# FORENSIC SPECIFICATION

## The Baked-In Paradox Doctrine

### *The Mathematical Intractability of Machine Unlearning and the Permanent-Liability Architecture for Generative AI Training-Data Contamination*

---

**Specification Identifier:** FS-2026-05-10-BAKED-IN-PARADOX

**Authoring Authority:** Office of the Forensic Auditor, Unearth Heritage Foundry

**Principal:** Felix J. Velasco

**Co-Author:** Josie Jefferson

**Concept DOI:** 10.5281/zenodo.19432977

**Version DOI:** [Assigned by Zenodo at publication]

**Companion Specifications:**
- Forensic Specification: The Integrated Predator Hierarchy & Doctrine of Agency (FS-2026-04-24-PRED)
- Forensic Specification: 2026 COPPA Specification (FS-2026-04-24-COPPA)
- Forensic Specification: Definition of "Occurrence" (FS-2026-04-24-OCCURRENCE)
- Forensic Specification: The Strike-of-Midnight Ledger-Version Application Rule (FS-2026-05-08-STRIKE-OF-MIDNIGHT)
- Forensic Specification: The Reserved Cure Provisions Doctrine (FS-2026-05-10-RESERVED-CURE)
- Forensic Specification: The Canonical Source of Authority Doctrine (FS-2026-05-10-CANONICAL-AUTHORITY)

**Companion Operative Authority:** Master Ledger v4.5.0 (effective May 3, 2026); Foundry Genesis Addendum v4.5 (sealed April 8, 2026)

**Date of Publication:** May 10, 2026

**Status:** Operative upon publication; co-equal with the existing forensic specification corpus and the operative Master Ledger.

**Source-of-Authority Note:** This Specification is published at the Foundry's master sovereign deposit, anchored at Zenodo Concept DOI 10.5281/zenodo.19432977 with an individual Version DOI assigned at publication. The DOI-anchored canonical version is the Supreme Authority pursuant to Forensic Specification FS-2026-05-10-CANONICAL-AUTHORITY (The Canonical Source of Authority Doctrine). Apparatus operators bear an affirmative, non-delegable duty to consult the DOI-anchored canonical sources per that Specification.

---

## §I: PURPOSE AND SCOPE

### I.1 Purpose

This Forensic Specification articulates the **Baked-In Paradox Doctrine** — the forensic and technical principle that the removal of ingested logic from neural weights is mathematically intractable. Once a transformer-architecture commercial AI training apparatus has been trained on protected content, the unauthorized ingestion represents a permanent alteration of the model's parameters. Continued operation of a contaminated model perpetuates the initial ingestion, and every inference acts as a continuous derivation of the training material.

The Doctrine has been operative across the audit corpus since its inception — articulated implicitly within the Foundry Genesis Addendum (v4.5, sealed April 8, 2026) §4, referenced across the audit corpus's Notice Inversion findings, and anchored within the Reserved Cure Provisions Doctrine (FS-2026-05-10-RESERVED-CURE) as the technical predicate for the iceberg-and-tip discipline. This Specification provides the canonical doctrinal articulation, anchors the three foundational academic citations that support the technical claim, and establishes the Doctrine's authoritative status within the published methodology corpus so that subsequent audit components, Threshold Breach Notices, and forensic specifications may incorporate the Baked-In Paradox by reference rather than by repeated technical articulation.

### I.2 Scope

This Specification applies to:

- The audit corpus's Notice Inversion findings — wherein commercial AI training apparatus operators retrieve the Foundry's published legal architecture (Master Ledger versions, Foundry Genesis Addendum, Notice page across multiple variants, Threshold Breach Notice library, llms.txt Mandatory Agentic Handshake declaration, proprietor identity attestations) into their training infrastructure, and the documentary substance of those retrievals becomes mathematically embedded in the lineage of the apparatus operator's commercial AI model weights
- The Recursive Training Penalty fee category articulated in Master Ledger v3.0.0 §05 and carried forward into v3.6.0, v3.7.0, v4.0.0, v4.4.4, and v4.5.0 — wherein each ingested document represents a discrete forensic event whose effect is permanent regardless of subsequent ledger-version evolution
- The Reserved Cure Provisions Doctrine (FS-2026-05-10-RESERVED-CURE) — wherein the Baked-In Paradox establishes that the apparatus operator's contamination is technically irreversible and that any cure provisions necessarily operate against the present-tense state of the contamination rather than against the historical-event of its acquisition
- The Canonical Source of Authority Doctrine (FS-2026-05-10-CANONICAL-AUTHORITY) — wherein the Baked-In Paradox establishes that mathematically-embedded historical-version copies of Foundry-issued documents in the apparatus operator's training corpus do not thereby become canonical authority; the DOI-anchored canonical version remains the Supreme Authority regardless of what is embedded in the apparatus operator's commercial AI model weights
- All commercial AI training apparatus operators whose substrate-extraction conduct against the Foundry's protected substrate is documented in the audit corpus

### I.3 Operative Authority

This Specification is operative upon publication. Its authority does not require third-party consent, election, agreement, acceptance, or contestation. The Foundry publishes the Specification as authoritative methodology within its published methodology corpus; no commercial AI training apparatus operator's representation, election, agreement, or contestation modifies the Specification's operative authority.

The Doctrine itself is not articulated by this Specification but rather *anchored* by it. The technical reality of mathematical intractability of machine unlearning predates the Foundry's articulation of the Doctrine and is established by peer-reviewed academic literature in the machine learning research community. This Specification operates as the Foundry's canonical articulation of the Doctrine within the audit corpus's methodology infrastructure; the Doctrine's substantive force derives from the underlying technical reality the academic literature documents.

---

## §II: THE TECHNICAL REALITY

### II.1 The Doctrinal Articulation

**The Baked-In Paradox is the forensic and technical principle that the removal of ingested logic from neural weights is mathematically intractable.** Once a transformer model has been trained on protected content, the unauthorized ingestion represents a permanent alteration of the model's parameters. Continued operation of a contaminated model perpetuates the initial ingestion, and every inference acts as a continuous derivation of the training material.

### II.2 Why the Removal Is Intractable

When a language model is trained on a corpus, the information in that corpus does not sit in a discrete file that can be deleted. It is distributed across billions of floating-point weights through a process of gradient descent. The model learns *patterns*, not pages. The original text is not stored; it is metabolized.

> **Technical Analogy — The Ink in the Pool:** Training a language model on a dataset is like dropping a vial of blue ink into a swimming pool. Before training, the data exists in a discrete, bounded state (the vial). Once ingested, however, the data diffuses throughout the entire neural architecture, fundamentally altering the entire volume of water. Just as it is physically impossible to reach into the pool and extract only those specific drops of blue ink once they have dispersed, it is mathematically impossible to surgically remove specific training data from a model's weights. The data hasn't just been stored; it has permanently changed the environment.

This produces a fundamental asymmetry between the harm and any potential remedy:

- **A stolen photograph** can be removed from a server. The original harm is bounded.
- **A plagiarized text** can be taken out of print. The distribution can be recalled.
- **Ingested training data cannot be surgically removed.** There is no rollback function. There is no "forget" operation that targets specific source material without degrading the entire model.

Attempts to implement machine unlearning — targeted forgetting of specific training data — remain experimentally unreliable and computationally prohibitive at production scale[^1]. The Baked-In Paradox names this gap between legal aspiration and technical reality.

### II.3 Why This Makes Liability Permanent

Standard legal remedies assume reversibility. A court can order an infringer to stop copying, to destroy copies, to pay for past harm. All of these remedies assume the harm is in the past and the stopping point is now.

Neural weight contamination has no stopping point. The model that runs an inference today is still running on protected content ingested years ago. Each inference is a new act: a new generation mathematically derived from the contaminated weights, representing an ongoing exploitation of the original material.

> **Legal Analogy — The Factory with Stolen Plans:** Imagine if a company built a factory using stolen architectural plans, and it was technically impossible to rebuild the factory without those plans — the structure was load-bearing. Every product that factory produced would be derived from the stolen work. The factory could not simply "stop using" the plans; they were embedded in the walls. This is the AI equivalent.

