P258 — AIEP — Deterministic Multimodal Ingestion Appliance

Applicant: Neil Grassby Classification: Patent Application — Confidential Priority: Claims priority from GB2519711.2 filed 20 November 2025 Architecture Layer: AIEP Phase 2 Data Ingestion Layer

Framework Context

[0001] This specification operates within an AIEP environment as defined in GB2519711.2 and GB2519798.9. The present specification defines a hardware-software appliance for the deterministic ingestion of multimodal data signals into the AIEP evidence ledger, providing a trusted, auditable entry point for sensor, document, and API data.

Field of the Invention

[0002] The present invention relates to deterministic multimodal data ingestion appliances for evidence-bound AI system data entry.

Background

[0003] Evidence-bound AI systems require that ingested data carry trustworthy provenance from the point of entry. Data arriving through general-purpose software stacks is vulnerable to processing artefacts, format normalisation inconsistencies, and logging gaps that reduce provenance trustworthiness. A purpose-built ingestion appliance that processes multimodal inputs (text, sensor, structured data, images) with deterministic normalisation and timestamp-locked admission provides a stronger provenance guarantee than software-only ingestion.

Summary of the Invention

[0004] The invention provides a Deterministic Multimodal Ingestion Appliance (DMIA) comprising: multimodal signal handlers for text, structured data, sensor streams, and image/video; a deterministic normalisation pipeline with fixed-version codec registry; a hardware timestamp unit providing GPS-disciplined timestamps with nanosecond precision; an admission hash engine computing content-addressable hashes in hardware; an ingestion receipt generator producing signed ingest records; and an evidence ledger write interface admitting processed artefacts to the AIEP evidence ledger.

ASCII Architecture

External Data Sources
  (text, sensors, APIs, images)
         |
         v
+------------------------------------------+
| Deterministic Multimodal Ingestion       |
|   Appliance (DMIA)                       |
|                                          |
|  Multimodal Signal Handlers             |
|  Deterministic Normalisation Pipeline   |
|  Hardware Timestamp Unit (GPS-locked)   |
|  Admission Hash Engine (hardware)       |
|  Ingestion Receipt Generator            |
+-------------------+----------------------+
                    |
                    v
  Evidence Artefact + Ingest Receipt
  → AIEP Evidence Ledger Admission

Detailed Description

[0005] Multimodal Signal Handlers. Separate handlers process each modality: text (encoding normalisation, language detection, format stripping); structured data (schema validation, canonical serialisation); sensor streams (unit normalisation, sampling rate standardisation); image/video (lossless format conversion, metadata extraction).

[0006] Codec Registry. The normalisation pipeline uses a versioned codec registry. Codec versions are fixed at appliance initialisation and recorded in every ingestion receipt. This ensures deterministic normalisation across identical inputs.

[0007] Hardware Timestamps. The hardware timestamp unit disciplines its clock to GPS time, providing provably accurate timestamps independent of system clock drift. Each ingested artefact receives a hardware timestamp.

[0008] Ingest Receipt. Each ingest event produces a receipt recording: input source identifier; signal handler version; codec registry version; GPS hardware timestamp; content hash; and appliance identity signature. The receipt is admitted to the ledger with the artefact.

Technical Effect

[0009] The invention provides deterministic, multi-party verifiable evidence artefact production from heterogeneous input sources, enabling the evidence basis of AI reasoning to be traced to auditable ingest events with hardware-certified timestamps. By fixing the codec registry at appliance initialisation and recording codec versions in every receipt, the appliance produces deterministically reproducible normalisation results from identical raw inputs, enabling replay verification of ingest events. GPS-disciplined hardware timestamps provide tamper-resistant temporal evidence independent of system clock manipulation, critical for regulatory accountability of evidence admission times.

Claims

1. A method of deterministic multimodal evidence ingest for an evidence-bound artificial intelligence system, comprising the steps of: (a) receiving an inbound signal and routing it to the appropriate modality handler: text handler (encoding normalisation, language detection, format stripping); structured data handler (schema validation, canonical serialisation); sensor stream handler (unit normalisation, sampling rate standardisation); or image/video handler (lossless format conversion, metadata extraction); (b) applying normalisation using a versioned codec registry fixed at appliance initialisation, with codec versions immutable during a deployment epoch and recorded in every ingest receipt; (c) assigning a hardware timestamp to the normalised artefact, the timestamp disciplined to GPS time independently of the system clock; (d) producing and signing an ingest receipt containing: input source identifier; signal handler version; codec registry version; GPS hardware timestamp; content hash of the normalised artefact; and appliance identity signature; (e) admitting the signed ingest receipt to the evidence ledger together with the normalised artefact.

2. The method of claim 1, wherein the appliance identity signature in the receipt is produced under a key bound to a hardware attestation module, enabling receiving nodes to verify appliance authenticity.

3. The method of claim 1, wherein the codec registry is versioned with a monotonically increasing epoch number, and appliances operating in the same deployment epoch use the same registry version, ensuring consistent normalisation across the federation.

4. The method of claim 1, wherein ingest receipts include the raw-input hash in addition to the normalised-artefact hash, enabling detection of normalisation codec regression when the same raw input is re-ingested with a different codec version.

5. The method of claim 1, wherein GPS-disciplined timestamps include a confidence interval derived from GPS signal quality measurement, enabling receiving nodes to assess temporal evidence reliability.

6. A deterministic multimodal ingestion appliance for an evidence-bound artificial intelligence system, comprising: modality-specific signal handlers; a versioned fixed codec registry; a GPS-disciplined hardware timestamp unit; an ingest receipt constructor and signer; and a ledger admission module.

7. A computer-readable medium carrying instructions for implementing the method of any preceding method claim.

Abstract

A deterministic multimodal ingestion appliance for evidence-bound artificial intelligence routes inbound signals to modality-specific handlers for text, structured data, sensor streams, and image/video, applying normalisation using a versioned codec registry fixed at appliance initialisation. A GPS-disciplined hardware timestamp unit provides provably accurate timestamps independent of system clock drift. Each ingest event produces a signed receipt containing input source identifier, handler and codec versions, GPS timestamp, and content hash, admitted to the evidence ledger together with the normalised artefact, enabling deterministic replay verification of all ingest events.