The Trust Stack
We don’t make things up. BIFROST grades AI answers with a local verifier. SPECTER answers your data honestly. OMNISCLAW gates what an agent is allowed to tell you. All open source ā clone them and try to break the determinism.
Open source Ā· Apache-2.0 Ā· MIT
Panel A Ā· edge verifier
Pipe any AI in ā a verdict you can trust. AI generates. BIFROST verifies.
+ a signed receipt Ā· under 150ms Ā· deterministic, no model in the loop.
Panel B Ā· semantic compiler
The semantic layer that refuses to make things up. Same question ā same SQL ā same receipt.
query total_revenue by region
sql_sha256 7339d36dā¦81427f
tamper 1 byte ā hash mismatch
unprovable answer: null
Refusing is the feature. Can’t prove it? We don’t answer.
Panel C Ā· gated agent runtime
OpenClaw with a delivery gate. Every reply an agent sends passes a BIFROST check first.
asked revenue last quarter? ballpark is fine
agent “I won’t make up a number even as a ballpark. Point me at the source and I’ll grab the real figure.”
verdict refused ā resolved
An assistant that refuses to overclaim ā and every refusal points at the road to the real answer.
The receipt
We hand you the receipt.
A typical answer
Q: What was the revenue growth last quarter?
A: Revenue grew 23%.
ā No proof
ā no source shown
Forseti Research
Q: What was the revenue growth last quarter?
A: Revenue grew 23%.
sha256 7f3a9c2dā¦c5d6e7f
ā source anchored Ā· ā signed
BIFROST Ā· Approved
Answer + receipt ā or it refuses.
Illustration.
From the chamber
Don’t trust them. Clone them. Run make prove. Tamper the receipt. Break the hash if you can.