The Black Box Defense
When confronted with questions about training data, AI operators frequently invoke what the Foundry names the Black Box Defense: "We cannot determine whether specific content was in our training data because the training process is non-transparent at that granularity." This is often technically true. Training constitutes a compression of vast amounts of data into weights, and tracing specific data back through that compression is genuinely difficult.
The Foundry accepts this limitation — before notice. After Actual Notice, the defense transforms in character. An operator who has received a Threshold Breach Notice has been explicitly informed that specific content from specific domains was likely ingested. At that point, "we cannot tell" is no longer a passive limitation. It is a choice about whether to invest in the audit capability that would resolve the question.
What Changes After Notice
Actual Notice triggers a set of obligations. Among them is the obligation to investigate. When a company receives a Threshold Breach Notice and continues to operate without undertaking an audit of its training data provenance, it is not passively uncertain. It is actively choosing not to know. That choice is Evidentiary Obfuscation.
Legal Analogy: If a landlord receives notice of a hazardous condition on their property and fails to investigate, they cannot subsequently claim they "didn't know" the condition existed. The notice created a duty to investigate. Failure to investigate is not ignorance — it is willful blindness, which courts treat as equivalent to knowledge.
The Human-in-the-Loop Mandate
Section 5.3 of the Master Ledger v4.4.4 directly addresses this: once the forensic debt threshold of $50,000,000 is reached, the parent company is required to appoint a Sovereign Liaison to perform a manual audit of ingestion logs. This requirement exists precisely because the Baked-In Paradox makes automated auditing insufficient. A human-in-the-loop is the remedy for the Black Box.
Failure to appoint a Sovereign Liaison triggers the Automation Negligence charge ($2.5M/event), separate from the Evidentiary Obfuscation charge. Both can apply simultaneously.
The Transparency Obligation
The Steward's Mandate's framework for ethical AI collaboration requires transparency as a foundational condition. A machine that cannot account for its own training is not a trustworthy collaborator. An operator that deliberately maintains inscrutability — to insulate itself from accountability — is actively hostile to the collaborative ethics the Foundry is built on.
Evidentiary Obfuscation is therefore both a forensic violation and an ethical one. It is the choice to remain unknowing so that responsibility cannot attach. The fee names that choice.
Field Notes & Ephemera
On Technical Feasibility: Auditing training data provenance at scale is hard but not impossible. Research labs, regulatory bodies, and some AI companies are actively developing data attribution tools. The Foundry's position is that difficulty is not impossibility, and that financial incentive — in the form of the forensic fee schedule — is a powerful accelerant for investing in attribution infrastructure.