Model Governance

AI Model Lineage

Track every model's ancestry, training method, data sources, and compliance status. Maintain a complete audit trail from base model to production deployment.

Training Methods

Supervised

Labeled training data with explicit input-output pairs.

Fine-Tuning

Adapting a pre-trained model on domain-specific data.

RLHF

Reinforcement learning from human feedback.

Distillation

Knowledge transfer from a larger teacher model.

Compliance Requirements by Method

Each training method carries different compliance implications for data licensing, consent requirements, and audit depth.

Supervised Learning

Requires documentation of labeling methodology, annotator agreements, and label quality metrics.

Compliance requirements: Dataset licensing, annotator consent, bias auditing

Fine-Tuning

Requires lineage of the base model, fine-tuning dataset provenance, and hyperparameter documentation.

Compliance requirements: Base model license, derivative rights, data consent

RLHF

Tracks reward model lineage, human preference datasets, and feedback loop iterations with annotator demographics.

Compliance requirements: Annotator agreements, preference data consent, bias mitigation

Distillation

Requires documentation of teacher model identity, distillation methodology, and capability retention metrics.

Compliance requirements: Teacher model license, capability disclosure, performance benchmarks

EU AI Act Article 53

High-risk AI systems must document the training methodologies used, including data selection, labeling, and cleaning techniques. DRD Model Lineage provides this documentation automatically.

Dataset Provenance

Every dataset linked to a model lineage record includes full provenance metadata.

nameHuman-readable dataset identifierYes

sourceOrigin URL or internal registry pathYes

licenseSPDX license identifier (e.g., CC-BY-4.0)Yes

versionSemantic version or hash of the dataset snapshotYes

recordCountNumber of records/samples in the datasetYes

collectedAtISO 8601 timestamp of data collectionYes

consentTypenone | opt-in | opt-out | contractualYes

piiCategoriesArray of PII types presentIf applicable

biasAuditReference to bias assessment reportRecommended

retentionPolicyData retention period and deletion rulesRecommended

Dependency Graphs

DRD automatically builds a directed acyclic graph (DAG) of model dependencies, showing the relationships between foundation models, fine-tuned variants, distilled models, and their datasets.

dependency-graph.json

{
  "model": "content-guardian-v3",
  "method": "fine-tuning",
  "parent": {
    "model": "llama-3-70b",
    "method": "supervised",
    "datasets": [
      { "name": "CommonCrawl-2025", "license": "CC-BY-4.0" },
      { "name": "Wikipedia-EN-2025", "license": "CC-BY-SA-4.0" }
    ]
  },
  "datasets": [
    { "name": "drd-content-safety-v2", "license": "proprietary", "consentType": "contractual" },
    { "name": "drd-policy-violations", "license": "proprietary", "consentType": "opt-in" }
  ],
  "children": [
    {
      "model": "content-guardian-v3-lite",
      "method": "distillation"
    }
  ]
}

License Propagation

DRD validates license compatibility across the dependency graph. If a parent model uses a copyleft license, all derivatives are flagged for compliance review.

Registering Models

lineage.ts

import { DRD } from '@drd/sdk';

const drd = new DRD({ apiKey: process.env.DRD_API_KEY });

const record = await drd.modelLineage.createRecord({
  modelId: 'custom-classifier-v3',
  modelName: 'Custom Content Classifier',
  version: '3.0.0',
  parentModelId: 'base-model-v2-id',
  trainingMethod: 'fine_tuning',
  datasetSize: 50000,
  license: 'Apache-2.0',
  trainingDataSources: {
    datasets: ['licensed-dataset-a', 'internal-dataset-b'],
    consentVerified: true,
  },
});

Compliance Audits

Create audits to verify model compliance with licensing, data usage, and regulatory requirements.

audit.ts

await drd.modelLineage.createAudit({
  recordId: record.id,
  auditType: 'compliance',
  findings: {
    dataSourcesVerified: true,
    licenseCompatible: true,
    consentChainValid: true,
  },
  complianceStatus: 'compliant',
  nextAuditDue: '2026-06-01T00:00:00Z',
});

