L3: Custody & Chain of Evidence

L3 provides chain-of-custody tracking for data and model artifacts throughout their lifecycle. Every piece of data that enters the AI pipeline and every model version used to process it has a verified, tamper-evident custody record -- who touched it, when, and what changed.

Why Chain of Custody?

In regulated industries, the question is never just "what did the AI decide?" -- it is "what data did it use, where did that data come from, who had access, and can you prove none of it was tampered with?"

L3 answers these questions with cryptographic certainty.

Capabilities

Data Lineage Tracking

L3 records the full lineage of data as it flows through the AI pipeline:

Data Lineage: record_id=DL-2026-04-10-00847
  Origin:     CRM export (Salesforce, 2026-04-10T08:12:00Z)
  Ingested:   Pipeline intake (user: j.martinez, 2026-04-10T08:14:22Z)
  Transform:  PII redaction applied (policy: HIPAA-SAFE, 2026-04-10T08:14:23Z)
  Processed:  Model inference (model: gpt-4o, 2026-04-10T08:14:25Z)
  Output:     Delivered to dashboard (user: j.martinez, 2026-04-10T08:14:26Z)
  
  Integrity:  SHA-256 hash chain VERIFIED at all 5 stages

Every stage is logged with:

• Who -- the user or system that handled the data
• When -- precise timestamp
• What -- the operation performed
• Hash -- cryptographic hash of the data at that stage

Model Artifact Custody

Tracks the custody of model artifacts -- weights, adapters, configurations, and serving endpoints:

Artifact	Tracked Fields
Model weights	Version, hash, storage location, access log
LoRA adapters	Training data reference, base model, merge history
Configuration	Temperature, system prompts, safety settings
Serving endpoint	Provider, region, API version, deployment timestamp

Regulatory Requirement

Under emerging AI regulations (EU AI Act, NIST AI RMF), organizations must demonstrate traceability of AI system components. L3 provides this traceability out of the box.

Evidence Chain Integrity

L3 uses hash-chained records to ensure evidence integrity. Each custody record includes a cryptographic hash of the previous record, creating an immutable chain. Any attempt to alter, delete, or insert records breaks the chain and is immediately detectable.

Record N:   hash(data_N + hash_N-1) = 0x8a3f...
Record N+1: hash(data_N+1 + 0x8a3f...) = 0x2b71...
Record N+2: hash(data_N+2 + 0x2b71...) = 0x9c04...

Tamper detection: If Record N is modified, hash_N changes,
which invalidates Record N+1, N+2, and every subsequent record.

Custody Transfer Logging

When data or artifacts are transferred between systems, teams, or providers, L3 records the transfer with:

• Source and destination identifiers
• Transfer authorization (who approved it)
• Data classification at time of transfer
• Encryption status during transit
• Acknowledgment of receipt

Custody Gaps

If a data asset appears in the pipeline without a verified origin record, L3 flags it as a custody gap. Custody gaps indicate data that entered the system outside governed channels and should be investigated.

Console Features

Lineage Explorer

Visual lineage graph showing the full path of any data asset from origin to output. Click any node to inspect the custody record, hash verification, and associated metadata.

Artifact Registry

Browse all tracked model artifacts with their custody history. Filter by model provider, version, deployment status, or date range.

Integrity Verification

Run on-demand integrity checks against the hash chain. Results show:

• Total records in chain
• Verification status (INTACT / BROKEN)
• If broken: the exact record where integrity fails
• Last verified timestamp

Transfer Log

Chronological log of all custody transfers with filtering by source, destination, data classification, and authorization status.

Operating Modes

Mode	Behavior
Monitor	Custody records are created and hash-chained. Gaps are logged but not enforced.
Advisory	Custody gaps produce HELD verdicts. Data with unverified lineage is flagged for review.
Enforce	Data without verified custody is BLOCKED from entering the AI pipeline.

Related Layers

• L2: Reasoning Governance -- uses L3 lineage data to verify the provenance of facts used in reasoning chains
• L6: Model Governance -- L3 tracks custody of the model artifacts that L6 governs
• L4: Content Verification -- L3 provides the evidence trail for content verification decisions