In under eight minutes: record an agent run from genesis to seal, trigger a fault, watch the policy engine catch and contain it on-record, then replay the tamper-evident chain and produce a regulator-ready audit certificate — all without trusting BLACKBOX itself.
The full system — anatomy of a sealed run, cryptographic proof, governance stakeholders, and audit features — in one view.
Real data. Real runs. Captured from promptblackbox.com — not mockups.
"The aviation black box doesn't make planes safer by itself. It makes crashes survivable as information.BLACKBOX does the same for AI agents."
hash(n) = SHA-256( hash(n−1) + canonical_JSON(event(n)) )Each node commits to every node before it. The FAULT event is sealed into the chain — not quietly removed. The SEAL event closes the run. Any post-hoc alteration to any event breaks the chain from that point forward, verifiable with stdlib sha256 alone.
Two lines of Python. Zero infrastructure changes. The first run is free.
pip install blackbox-sdk import blackbox_sdk as bb bb.configure(api_url="https://your-blackbox.railway.app") bb.instrument() # every LLM call now recorded automatically