REST API

POST /v1/models/lineage

curl -X POST https://api.drd.io/v1/models/lineage \
  -H "Authorization: Bearer drd_ws_sk_live_Abc123..." \
  -H "Content-Type: application/json" \
  -d '{
    "modelId": "mdl_01JM7XBN4RTYP",
    "name": "content-guardian-v3",
    "version": "3.0.0",
    "method": "fine-tuning",
    "parentModelId": "mdl_00BASE70BLLAMA",
    "datasets": [
      {
        "name": "drd-content-safety-v2",
        "source": "s3://drd-datasets/content-safety-v2",
        "license": "proprietary",
        "version": "2.1.0",
        "recordCount": 2500000,
        "consentType": "contractual"
      }
    ],
    "hyperparameters": {
      "learningRate": 2e-5,
      "epochs": 3,
      "batchSize": 32,
      "warmupSteps": 500
    },
    "metrics": {
      "accuracy": 0.96,
      "f1": 0.94,
      "latencyP99Ms": 45
    }
  }'

// Response
{
  "ok": true,
  "data": {
    "id": "lin_01JM7XBN4RTYP",
    "modelId": "mdl_01JM7XBN4RTYP",
    "name": "content-guardian-v3",
    "method": "fine-tuning",
    "parentModelId": "mdl_00BASE70BLLAMA",
    "datasetsCount": 1,
    "createdAt": "2026-02-14T12:00:00Z",
    "signedHash": "sha256:a1b2c3d4..."
  }
}

GET /v1/models/lineage/:modelId/graph

curl "https://api.drd.io/v1/models/lineage/mdl_01JM7XBN4RTYP/graph?depth=3" \
  -H "Authorization: Bearer drd_ws_sk_live_Abc123..."

// Response
{
  "ok": true,
  "data": {
    "root": "mdl_01JM7XBN4RTYP",
    "nodes": [
      { "id": "mdl_01JM7XBN4RTYP", "name": "content-guardian-v3", "method": "fine-tuning" },
      { "id": "mdl_00BASE70BLLAMA", "name": "llama-3-70b", "method": "supervised" },
      { "id": "mdl_02LITE7XBN4RT", "name": "content-guardian-v3-lite", "method": "distillation" }
    ],
    "edges": [
      { "from": "mdl_00BASE70BLLAMA", "to": "mdl_01JM7XBN4RTYP", "relation": "parent" },
      { "from": "mdl_01JM7XBN4RTYP", "to": "mdl_02LITE7XBN4RT", "relation": "parent" }
    ]
  }
}

Full SDK Example

lineage-workflow.ts

import { DRD } from '@drd/sdk';

const drd = new DRD({ apiKey: process.env.DRD_API_KEY });

// Register a new model lineage record
const lineage = await drd.modelLineage.createRecord({
  modelId: 'mdl_01JM7XBN4RTYP',
  modelName: 'content-guardian-v3',
  version: '3.0.0',
  parentModelId: 'mdl_00BASE70BLLAMA',
  trainingMethod: 'fine_tuning',
  datasets: [{
    name: 'drd-content-safety-v2',
    source: 's3://drd-datasets/content-safety-v2',
    license: 'proprietary',
    recordCount: 2_500_000,
    consentType: 'contractual',
  }],
});

// Query the dependency graph
const graph = await drd.modelLineage.graph(lineage.modelId, { depth: 3 });
console.log(`Nodes: ${graph.nodes.length}, Edges: ${graph.edges.length}`);

// Check license compatibility
const licenseCheck = await drd.modelLineage.checkLicenses(lineage.modelId);
if (licenseCheck.conflicts.length > 0) {
  console.warn('License conflicts found:', licenseCheck.conflicts);
}

// Get full audit trail
const trail = await drd.modelLineage.auditTrail(lineage.modelId);
for (const entry of trail) {
  console.log(`${entry.timestamp}: ${entry.action} by ${entry.actor}`);
}

Next Steps

Provenance

Track model decisions

Learn more →

Royalty Splitting

Compensate data providers

Learn more →

Explainability

Explain model decisions

Learn more →

Training Verification

Verify training data provenance and consent

Learn more →

Model Governance

AI Model Lineage

Track every model's ancestry, training method, data sources, and compliance status. Maintain a complete audit trail from base model to production deployment.

