Seven fields, cryptographically bound together by one digital signature. Here is what each field captures, why it matters, and how they work together to create
When a human professional makes a consequential decision, there is a paper trail: who made the call, when, what information they reviewed, and what rules governed the decision. AI systems making equivalent decisions should carry equivalent proof.
InferTrust™'s signed decision record captures seven fields in a single cryptographic signature. Together, they answer every question an auditor, regulator, or court might ask about an AI decision.
A cryptographic hash of the exact model binary that produced the inference. This proves which version of the AI made the decision, not just which model name was deployed, but the precise weights, architecture, and training state that generated the output.
The model's reported confidence at inference time, recorded to six decimal places. This value drives confidence-gated decision-making: whether the AI acts autonomously or escalates to a human reviewer. Recording it cryptographically proves the model's certainty at the moment of decision.
A hash of the organizational policy ruleset active when the decision was made. Policies change over time, and proving which rules governed a specific decision is critical for compliance. If a policy was updated after an adverse event, the hash proves whether the old or new policy was in effect.
A hash of the input data features the model processed. This links the decision to its inputs without storing raw data. For privacy-sensitive applications like healthcare imaging or financial records, the hash proves what the model saw without exposing the underlying data.
The exact time of inference, cryptographically bound to the record. Unlike traditional timestamps that can be backdated, this timestamp is sealed inside the signature. Altering it invalidates the entire record.
The specific action the model recommended or executed: APPROVE, DENY, ESCALATE, FLAG, or DEFER. This is what the AI actually decided, recorded before any downstream process can modify or override it.
A strictly incrementing counter. Every inference gets the next number in sequence. Gaps reveal deleted records. Out-of-order entries reveal tampering. The sequence proves the audit trail is complete and unbroken.
All seven fields are bound together by a single cryptographic signature using a device-bound hardware key. This means changing any single field invalidates the entire record. You cannot alter the confidence score without also invalidating the model version, timestamp, and decision action. The record is authentic in its entirety, or detectably tampered.