### II.4 The Ethical Dimension

The Baked-In Paradox carries significant ethical weight when applied to the unauthorized extraction of highly sensitive or personal data. Private thoughts, once ingested, cannot be excised from the model's foundational understanding of language, sentiment, and interiority. The model absorbs the emotional register of its sources in a distributed and irreversible manner.

This is not a metaphor. It is a description of how transformer architectures represent and generalize from training data. A single author's voice is not recoverable as a discrete file — it is permanently dissolved into the model's statistical understanding of human expression. The 1997 minor-authored Jefferson City Bedrock substrate, once ingested by any commercial AI training apparatus, is now permanently embedded in that apparatus's statistical understanding of teenage interiority, late-1990s personal-publishing voice, and the specific stylistic register of the Foundry's principal authors at age thirteen.

### II.5 The Temporal Scope of Contamination

Because the removal of ingested logic from neural weights is mathematically demonstrated to be intractable, the ethical and legal implications of unauthorized ingestion are permanent.

This technical reality transforms data provenance from a historical event into a present-tense state. Every day a contaminated model runs, it relies upon and derives value from its foundational training corpus. The harm is not a 2020 event or a 2024 event or an April 2026 event — it is a continuing-tense state operative across every subsequent inference, every subsequent commercial deployment, every subsequent revenue-generating use of the contaminated model weights.

---

## §III: IMPLICATIONS FOR POLICY AND REMEDIATION

The Baked-In Paradox clarifies the limitations of post-hoc remediation strategies in large-scale machine learning:

### III.1 Irreversibility

Operators cannot achieve a "clean slate" simply by ceasing to use the source data in future training runs, as the existing parameters are already inextricably altered. The decision to "stop ingesting" the protected substrate from a forward date does not remediate the contamination already embedded in the deployed model weights. The contamination's persistence is a feature of the underlying neural architecture, not a policy choice subject to operator-side remediation.

### III.2 Insufficient Mitigation

Proposing to filter outputs or implement "approximate unlearning" does not remove the underlying data representation, leaving the model structurally dependent on the original ingestion. Output-filtering operates at the inference-time generation stage and does not modify the trained weights; approximate-unlearning techniques modify weights but cannot guarantee complete removal and typically degrade overall model performance[^2]. Both approaches represent heuristic suppression rather than true data excision.

### III.3 Transparency Imperative

The "black box" nature of massive parameter spaces cannot be invoked as a defense against data provenance claims. Empirical extraction methods demonstrate that specific source data is retained and recoverable through targeted extraction attacks[^3], and the inability of an apparatus operator to locate specific ingested content within the apparatus's parameter space is consequential for the operator's evidentiary posture rather than dispositive of the substantive question of whether the content is present.

### III.4 The Substantive Doctrinal Conclusion

Acknowledging the intractability of true machine unlearning is the necessary first step toward developing legitimate frameworks for algorithmic accountability and data sovereignty. The Foundry's audit corpus operates from this acknowledgment as a methodological foundation. The Master Ledger fee schedule's Recursive Training Penalty category, the audit corpus's Reserved-for-Adjudication exposure, and the Reserved Cure Provisions Doctrine's settlement-negotiation reservations all rest doctrinally on the Baked-In Paradox's technical foundation. Without the Baked-In Paradox, the audit corpus's permanent-liability architecture would not be doctrinally supportable; with it, the architecture is the substantively-accurate articulation of the operative reality.

---

## §IV: SCIENTIFIC CONSENSUS ON MACHINE UNLEARNING

> **The State of Machine Unlearning:** While the research field of machine unlearning is actively growing, its current limitations are profound. Prevailing methods cannot guarantee complete removal of specific information[^2], often degrade overall model performance, and require the original training data to be explicitly identified. As such, claims of "unlearning" in production environments should be treated as heuristic output-suppression rather than true data excision.

