Most spatial-recognition systems collapse a recording into one number. One number is the confirmation. One number is the explanation. One number is the audit log. A single-embedding system has a hard time when the input is adversarial. It also has a hard time when an operator asks it to explain a decision. The number does not decompose. The system reports "yes" or "no" and nothing in between.

We chose a different path. The engine stack does not collapse. It computes forty-nine separate confirmations from one recording. Each confirmation is independent. Each confirmation is auditable. The aggregate is the confidence score. The agreement count is the decision.

The foundation is a modified-ArcFace network. The output is a (1, 512, 7, 7) feature tensor. That tensor is partitioned into a seven-by-seven grid of spatial regions. Each region is one of the forty-nine confirmations. Each region is computed in parallel. Each region holds its own pass-or-fail bit.

The lattice is the substrate of the confirmation. It is also the explanation. An operator looking at a confirmation sees forty-nine cells. Green cells agreed. Red cells did not. The pattern of agreement is the audit trail. There is nothing more to inspect. The decision is explainable by construction, not after the fact.

An attack on a single-embedding system needs to fool one number. An attack on the lattice needs to fool forty-nine numbers, each computed independently, each with its own spatial geometry, each from a different region of the input. The keys that fool one region do not fool the others. The attack surface multiplies. The cost of evasion goes from "one model inversion" to "forty-nine simultaneous model inversions, each agreeing." That has not been demonstrated in the literature against architectures of this shape.

Parallel evidence is the architectural answer to adversarial input. It is not a heuristic. It is not a confidence threshold. It is a structural property of the computation. The lattice is the confirmation. The lattice is the defense.

The substrate carries no person. It carries no recording. It carries no template. What the registry holds is a derived shape: the seven-by-seven grid of confirmations from one spatial recording, encrypted with a post-quantum hybrid handshake, stored on partner infrastructure that the partner controls. The person is never stored. The person is only confirmed against.

This is the architectural commitment. There is no credential store to compromise. There is no shared secret to leak. The breach surface that drives most enterprise compromises does not exist in this architecture. That is the point of the engine.

Two things change if the lattice is the substrate of confirmation. One: the credential store is no longer on the threat model. There is nothing to steal. There is nothing to leak. Two: the explainability story is no longer a separate workstream. The lattice is the audit trail. The operator sees the same shape the engine saw. The decision is decomposable into forty-nine independent confirmations. Compliance does not require a separate ML-interpretability layer.

The cost of switching is real. Deployments live on partner infrastructure. The partner controls the integration. TASCET does not sell direct. The engine is deployed by cleared partners under existing certification regimes. If your organization is already running a partner channel for security infrastructure, the integration is short. If not, the partner conversation is the first step.