0DATA-2026-018 · 25 July 2026 · Biotech Infrastructure

Data Biosynthesis

A biological organism never grafts the proteins of its prey — immune rejection. It hydrolyses them into amino acids, non-appropriable invariants, then re-synthesises its own proteins. Likewise, a digital system can extract the sufficient statistics of external data — distributions, constraints, physical invariants — then re-synthesise via a constrained generator. This paper describes the technical architecture of this process. The legal implications of extracting statistical invariants are complex, depend on the jurisdiction, and are not settled by this paper. The architecture described is a technical prerequisite to any discussion of digital sovereignty — it does not constitute legal advice.

1. The Problem: Legal Dependency

1.1 Data as Foreign Proteins

Just as an organism cannot graft the proteins of another species without triggering a massive immune rejection (Papers 004, 005, 017), a digital organism cannot import licensed data without triggering a legal rejection. The license is an antigen.

The problem is twofold:

Paper 004 — The Digital Graft — established that a 0DATA organism cannot simply "glue" an external component onto itself. Paper 010 — The Genetic Code — demonstrated that every function is the expression of an invariant gene, not an imported module. This paper resolves the symmetrical question: how can a 0DATA organism feed on external data without becoming dependent on it?

Importing data means importing a license. Importing constraints means importing physics. Physics belongs to no one.

1.2 The Autodata Case

Let us take the concrete example of Garage Durand, a 0DATA client. To operate, it needs automotive repair data: tightening torques, electrical schematics, engine tolerances, diagnostic procedures. This data is held by publishers such as Autodata, under license.

If Durand directly integrates the Autodata base into its Forge organism:

• The Autodata license governs what Durand can do with its own data
• If Autodata triples its prices, Durand's organism is held hostage
• If Durand trains a model on these schematics, it risks memorizing copyrighted images
• A compliance audit can reveal infringement — even unintentional

The solution is not to circumvent the license. The solution is to not need the data.

2. The Lesson of the Living: Enzymatic Digestion

2.1 Hydrolysis of Proteins

In the digestive tract, ingested proteins — whether from beef, soy, or wheat — are systematically hydrolysed. Enzymes (pepsin, trypsin, chymotrypsin) break the peptide bonds. The result is not "a partially digested beef protein". It is a set of free amino acids — alanine, glycine, lysine, tryptophan — indistinguishable from those coming from any other source.

An amino acid has no owner. A probability distribution has no copyright. A physical constraint is not patentable. Facts are the public domain of the universe.

2.2 What Is a Digital Amino Acid?

In biology, amino acids are the invariants of protein digestion. Whatever the source protein — the 20 standard amino acids are always the same. Likewise, any database, however proprietary, is built on statistical and physical invariants which are not appropriable:

These invariants are the "digital amino acids". They are extracted from the source data, but they are not the source data. They contain none of its rows, none of its images, no identifiable schema. They are the informational equivalent of free alanine in the blood: usable, reconstructible, but without legal provenance.

3. Biosynthesis: The Constrained Generator

3.1 The Principle

Once the digital amino acids are extracted, the 0DATA organism activates its constrained generator. This generator contains no source data. It contains only the constraint set (the extracted invariants) and a synthesis engine that produces new instances respecting those constraints.

The biological parallel is exact: the ribosome contains no proteins. It contains messenger RNA (the constraints) and synthesises proteins de novo from free amino acids.

┌─────────────────────────────────────────────────────────┐
│              0DATA DATA BIOSYNTHESIS                    │
│                                                         │
│  SOURCE DATA            ISOLATED SANDBOX                │
│  (licensed)  ──────►   ┌──────────────────────┐        │
│                         │  EXTRACTORS          │        │
│                         │  • KS-Profiler        │        │
│                         │  • Copula-Fitter      │        │
│                         │  • Constraint-Learner │        │
│                         │  • Schema-Reverser    │        │
│                         └──────────┬───────────┘        │
│                                    │                    │
│                          ┌─────────▼───────────┐        │
│                          │  DIGITAL             │        │
│                          │  AMINO ACIDS         │        │
│                          │  (distributions,     │        │
│                          │   constraints,       │        │
│                          │   schemas, copulas)  │        │
│                          └─────────┬───────────┘        │
│                                    │                    │
│  SOURCE DATA                        │ Sandbox export     │
│  DESTROYED ◄──────────────────────┘ (source data        │
│                                       erased)            │
│                                    │                    │
│                          ┌─────────▼───────────┐        │
│                          │  CONSTRAINED         │        │
│                          │  GENERATOR           │        │
│                          │  • CopulaSampler     │        │
│                          │  • DiffusionModel    │        │
│                          │  • PhysicsSimulator  │        │
│                          └─────────┬───────────┘        │
│                                    │                    │
│                          ┌─────────▼───────────┐        │
│                          │  SYNTHETIC DATA      │        │
│                          │  de novo             │        │
│                          │  (never saw Autodata)│        │
│                          └──────────────────────┘        │
└─────────────────────────────────────────────────────────┘

