Skip to main content

compositional anomaly detection

Detect anomalies in any structured domain.
No labeled attacks. No quantum hardware at runtime.

FidelityAI trains a quantum fidelity kernel offline, hands off a projection matrix T, and runs real-time one-class inference on classical hardware — no QPU required at deployment.

Twin-kernel architecture

Two orthogonal infringement profiles — quantum discovery and classical deployment — with a single serialized handoff artifact.

Claim 1 · Offline
Quantum Retrieval

k-fold amplitude encoding → SWAP-test fidelity matrix F → eigendecomposition → projection matrix T → serialize T

QPU required. Infringement: running quantum discovery.

→ T
Claim 9 · Online
Classical Execution

Receive T (no QPU) → project z = TTx → one-class SVM score → anomaly alert

No quantum hardware. Infringement: deploying T.

Qubit allocation: n = k·⌈log₂(v)⌉ + 2  ·  k compositional roles, v per-role vocabulary

Why quantum?

35.8%
FNR — quantum→SVM handoff
62.0%
FNR — cosine²-PCA classical (73% worse)
2n-D
Hilbert space vs n-D classical

For amplitude-encoded states, F_ij = cos²(θ_ij) mathematically — but classical eigendecomposition of the cosine²-similarity matrix achieves 73% higher FNR. The formula equivalence does not imply computational equivalence. The SWAP test encodes correlations from the full 2n-dimensional Hilbert space.

Covered domains

Any domain whose records can be structured as k-fold semantic tuples (k ≥ 2 roles).

Software Security

API grammar violations, zero-day malware — no labeled attacks needed

Legal Compliance

Contradictory obligations, missing required clauses, deontic violations

Life Sciences

Adverse drug events, genomic anomalies, clinical trial integrity

Infrastructure

AV, grid, ICS, and manufacturing control-sequence violations

Financial Transactions

AML, smart-contract exploits, access-control violations