Training Methods

Supervised

Labeled training data with explicit input-output pairs.

Fine-Tuning

Adapting a pre-trained model on domain-specific data.

RLHF

Reinforcement learning from human feedback.

Distillation

Knowledge transfer from a larger teacher model.

Compliance Requirements by Method

Each training method carries different compliance implications for data licensing, consent requirements, and audit depth.

Supervised Learning

Requires documentation of labeling methodology, annotator agreements, and label quality metrics.

Compliance requirements: Dataset licensing, annotator consent, bias auditing

Fine-Tuning

Requires lineage of the base model, fine-tuning dataset provenance, and hyperparameter documentation.

Compliance requirements: Base model license, derivative rights, data consent

RLHF

Tracks reward model lineage, human preference datasets, and feedback loop iterations with annotator demographics.

Compliance requirements: Annotator agreements, preference data consent, bias mitigation

Distillation

Requires documentation of teacher model identity, distillation methodology, and capability retention metrics.

Compliance requirements: Teacher model license, capability disclosure, performance benchmarks

EU AI Act Article 53

High-risk AI systems must document the training methodologies used, including data selection, labeling, and cleaning techniques. DRD Model Lineage provides this documentation automatically.

Dataset Provenance

Every dataset linked to a model lineage record includes full provenance metadata.

nameHuman-readable dataset identifierYes

sourceOrigin URL or internal registry pathYes

licenseSPDX license identifier (e.g., CC-BY-4.0)Yes

versionSemantic version or hash of the dataset snapshotYes

recordCountNumber of records/samples in the datasetYes

collectedAtISO 8601 timestamp of data collectionYes

consentTypenone | opt-in | opt-out | contractualYes

piiCategoriesArray of PII types presentIf applicable

biasAuditReference to bias assessment reportRecommended

retentionPolicyData retention period and deletion rulesRecommended

Dependency Graphs

DRD automatically builds a directed acyclic graph (DAG) of model dependencies, showing the relationships between foundation models, fine-tuned variants, distilled models, and their datasets.

dependency-graph.json

{
  "model": "content-guardian-v3",
  "method": "fine-tuning",
  "parent": {
    "model": "llama-3-70b",
    "method": "supervised",
    "datasets": [
      { "name": "CommonCrawl-2025", "license": "CC-BY-4.0" },
      { "name": "Wikipedia-EN-2025", "license": "CC-BY-SA-4.0" }
    ]
  },
  "datasets": [
    { "name": "drd-content-safety-v2", "license": "proprietary", "consentType": "contractual" },
    { "name": "drd-policy-violations", "license": "proprietary", "consentType": "opt-in" }
  ],
  "children": [
    {
      "model": "content-guardian-v3-lite",
      "method": "distillation"
    }
  ]
}

License Propagation

DRD validates license compatibility across the dependency graph. If a parent model uses a copyleft license, all derivatives are flagged for compliance review.

Registering Models

lineage.ts

import { DRD } from '@drd/sdk';

const drd = new DRD({ apiKey: process.env.DRD_API_KEY });

const record = await drd.modelLineage.createRecord({
  modelId: 'custom-classifier-v3',
  modelName: 'Custom Content Classifier',
  version: '3.0.0',
  parentModelId: 'base-model-v2-id',
  trainingMethod: 'fine_tuning',
  datasetSize: 50000,
  license: 'Apache-2.0',
  trainingDataSources: {
    datasets: ['licensed-dataset-a', 'internal-dataset-b'],
    consentVerified: true,
  },
});

Compliance Audits

Create audits to verify model compliance with licensing, data usage, and regulatory requirements.

audit.ts

await drd.modelLineage.createAudit({
  recordId: record.id,
  auditType: 'compliance',
  findings: {
    dataSourcesVerified: true,
    licenseCompatible: true,
    consentChainValid: true,
  },
  complianceStatus: 'compliant',
  nextAuditDue: '2026-06-01T00:00:00Z',
});

