AI Data Processing & Model Non-Retention Addendum

Document Reference: AUC-AIG-ADD-001
Version: 2.4.0
Effective Date: October 24, 2026
Last Updated: October 24, 2026
Classification: Legal / AI Governance
Companion Documents: Master Services Agreement | Data Processing Addendum | Schedule 1: AI Service-Specific Terms | Sub-processor Schedule

Executive Summary (ISO 24495-1 Plain-Language Overview)

[!IMPORTANT] Legal Precedence Notice: This plain-language summary is provided solely for reader convenience and technical audit navigation pursuant to ISO 24495-1:2023. It does not constitute binding legal terms, warranties, or covenants. In the event of any ambiguity, conflict, or inconsistency between this summary and the operative articles below, the operative articles strictly govern and control.

• Absolute Zero-Training Guarantee: Under no circumstances will Auctoric, or our upstream model providers, train, retrain, or RLHF base models using your prompts, completions, context embeddings, or agent traces.
• Ephemeral Processing Enclaves: Real-time inference executes in volatile RAM enclaves and is purged immediately upon completion unless customer-side audit logging is actively configured.
• Customer Ownership of Adapters: You exclusively own all custom PEFT adapters, LoRA weights, and domain-specific vector embeddings generated on the platform.
• Rapid AI Incident Notification: Auctoric will alert you in writing within 24 hours of confirming any model extraction, inversion attack, or cross-tenant prompt injection breach.
• Data Sanitization SLA: Deprovisions and sanitizes custom model artifacts within maximum 60 days following termination per NIST SP 800-88 Rev. 2.

High-Intent Procurement FAQ (AEO Direct Answers)

Can Auctoric or upstream providers use customer prompts to train foundation models?

No. Under Section 1.1, Auctoric legally and technically guarantees that customer content is never used to pre-train, fine-tune, RLHF, or distill public or multi-tenant models. All upstream inference utilizes enterprise-tier zero-data-retention (ZDR) endpoints.

Who owns fine-tuned model weights and LoRA adapters created on Auctoric?

The Customer owns all custom adapters. Under Section 2.1, customer retains exclusive IP title, ownership, and interest in all parameter-efficient fine-tuning (PEFT) weights, LoRA matrices, and embeddings generated from customer data.

What is the reporting SLA for AI-specific security breaches and jailbreaks?

Under Section 6, Auctoric commits to notifying customer within twenty-four (24) hours of confirming any model inversion, weight extraction attack, or cross-tenant prompt injection breach.

This AI Data Processing & Model Non-Retention Addendum (“AI Addendum”) supplements the Master Services Agreement and Data Processing Addendum entered into between Auctoric, LLC (“Auctoric”) and the contracting customer entity (“Customer”).

This AI Addendum establishes legally binding safeguards and technical guarantees regarding the isolation, processing, non-retention, and intellectual property ownership of Customer Content within the Auctoric AI Platform.

1. Zero Model Training & Non-Retention Commitment

1.1 Strict Prohibition on Public Model Training

Auctoric contractually guarantees and warrants that:

1. No Foundation Model Training: Auctoric will NOT use Customer Content (including prompts, inputs, vector embeddings, context attachments, synthetic data, agent execution traces, or generated outputs) to train, retrain, pre-train, RLHF (reinforcement learning from human feedback), DPO (direct preference optimization), or distill any public, multi-tenant, or foundational AI models.
2. Third-Party Model Isolation: When Auctoric orchestrates inference via third-party foundation model APIs (e.g., as detailed in the authoritative Sub-processor Schedule), Auctoric utilizes enterprise-tier zero-data-retention (ZDR) endpoints where providers are bound not to train on Customer inputs or outputs.

1.2 Ephemeral Processing Enclaves

By default, real-time inference processing operates in volatile memory enclaves. Customer Content (prompts and completions) is not retained in persistent model storage following completion of the response, unless Customer explicitly configures persistent audit logging or custom storage buckets. Non-content operational telemetry (token counts, latency, and cryptographic audit hashes) is retained for billing and security monitoring.