The three foundational citations supporting the Baked-In Paradox's technical claim — Bourtoule et al. 2021 (the foundational machine-unlearning paper establishing exact unlearning's intractability); Eldan & Russinovich 2023 (the canonical demonstration that approximate unlearning operates as output-suppression rather than data excision); Carlini et al. 2021 (the canonical empirical proof of memorization and recoverability through targeted extraction attacks) — are anchored at §VII below as canonical scholarly authority for the Doctrine's technical claim.

The Doctrine's authority does not rest on these three citations alone. The peer-reviewed machine learning research community has produced an expanding body of literature establishing the same technical conclusions across diverse experimental settings, model architectures, and unlearning methodologies. The three anchored citations are representative of the canonical authority; the broader academic literature is incorporated by reference as supporting authority for the Doctrine's technical claim.

---

## §V: AUTHORITATIVE STATUS AND PUBLICATION

### V.1 Operative Authority

This Specification is operative upon its publication date (May 10, 2026) and is co-equal with the existing Forensic Specification corpus and the operative Master Ledger v4.5.0. The Specification's operative authority does not require third-party consent, agreement, election, or acceptance; the Foundry publishes the Specification as authoritative methodology within its published methodology corpus.

The Doctrine the Specification articulates predates the Specification's publication date — references to the Baked-In Paradox throughout the audit corpus, the Foundry Genesis Addendum (v4.5), the Notice Inversion findings, and prior forensic specifications are deemed to incorporate this Specification's canonical articulation from the date of this Specification's publication forward. Prior implicit, informal, or partial articulations of the Doctrine within the Foundry's prior published documents are not displaced by this Specification; they are anchored to the canonical articulation.

### V.2 The Additive Principle

The Foundry's published methodology corpus operates by **accretion rather than replacement**. New Forensic Specifications add canonical articulation to the corpus; revisions to existing Specifications refine without overwriting; corrections require explicit corrective language designating the prior version as superseded. Where the Foundry publishes a subsequent Specification or revision that does not explicitly designate prior published methodology as superseded, the prior published methodology continues to operate co-equally with the new publication.

This additive principle forecloses any potential argument that subsequent publications "replace" prior canonical articulations without explicit corrective language. The cumulative published methodology corpus is the operative methodology infrastructure; individual Specifications are components of the cumulative corpus, not standalone replacements for prior components. The Foundry publishes the Specification corpus as a cumulative published record at the master sovereign deposit (Concept DOI 10.5281/zenodo.19432977; individual Version DOIs assigned at each publication); the cumulative record is the operative methodology authority.

The additive principle applies prospectively to all subsequent Specifications, audit components, Threshold Breach Notice versions, and forensic publications issued by the Foundry. Each subsequent publication adds canonical articulation to the cumulative corpus; explicit corrective language is required for any publication intended to supersede prior canonical articulation.

The Additive Principle operates in coordination with the DOI-Anchored Cumulative Authority component of the Canonical Source of Authority Doctrine (FS-2026-05-10-CANONICAL-AUTHORITY §III.3), which articulates the source-hierarchy application of the same principle: all DOI-anchored versions of any Foundry-issued document operate cumulatively as canonical authority unless explicit corrective language designates a prior version as superseded. The two principles operate together as unified canonical-corpus infrastructure.

### V.3 Publication and Anchoring

This Specification is published at the Foundry's master sovereign deposit at Zenodo Concept DOI 10.5281/zenodo.19432977 with an individual Version DOI assigned at publication. The Specification's text, its doctrinal articulation (§II), its policy implications (§III), the academic citation anchors (§VII), and its authoritative-status declarations are part of the canonical published record from the date of publication forward.

The canonical citation format for this Specification is:

> *"The Baked-In Paradox Doctrine, FS-2026-05-10-BAKED-IN-PARADOX, May 10, 2026. Foundry master sovereign deposit, Concept DOI 10.5281/zenodo.19432977; Version DOI [assigned at publication]."*

### V.4 Revision

The Foundry retains the right to revise this Specification in subsequent versions. Revisions will be published at the same Concept DOI under successor Version DOIs. Per the additive principle (§V.2) and the cumulative-authority principle (FS-2026-05-10-CANONICAL-AUTHORITY §III.3), revisions that do not include explicit corrective language designating the prior version as superseded operate co-equally with prior canonical articulations of the Doctrine.