REST API

POST /v1/models/lineage

curl -X POST https://api.drd.io/v1/models/lineage \
  -H "Authorization: Bearer drd_ws_sk_live_Abc123..." \
  -H "Content-Type: application/json" \
  -d '{
    "modelId": "mdl_01JM7XBN4RTYP",
    "name": "content-guardian-v3",
    "version": "3.0.0",
    "method": "fine-tuning",
    "parentModelId": "mdl_00BASE70BLLAMA",
    "datasets": [
      {
        "name": "drd-content-safety-v2",
        "source": "s3://drd-datasets/content-safety-v2",
        "license": "proprietary",
        "version": "2.1.0",
        "recordCount": 2500000,
        "consentType": "contractual"
      }
    ],
    "hyperparameters": {
      "learningRate": 2e-5,
      "epochs": 3,
      "batchSize": 32,
      "warmupSteps": 500
    },
    "metrics": {
      "accuracy": 0.96,
      "f1": 0.94,
      "latencyP99Ms": 45
    }
  }'

// Response
{
  "ok": true,
  "data": {
    "id": "lin_01JM7XBN4RTYP",
    "modelId": "mdl_01JM7XBN4RTYP",
    "name": "content-guardian-v3",
    "method": "fine-tuning",
    "parentModelId": "mdl_00BASE70BLLAMA",
    "datasetsCount": 1,
    "createdAt": "2026-02-14T12:00:00Z",
    "signedHash": "sha256:a1b2c3d4..."
  }
}

GET /v1/models/lineage/:modelId/graph

curl "https://api.drd.io/v1/models/lineage/mdl_01JM7XBN4RTYP/graph?depth=3" \
  -H "Authorization: Bearer drd_ws_sk_live_Abc123..."

// Response
{
  "ok": true,
  "data": {
    "root": "mdl_01JM7XBN4RTYP",
    "nodes": [
      { "id": "mdl_01JM7XBN4RTYP", "name": "content-guardian-v3", "method": "fine-tuning" },
      { "id": "mdl_00BASE70BLLAMA", "name": "llama-3-70b", "method": "supervised" },
      { "id": "mdl_02LITE7XBN4RT", "name": "content-guardian-v3-lite", "method": "distillation" }
    ],
    "edges": [
      { "from": "mdl_00BASE70BLLAMA", "to": "mdl_01JM7XBN4RTYP", "relation": "parent" },
      { "from": "mdl_01JM7XBN4RTYP", "to": "mdl_02LITE7XBN4RT", "relation": "parent" }
    ]
  }
}

Full SDK Example

lineage-workflow.ts

import { DRD } from '@drd/sdk';

const drd = new DRD({ apiKey: process.env.DRD_API_KEY });

// Register a new model lineage record
const lineage = await drd.modelLineage.createRecord({
  modelId: 'mdl_01JM7XBN4RTYP',
  modelName: 'content-guardian-v3',
  version: '3.0.0',
  parentModelId: 'mdl_00BASE70BLLAMA',
  trainingMethod: 'fine_tuning',
  datasets: [{
    name: 'drd-content-safety-v2',
    source: 's3://drd-datasets/content-safety-v2',
    license: 'proprietary',
    recordCount: 2_500_000,
    consentType: 'contractual',
  }],
});

// Query the dependency graph
const graph = await drd.modelLineage.graph(lineage.modelId, { depth: 3 });
console.log(`Nodes: ${graph.nodes.length}, Edges: ${graph.edges.length}`);

// Check license compatibility
const licenseCheck = await drd.modelLineage.checkLicenses(lineage.modelId);
if (licenseCheck.conflicts.length > 0) {
  console.warn('License conflicts found:', licenseCheck.conflicts);
}

// Get full audit trail
const trail = await drd.modelLineage.auditTrail(lineage.modelId);
for (const entry of trail) {
  console.log(`${entry.timestamp}: ${entry.action} by ${entry.actor}`);
}

Next Steps

Provenance

Track model decisions

Learn more →

Royalty Splitting

Compensate data providers

Learn more →

Explainability

Explain model decisions

Learn more →

Training Verification

Verify training data provenance and consent

Learn more →