswisdom.ai/proof/ into a single working directory on your machine. Keep them together — the proof script references all four by name.
ai-quality-agent-wisdom-v1-eval.tar # the sealed agent image (~1.5 GB) prove.sh # the proof script you run prove_checks.py # helper called by prove.sh SHA256SUMS # checksums to verify the download
What PROOF_REPORT.md looks like
PROOF_REPORT.md
## Machine
PlatformmacOS 14.2 (Darwin 23.2.0)DockerDesktop 4.27.2 (Engine 25.0.3)Archarm64Run IDc3f4a8b2-7d1e-4f09-a2c6-8b5e3d0f1a4eTimestamp2025-01-15T10:23:44Z
## Results
Check 1: image digest matches release-eval.jsonPASS
Check 2: container starts with --network nonePASS
Check 3: vault-doctor 6/6PASS
Check 4: egress blocked - DNS lookupPASS
Check 5: egress blocked - TCP to 1.1.1.1:443PASS
Check 6: all 16 sealed routes return 503PASS
Check 7: /vault/verify hash chainPASS
Check 8: tamper detectionPASS
Check 9: /analyze returns scored resultPASS
Check 10: evidence savedPASS
## Known Gaps
Windows Docker Desktop: --network none may not support port publishing (recorded as known gap, not FAIL)
Eval corpus is demo wisdom — scoring demonstrates the mechanism, not a customer-specific policy
This proof covers the eval image; the customer image uses identical sealing with a different corpus
Ready to talk?
Complete the walkthrough above, then bring your PROOF_REPORT.md to a 30-minute conversation. The call starts at domain selection — not trust-building.
aj@syswisdom.aiNo form. No funnel. Email Aaron directly.