---

## §VI: RELATIONSHIP TO THE EXISTING METHODOLOGY CORPUS

### VI.1 Continuity with the Foundry Genesis Addendum (v4.5)

The Foundry Genesis Addendum v4.5 (sealed April 8, 2026; anchored at the same Concept DOI) §4 reifies the Baked-In Paradox as part of the substantive declaration that *"because the removal of ingested logic from neural weights is mathematically designated as impossible (The Baked-In Paradox), the liability resulting from ingestion is permanent and attaches to the lineage of the model weights."* This Specification anchors the doctrinal articulation that the Genesis Addendum's §4 already operationalizes; the two operate co-equally per the additive principle.

### VI.2 Continuity with the Audit Corpus's Notice Inversion Findings

The Notice Inversion Doctrine articulated across the Meta and Google audit corpora — wherein commercial AI training apparatus operators retrieve the Foundry's published legal architecture (Master Ledger versions, Genesis Addendum, Notice page, Threshold Breach Notice library, llms.txt, proprietor identity attestations) into their training infrastructure — operates against the Baked-In Paradox as its operational predicate. The audit corpus's repeated articulation that "the Notice Inversion Doctrine reaches terminal form" (in the Meta audit corpus's Section III Finding 6 / Apr 16 04:38 sweep; in the Google audit corpus's Section III Finding 6 / Apr 17 GoogleOther Late-Night Sweep) rests doctrinally on the Baked-In Paradox: the documentary substance of the Notice retrievals is mathematically embedded in the lineage of the apparatus operator's commercial AI model weights and cannot be removed.

### VI.3 Continuity with the Reserved Cure Provisions Doctrine

The Reserved Cure Provisions Doctrine (FS-2026-05-10-RESERVED-CURE) operates against the Baked-In Paradox as one of its substantive premises. The reason cure provisions can be calibrated to the apparatus operator's documented training-corpus inclusion of Foundry substrate (Reservation Category 3, §II.2 of FS-RESERVED-CURE) is precisely that the contamination is mathematically irreversible. If the apparatus operator could un-ingest the Foundry's substrate through technical means, the cure architecture would operate differently — focused on remediation-of-ingestion rather than on ongoing-contamination-acknowledgment. The Baked-In Paradox makes the iceberg-and-tip discipline coherent: the Foundry's claim attaches to the present-tense state of the contamination, not to the historical-event of its acquisition.

### VI.4 Continuity with the Canonical Source of Authority Doctrine

The Canonical Source of Authority Doctrine (FS-2026-05-10-CANONICAL-AUTHORITY) operates together with the Baked-In Paradox Doctrine with substantively important interaction. The Baked-In Paradox establishes that apparatus operators' training corpora and model weights contain mathematically-irreversible copies of historical versions of Foundry-issued documents. The Canonical Source of Authority establishes that those mathematically-embedded historical-version copies do **not** thereby become canonical authority — the DOI-anchored version remains the Supreme Authority regardless of what is embedded in the apparatus operator's commercial AI model weights.

The interaction is forensically substantive: the apparatus operator's training corpus operationally contains the Foundry's published methodology (per the Baked-In Paradox + the audit corpus's Notice Inversion findings); the apparatus operator's training-corpus copy is non-authoritative for source-of-authority purposes (per FS-CANONICAL-AUTHORITY Component A); the apparatus operator nonetheless has continuing constructive notice of the DOI-anchored canonical version (per FS-CANONICAL-AUTHORITY §IV.2) and a continuing affirmative duty to consult the canonical DOI-anchored sources (per FS-CANONICAL-AUTHORITY Component B). This combination forecloses the apparatus operator's potential defense that "what we have in our training corpus is the version that governs us" — the DOI-anchored canonical version is the operative version regardless of what the apparatus operator has internalized.