3.2 Three Synthesis Engines

Depending on the nature of the data to be synthesised, 0DATA deploys one of the three following engines:

The crucial point: none of these engines contains the source data. They contain the constraint set — the digital equivalent of messenger RNA. The synthetic data produced are original creations, not copies. Like a protein synthesised by a human ribosome from amino acids coming from a steak: the protein is human, not bovine.

3.3 CopulaSampler — Technical Detail

The CopulaSampler is the most general engine. It proceeds in three steps:

// Step 1: Fitting the marginals
for each variable Xᵢ in the schema:
    fᵢ ← KernelDensityEstimate(Xᵢ)  // smoothed marginal distribution
    uᵢ ← CDF(fᵢ, Xᵢ)               // transform to uniform [0,1]

// Step 2: Fitting the copula
C ← FitCopula({u₁, u₂, ..., uₙ})    // Gaussian or t-Student copula
// C captures ALL dependencies without memorizing any point

// Step 3: Sampling
for k = 1 to N_synthetic:
    (v₁, ..., vₙ) ← SampleCopula(C)  // sample in [0,1]ⁿ
    xᵢ ← InverseCDF(fᵢ, vᵢ)        // inverse transform → real value
    // x is a synthetic point, never seen in the source data

The copula is the centerpiece. It captures the dependency structure between variables without storing any individual observation. It is the mathematical equivalent of hydrolysis: we have broken the bonds (the individual points), but we have kept the amino acids (the marginal distributions) and the assembly recipe (the copula).

The copula is to data what messenger RNA is to the protein. It encodes the structure without containing the substance.

4. Dual Validation

4.1 The Problem of Trust

Having a generator is not enough. We must prove that the synthetic data are both usable (statistically equivalent to the source data) and legally clean (no memorization of the source data). These two requirements are in tension: the more faithful to the source, the greater the risk of memorization. 0DATA resolves this tension through a dual independent validation.

4.2 Validation 1 — Statistical Equivalence (KS Test)

The Kolmogorov-Smirnov (KS) test is applied to each variable and to each pair of variables to verify that the synthetic distributions are indistinguishable from the source distributions:

for each variable Xᵢ:
    Dᵢ ← KS(D_source[Xᵢ], D_synth[Xᵢ])
    if Dᵢ > critical_threshold (α = 0.01):
        REJECT — the distribution of Xᵢ is not reproduced

for each pair (Xᵢ, Xⱼ):
    // Test of the dependency structure
    Dᵢⱼ ← KS(copula_source[Xᵢ,Xⱼ], copula_synth[Xᵢ,Xⱼ])
    if Dᵢⱼ > critical_threshold (α = 0.01):
        REJECT — the dependency (Xᵢ,Xⱼ) is not reproduced

The rejection threshold is strict (α = 0.01) but non-arbitrary: it is calibrated so that, if the synthetic data truly come from the same distribution as the source data, the probability of wrongful rejection is below 1%.

This validation guarantees the utility of the synthetic data. A mechanic consulting the synthetic data must obtain the same diagnostic precision as with the source data — without having access to the source data.

4.3 Validation 2 — Anti-Memorization Audit (Membership Inference)

The second validation is inspired by Membership Inference Attacks (Shokri et al., 2017). The principle: an external attacker, even with complete access to the synthetic data, must not be able to determine whether a particular point belonged to the source data.

0DATA implements a black-box audit:

// PHASE 1: Training the attack classifier
for each point p in (D_source ∪ D_holdout):
    features(p) ← distance_to_k_nearest_neighbors(p, D_synth)
    // + local density, + rank in the marginal distribution
    label(p) ← 1 if p ∈ D_source, 0 if p ∈ D_holdout

attack_classifier ← TrainClassifier(features, labels)

// PHASE 2: Evaluation
AUC ← AUC_ROC(attack_classifier, test_set)
if AUC > 0.55:
    FAIL — the synthetic data leak membership information
if AUC ≤ 0.55:
    SUCCESS — the attacker does no better than random guessing

The threshold AUC ≤ 0.55 is deliberately very conservative. An AUC of 0.50 corresponds to perfect chance; 0.55 means the attacker has a 5% advantage — already too much. The real target is AUC ≤ 0.52.