2. Customer Fine-Tuned Models & Custom Adapters

2.1 Ownership of Customization Artifacts

Where Customer engages Auctoric Services to train parameter-efficient fine-tuning (PEFT) adapters (e.g., LoRA, QLoRA) or generate specialized vector embeddings:

• Customer Ownership of Adapters: Customer retains exclusive intellectual property ownership, title, and interest in newly created fine-tuning adapters, LoRA matrices, and embeddings generated specifically for Customer from Customer Content.
• Underlying Provider IP: Pre-existing foundational models, base weights, and provider-managed model architectures remain subject to their applicable third-party license terms (such as Google Vertex AI Modified Model terms).
• Cryptographic Isolation: Custom model weights and adapters are encrypted with dedicated keys (including Customer-Managed Encryption Keys (CMEK) where supported). No other tenant can access or decrypt these weights.

2.2 Model Portability, Deprovisioning, and Sanitization

Upon thirty (30) days’ written notice or upon termination of the Services:

1. Customer may request an export of extractable, customer-owned parameter adapters (e.g., LoRA matrices) within thirty (30) calendar days;
2. Following such 30-day export window, or upon customer’s confirmed completion of export, Auctoric will permanently deprovision and sanitize all instances of the custom model artifacts within thirty (30) calendar days (total maximum 60 days from termination), in accordance with NIST SP 800-88 Revision 2 media sanitization standards.

3. Retrieval-Augmented Generation (RAG) and Vector Stores

3.1 Namespace Isolation

All vector embeddings, chunked document indexes, and semantic search stores maintained in the Auctoric Knowledge Engine are logically and cryptographically partitioned at the tenant database level.

3.2 Access Token Scoping

Downstream autonomous agents querying vector stores operate under strict OAuth 2.0 / RBAC tokens scoping queries only to Customer’s authorized organizational indexes.

4. Technical Safeguards & Architecture

flowchart LR
    subgraph CustomerEnv["Customer Environment"]
        App["Customer Application"]
    end

    subgraph AuctoricSecureEnclave["Auctoric Governed AI Enclave"]
        WAF["WAF & mTLS Gateway"]
        PII["Real-Time PII Masking Filter"]
        Engine["Inference Orchestration Engine"]
        Memory["Volatile RAM Enclave (Purged Post-Inference)"]
        Guard["Safety & Toxicity Guardrail"]
    end

    subgraph FoundationModel["Isolated Inference Endpoint (Zero-Retention)"]
        Model["Frozen Weights (Read-Only)"]
    end

    App -->|TLS 1.3 Request| WAF
    WAF --> PII
    PII --> Engine
    Engine <--> Memory
    Engine --> Guard
    Guard -->|Inference Call| Model
    Model -->|Generated Tokens| Guard
    Guard --> Engine
    Engine -->|TLS 1.3 Response| App

4.1 Cryptographic Controls

• All network traffic is encrypted in transit using TLS 1.3 with forward secrecy.
• Static data at rest is encrypted using AES-256 with automated key management.

4.2 Automated PII Redaction

Customer may activate Auctoric’s real-time PII Redactor, which masks credit card numbers, Social Security Numbers, phone numbers, and protected health information (PHI) before prompts hit downstream model inference.

5. Control Verification and Compliance Alignment

Auctoric designs and operates its technical safeguards to comply with the criteria established by ISO/IEC 42001:2023 (AIMS), ISO/IEC 27001:2022, and SOC 2 Type II (Security and Privacy Trust Services Criteria). Auctoric maintains documented policies, evidence logs, and testing summaries verifying adherence to these non-retention and data isolation commitments.

6. Model Inversion & AI Security Incidents

In addition to the breach notifications specified in the DPA, Auctoric will notify Customer within twenty-four (24) hours of becoming aware of:

1. Any verified model extraction, inversion, or gradient reconstruction attack targeting Customer’s dedicated models or vector embeddings;
2. Any confirmed prompt injection exploit that bypassed perimeter guardrails and exfiltrated Customer Content across tenant boundaries.

7. Form of Execution & Counterparts

(This standard form AI Addendum is incorporated into and executed via Customer Order Form, bilateral electronic signature, or digital enrollment through the Auctoric Trust Center.)