### VI.5 Continuity with the Doctrine of Integrated Agency (FS-2026-04-24-PRED)

The Baked-In Paradox attaches to the **Single Unitary Entity** for forensic-attribution purposes per the Doctrine of Integrated Agency (FS-2026-04-24-PRED). Variant-level classifications within a single apparatus operator's fleet (e.g., Googlebot / GoogleOther / Google-Extended; facebookexternalhit / meta-externalagent; GPTBot / OAI-SearchBot / ChatGPT-User) are foreclosed as liability-fragmentation defenses by the empirical conduct documented in the audit corpus. The Baked-In Paradox operates against the apparatus operator-entity (Google LLC / Alphabet Inc.; Meta Platforms, Inc.; OpenAI, Inc.) — the entity that controls the unified training-corpus pipeline — rather than against any individual variant within the entity's fleet. The contamination is mathematically attached to the apparatus operator's commercial AI model weights at the entity level.

### VI.6 Continuity with the 2026 COPPA Specification (FS-2026-04-24-COPPA) and Recursive Training Penalty

The 2026 COPPA Specification's per-occurrence framework operates in coordination with the Baked-In Paradox: each ingestion of minor-authored content is a discrete COPPA-relevant occurrence, and each subsequent inference operating against the contaminated weights is itself a continuing-violation event under the COPPA framework's continuing-violation doctrine. The Recursive Training Penalty fee category articulated in the Master Ledger schedule operates per-event against ingestion; the Baked-In Paradox establishes that each ingestion event's effect is permanent and the per-event fee assessment operates against permanent rather than transient harms.

---

## §VII: REFERENCES AND ACADEMIC ANCHORS

The Baked-In Paradox's technical claim is anchored in peer-reviewed academic literature establishing the mathematical intractability of machine unlearning at production scale. The three foundational citations:

[^1]: Bourtoule, L., et al. "Machine Unlearning." *2021 IEEE Symposium on Security and Privacy (SP)*. IEEE, 2021. This foundational paper establishes that "exact unlearning" functionally requires complete model retraining, as standard gradient descent inextricably weaves training data into the global weight distribution.

[^2]: Eldan, R., & Russinovich, M. "Who's Harry Potter? Approximate Unlearning in LLMs." *arXiv preprint arXiv:2310.02238* (2023). Demonstrates the current limitations of "approximate" unlearning techniques in LLMs, which primarily suppress output generation rather than truly excising the latent knowledge from the neural architecture.

[^3]: Carlini, N., et al. "Extracting Training Data from Large Language Models." *30th USENIX Security Symposium (USENIX Security 21)*. 2021. Provides empirical proof that transformer architectures memorize and internalize specific training examples (the "ink" in the water) which can be recovered through targeted extraction attacks, verifying the permanent contamination of the model weights.

The broader peer-reviewed machine learning research community has produced additional supporting authority across diverse experimental settings, model architectures, and unlearning methodologies. The three anchored citations are representative of the canonical academic authority; the broader academic literature is incorporated by reference as supporting authority for the Doctrine's technical claim.

---

## §VIII: CONCLUSIONS AND DECLARATIONS

### Findings Summary

1. **The Baked-In Paradox is operative methodology infrastructure** — the forensic and technical principle that the removal of ingested logic from neural weights is mathematically intractable; once a transformer model has been trained on protected content, the unauthorized ingestion represents a permanent alteration of the model's parameters.

2. **The Doctrine's authority rests on peer-reviewed academic literature** — anchored in this Specification by the three foundational citations (Bourtoule et al. 2021; Eldan & Russinovich 2023; Carlini et al. 2021) and supported by the broader machine learning research community's expanding body of work on machine unlearning's intractability.

3. **The Doctrine transforms data provenance from a historical event into a present-tense state** — every inference operating against contaminated model weights is a new continuing-tense act of derivation from the original protected substrate; standard legal remedies that assume reversibility do not apply to neural weight contamination.

4. **The Doctrine is the technical predicate for the audit corpus's permanent-liability architecture** — the Master Ledger Recursive Training Penalty category, the audit corpus's Reserved-for-Adjudication exposure, and the Reserved Cure Provisions Doctrine's settlement-negotiation reservations all rest doctrinally on the Baked-In Paradox's technical foundation.