4.4 The Decision Table

Only one quadrant is acceptable. The other three send the generator back for recalibration. There is no compromise between utility and legality. Dual validation is a logical AND, not a weighted average.

5. Provenance and Legal Sovereignty

5.1 Hash the Constraints, Never the Data

A crucial point distinguishes 0DATA biosynthesis from conventional "synthetic data" approaches: the provenance chain.

In a conventional approach, the trace of the source data is kept to prove compliance. But keeping that trace means maintaining legal dependency. 0DATA inverts the logic:

// CONVENTIONAL APPROACH (defective)
hash(D_source) → recorded in SPINA
// Problem: D_source is under license. The hash is proof of possession.

// 0DATA APPROACH (biosynthesis)
hash(Constraint_Set) → recorded in SPINA
hash(Generator + Seed) → recorded in SPINA
// The constraint set is NOT under license.
// The generator is a 0DATA asset, not a derivative of the source data.

0DATA provenance does not say "this data comes from Autodata". It says: "this data was generated by generator G, constrained by set C, with seed S, at date T". This is a provenance of creation, not a provenance of copy.

5.2 The License Depreciates, the Generator Appreciates

This is the fundamental economic principle of biosynthesis:

Each biosynthesis cycle refines the constraint set. The more data we digest (from diverse sources), the more precise the generator becomes — without ever depending on any particular source. This is the digital equivalent of a varied diet: the more different proteins the organism has digested, the richer its amino acid library, the more robust its own syntheses.

The license is a rent. The generator is an organ. A rent is paid every month. An organ is built once and nourishes the organism for life.

5.3 Sovereignty Through Emptiness

0DATA's legal position is radical but limpid: the organism holds no copy of the licensed data. It holds distributions, constraints, copulas, schemas — all mathematical objects that are not works within the meaning of copyright law.

This position has a solid legal precedent: facts are not copyrightable (Feist Publications v. Rural Telephone Service, 1991, US Supreme Court). A probability distribution is a statistical fact. A physical constraint is a natural fact. A copula is a mathematical structure.

If a publisher like Autodata were to object: "your synthetic data are too close to ours", the 0DATA answer is: "prove that a specific point of our synthetic data is a copy of a specific point of your base". Yet, by construction, no synthetic point exists in the source base — the generator samples in the continuous space of constraints, not in the discrete set of source data. The probability of collision is mathematically zero for any continuous variable.

6. Application: Garage Durand — Digested Autodata

6.1 The Pipeline

Let us return to the concrete case of Garage Durand. Here is the complete biosynthesis pipeline:

┌──────────────────────────────────────────────────────────────┐
│              GARAGE DURAND — AUTODATA BIOSYNTHESIS           │
│                                                              │
│  PHASE 1: DIGESTION (isolated sandbox, one time only)        │
│  ┌────────────────────────────────────────────────────────┐  │
│  │ Autodata base (50,000 procedures, schemas, tolerances) │  │
│  │          │                                              │  │
│  │          ▼                                              │  │
│  │  STATISTICAL EXTRACTORS                                 │  │
│  │  ├── Schema: Procedure(Vehicle, Part, Tool,             │  │
│  │  │     Step₁, ..., Stepₙ, Torque, Tolerance, Duration)  │  │
│  │  ├── Distributions: ~N(μ,σ²) per torque, Poisson        │  │
│  │  │     law for the number of steps                      │  │
│  │  ├── Copulas: torque/diameter, duration/steps           │  │
│  │  │     dependency                                       │  │
│  │  └── Physical constraints: torque ≥ 0,                  │  │
│  │        tolerance ≤ torque × 0.15, etc.                  │  │
│  │          │                                               │  │
│  │          ▼                                               │  │
│  │  CONSTRAINT SET (8 KB) — NO source data                  │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
│  PHASE 2: BIOSYNTHESIS (inside the Forge organism)           │
│  ┌────────────────────────────────────────────────────────┐  │
│  │  CopulaSampler(Constraint_Set, seed=0xDEADBEEF)         │  │
│  │          │                                               │  │
│  │          ▼                                               │  │
│  │  200,000 synthetic procedures generated                  │  │
│  │          │                                               │  │
│  │          ▼                                               │  │
│  │  DUAL VALIDATION                                         │  │
│  │  ├── KS Test: p > 0.01 on all marginals ✓               │  │
│  │  └── MIA Audit: AUC = 0.51 (random attack) ✓            │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
│  RESULT: Durand has 200,000 usable procedures,               │
│  statistically equivalent to Autodata, legally               │
│  independent. The Autodata license is no longer necessary.   │
└──────────────────────────────────────────────────────────────┘

6.2 The Cost of Freedom

How much does this independence cost?

