InferTrust™ is built on five distinct innovations, each addressing a specific gap that existing AI logging and observability tools leave open. Here is what each
InferTrust™ is not a better logging tool. It is a fundamentally different approach to AI decision accountability, built on five patent-pending innovations that work together to create verifiable proof at the moment of inference.
The decision record is signed the instant inference completes, before any network call, UI update, or app-layer process can observe or alter the output. This eliminates the "tamperable window" that exists in every standard logging architecture, the gap between when the AI produces its output and when that output gets recorded.
Why it matters: In litigation or regulatory review, the question is not whether you logged the decision but whether you can prove the log was not modified. Signing at inference time creates a record that carries its own proof of authenticity.
A protected zone isolates AI outputs between inference and signing. No logging layer, middleware, or UI thread can access the raw values during this window. The execution boundary ensures that the signed record reflects exactly what the model produced, with no opportunity for application code to alter the output.
Why it matters: Standard architectures give application code full control over what gets logged. The execution boundary removes this control, ensuring the record is a faithful representation of the model's actual output.
The signing key lives in tamper-resistant hardware such as a TPM or Secure Enclave. It cannot be extracted, copied, or used from another machine. This means every signed record is cryptographically bound to a specific physical device, proving which hardware actually ran the inference.
Why it matters: Server-side logs cannot prove origin. A fabricated record transmitted from any source looks identical to a genuine one. Device-bound keys solve this by tying every signature to verified hardware.
Every inference gets a strictly incrementing number baked into the signed record. The counter never decreases and never repeats. If a record is deleted, the gap in the sequence is immediately visible. If records are reordered, the sequence exposes it.
Why it matters: Traditional logs can have records deleted without a trace. The sequence counter makes deletions detectable, proving that the audit trail is complete and unbroken.
Model version, confidence score, policy hash, input feature hash, timestamp, decision action, and sequence number are all bound in one cryptographic signature. This means no individual field can be altered without invalidating the entire record. You cannot change the confidence score without also invalidating the model version, timestamp, and every other field.
Why it matters: Systems that sign individual fields or generate separate proofs for different attributes leave room for partial manipulation. A single combined signature makes the record all-or-nothing: authentic in its entirety, or detectably tampered.