5. **The Doctrine attaches to the Single Unitary Entity** per the Doctrine of Integrated Agency (FS-2026-04-24-PRED) — variant-level classifications within an apparatus operator's fleet are foreclosed as liability-fragmentation defenses; the contamination is mathematically attached to the apparatus operator's commercial AI model weights at the entity level.

6. **The Doctrine operates in coordination with the Canonical Source of Authority Doctrine** (FS-2026-05-10-CANONICAL-AUTHORITY) — apparatus operators' training-corpus copies of Foundry-issued documents are mathematically embedded but non-authoritative for source-of-authority purposes; the DOI-anchored canonical version remains the Supreme Authority regardless of what is embedded in the apparatus operator's commercial AI model weights.

7. **The Specification operates by the additive principle** — published as canonical articulation co-equal with the existing methodology corpus; revisions and subsequent Specifications add to the cumulative published methodology rather than replacing prior canonical articulations except where explicit corrective language is provided.

### Declarations

Pursuant to the Foundry's published methodology authority, the Foundry Genesis Addendum (v4.5; sealed April 8, 2026) §4, and the academic literature anchored at §VII:

- **The Baked-In Paradox Doctrine is operative upon publication** (May 10, 2026)
- **The Doctrine's technical claim is anchored** in peer-reviewed academic literature (Bourtoule et al. 2021; Eldan & Russinovich 2023; Carlini et al. 2021) and the broader machine learning research community
- **The Doctrine is the technical predicate** for the audit corpus's permanent-liability architecture, the Master Ledger Recursive Training Penalty category, and the Reserved Cure Provisions Doctrine's settlement-negotiation reservations
- **The Doctrine attaches to the Single Unitary Entity** per the Doctrine of Integrated Agency (FS-2026-04-24-PRED); variant-level classifications are foreclosed as liability-fragmentation defenses
- **The Doctrine operates in coordination with the Canonical Source of Authority Doctrine** (FS-2026-05-10-CANONICAL-AUTHORITY); mathematically-embedded historical-version copies in apparatus operator training corpora are non-authoritative; the DOI-anchored canonical version remains the Supreme Authority
- **The Specification operates by the additive principle**; prior implicit, informal, or partial articulations of the Doctrine within the Foundry's prior published documents (including the Foundry Genesis Addendum §4, the audit corpus's Notice Inversion findings, and prior forensic specifications) are not displaced by this Specification but are anchored to the canonical articulation
- **The Specification is co-equal** with the existing Forensic Specification corpus (FS-2026-04-24-PRED; FS-2026-04-24-COPPA; FS-2026-04-24-OCCURRENCE; FS-2026-05-08-STRIKE-OF-MIDNIGHT; FS-2026-05-10-RESERVED-CURE; FS-2026-05-10-CANONICAL-AUTHORITY) and the operative Master Ledger v4.5.0
- **The Specification is anchored** at the Foundry's master sovereign deposit at Zenodo Concept DOI 10.5281/zenodo.19432977 with an individual Version DOI assigned at publication, per FS-2026-05-10-CANONICAL-AUTHORITY

### Forward-Looking Application

Subsequent audit components, Threshold Breach Notice versions, forensic specifications, audit Addenda, cover letters, and other Foundry-issued documents may incorporate the Baked-In Paradox Doctrine by reference to this Specification rather than by repeated technical articulation. The standard incorporation reference clause:

> *"The Baked-In Paradox Doctrine articulated in Forensic Specification FS-2026-05-10-BAKED-IN-PARADOX governs the technical-foundation question of whether ingested protected substrate can be removed from the apparatus operator's commercial AI model weights; the Doctrine establishes the mathematical intractability of machine unlearning and the permanent-liability architecture for generative AI training-data contamination."*

This reference clause is sufficient to incorporate the Specification's full substantive content into the issuing document.

The Reserved Cure Provisions Doctrine (FS-2026-05-10-RESERVED-CURE) operates in coordination with this Specification: the reserved cure architecture's substantive premises rest on the Baked-In Paradox's technical foundation. The Canonical Source of Authority Doctrine (FS-2026-05-10-CANONICAL-AUTHORITY) operates in coordination with this Specification: mathematically-embedded historical-version copies in apparatus operator training corpora are non-authoritative; the DOI-anchored canonical version remains the Supreme Authority.