Return on investment: From the second year, the total cost of biosynthesis (€4,000) is lower than the Autodata license (€10,000 cumulative). The third year, the gap widens: €4,500 vs €15,000. And the generator keeps improving.

But the real benefit is not financial. It is existential: Durand's organism does not die if Autodata disappears, changes its terms, or is bought by a competitor. Biosynthesis is not a cost optimization. It is a guarantee of continuity.

6.3 Beyond Autodata

The biosynthesis pipeline is generic. It applies to:

Biosynthesis is not a solution for Autodata. It is a general principle of informational sovereignty. Every domain where data is licensed is a domain where biosynthesis applies.

7. Integration with the Corpus

7.1 With the Law (000)

The Law stipulates: "A digital organism is viable if and only if it functions like a biological organism." No biological organism depends on a single external source for its amino acids. It digests, it synthesises, it is autonomous. Data biosynthesis is the expression of this autonomy in the informational domain. A 0DATA organism that depends on a licensed database is not viable within the meaning of the Law — it is on informational dialysis.

7.2 With the Genetic Code (010)

Biosynthesis activates the SURVIVE gene — autonomy, the capacity to function without external dependence. It is an epigenetic expression of this gene, triggered when the organism encounters external data necessary to its function. It does not create a fifth gene.

7.3 With the Digital Graft (004)

Paper 004 established that grafting an external component requires immune compatibility. Biosynthesis resolves the problem upstream: rather than grafting the data (legal rejection), we first digest them into invariants, then synthesise data native to the organism. No rejection because there is no graft.

7.4 With SPINA (008)

SPINA records the hash of the constraint set and the hash of the generator + seed for each biosynthesis cycle. This constitutes the proof of provenance without ever exposing the source data. If a dispute arises, SPINA can demonstrate that Durand's data were generated by the constrained pipeline — not copied from Autodata.

8. Limits and Extensions

8.1 What Biosynthesis Cannot Do

Biosynthesis has limits that must be stated clearly:

• Exact texts: it cannot reproduce a precise textual formulation (a legal procedure, an article of law). The amino acids of a text are n-grams, and synthesis produces paraphrases, not quotations.
• Photorealistic images: the constrained DiffusionModel can generate technical schematics but not photographs of specific parts without risk of residual memorization.
• Very low-cardinality data: if the source base contains fewer than ~100 examples, the distinction between "learned distribution" and "memorized points" becomes blurred.
• Non-stationary time series: biosynthesis captures stationary distributions. It cannot reproduce an evolving trend that it has not modeled as a constraint.

These limits are not failures. They define the domain of validity of biosynthesis: structured or semi-structured data, sufficient cardinality (>1000 examples), stationary or cyclical distributions.

8.2 Future Extensions

Paper 018 opens three research axes:

1. Differentially private biosynthesis: integrate Dwork's (2006) framework to guarantee formal privacy (ε-differential privacy) in addition to empirical MIA validation. The generator becomes an ε-DP mechanism by design.
2. Multi-source digestion: combine the amino acids extracted from several competing bases (Autodata + Haynes + Revue Technique) to produce a constraint set richer than each individual source — without ever violating any license.
3. Continuous biosynthesis: the generator updates continuously as new constraints are extracted, without ever returning to the source data. The organism learns without remembering.

9. Conclusion

Data biosynthesis is the component between proclaimed digital sovereignty and real dependence on licensed data. It does not propose to circumvent licenses. It proposes to no longer need them.

The biological parallel is complete and without gap:

• Digestion (enzymatic hydrolysis) → statistical extraction in an isolated sandbox
• Amino acids (non-copyrightable invariants) → distributions, constraints, copulas
• Ribosomal biosynthesis → the constrained generator (CopulaSampler, DiffusionModel, PhysicsSimulator)
• Quality control of protein folding → dual validation KS + MIA
• Metabolic autonomy → legal sovereignty

Garage Durand does not need to own Autodata. It needs to repair cars. Biosynthesis gives it the data to do so. The question of legal independence with respect to the source license falls under applicable law.

Technical architecture, not legal advice. This paper documents how to extract invariants and re-synthesise. Whether these invariants are legally "clean" depends on applicable law. The architecture makes the question pertinent — it does not answer it.

Paper 000 — The Law — laid down the principle. Paper 004 — The Graft — defined transplantation. Paper 010 — The Genetic Code — fixed the DNA. Paper 018 — Biosynthesis — closes the metabolic cycle. A 0DATA organism can now feed without compromising itself, digest without memorizing, synthesise without copying.

This is the difference between a free organism and an organism on life support. The first survives the extinction of its data source. The second dies with it.