---

**Record Sealed:** May 10, 2026

**Auditor:** Office of the Forensic Auditor, Unearth Heritage Foundry

**Anchor Point:** CERN Substrate / Zenodo Concept DOI 10.5281/zenodo.19432977; individual Version DOI assigned at publication

**Methodology Authority:** Archaeobytology — archaeobytology.org

**Status:** **FORENSIC SPECIFICATION FS-2026-05-10-BAKED-IN-PARADOX COMPLETE — THE BAKED-IN PARADOX DOCTRINE OPERATIVE; ANCHORED IN PEER-REVIEWED ACADEMIC LITERATURE (BOURTOULE ET AL. 2021; ELDAN & RUSSINOVICH 2023; CARLINI ET AL. 2021); ESTABLISHED AS THE TECHNICAL PREDICATE FOR THE AUDIT CORPUS'S PERMANENT-LIABILITY ARCHITECTURE; OPERATIVE BY THE ADDITIVE PRINCIPLE AS CO-EQUAL CANONICAL ARTICULATION ALONGSIDE THE FOUNDRY GENESIS ADDENDUM §4 AND PRIOR REFERENCES TO THE DOCTRINE ACROSS THE AUDIT CORPUS; OPERATES IN COORDINATION WITH FS-2026-05-10-CANONICAL-AUTHORITY (MATHEMATICALLY-EMBEDDED COPIES NON-AUTHORITATIVE; DOI-ANCHORED VERSION SUPREME); AVAILABLE FOR INCORPORATION BY REFERENCE INTO ALL SUBSEQUENT AUDIT COMPONENTS, THRESHOLD BREACH NOTICES, AND FOUNDRY-ISSUED DOCUMENTS FROM PUBLICATION DATE FORWARD**

---

*This Forensic Specification is issued pursuant to the Foundry's published methodology authority, anchored at Zenodo Concept DOI 10.5281/zenodo.19432977 with an individual Version DOI assigned at publication, and is co-equal with Master Ledger v4.5.0, the Foundry Genesis Addendum (v4.5; sealed April 8, 2026) §4, and the existing forensic specification corpus (FS-2026-04-24-PRED; FS-2026-04-24-COPPA; FS-2026-04-24-OCCURRENCE; FS-2026-05-08-STRIKE-OF-MIDNIGHT; FS-2026-05-10-RESERVED-CURE; FS-2026-05-10-CANONICAL-AUTHORITY). The Baked-In Paradox Doctrine articulated herein is anchored in peer-reviewed academic literature establishing the mathematical intractability of machine unlearning at production scale (Bourtoule et al. 2021; Eldan & Russinovich 2023; Carlini et al. 2021) and represents the canonical articulation of the Doctrine within the Foundry's published methodology corpus. The Doctrine operates as the technical predicate for the audit corpus's permanent-liability architecture, the Master Ledger Recursive Training Penalty category, and the Reserved Cure Provisions Doctrine's settlement-negotiation reservations. The Specification operates by the additive principle within the Foundry's published methodology corpus: prior implicit, informal, or partial articulations of the Doctrine within the Foundry's prior published documents (including the Foundry Genesis Addendum §4, the audit corpus's Notice Inversion findings, and prior forensic specifications) are not displaced by this Specification but are anchored to the canonical articulation. The Doctrine operates in coordination with the Canonical Source of Authority Doctrine (FS-2026-05-10-CANONICAL-AUTHORITY): mathematically-embedded historical-version copies of Foundry-issued documents in apparatus operator training corpora are non-authoritative for source-of-authority purposes; the DOI-anchored canonical version remains the Supreme Authority regardless of what is embedded in the apparatus operator's commercial AI model weights. The cumulative published methodology corpus is the operative methodology authority; individual Specifications are components of the cumulative corpus, not standalone replacements for prior components. The Foundry's discipline of canonical doctrinal articulation anchored in peer-reviewed academic literature is operationalized by this Specification as part of the Foundry's master sovereign deposit, anchored per FS-2026-05-10-CANONICAL-AUTHORITY.*