Content Standards Implementation

Step-by-step guide to implementing AdCP Content Standards — calibration flow, artifact generation, delivery validation, and integration patterns.

This guide covers implementation patterns for the Content Standards Protocol from three perspectives:

1. Sales agents accepting and enforcing brand suitability standards
2. Orchestrators coordinating content standards across publishers
3. Governance agents providing content evaluation services

Roles Overview

Before diving in, understand who does what:

Role	Examples	Responsibilities
Orchestrator	DSP, trading desk, agency platform	Coordinates media buying; passes standards refs to sellers; receives artifacts for validation
Sales Agent	Publisher ad server, SSP	Accepts standards; calibrates local model; enforces during delivery; pushes artifacts
Governance Agent	IAS, DoubleVerify, brand suitability service	Hosts standards; implements calibrate_content and validate_content_delivery

The typical flow:

1. Brand sets up standards with governance agent (via orchestrator)
2. Orchestrator sends standards_ref with get_products/create_media_buy
3. Sales agent accepts or rejects based on capability
4. Sales agent calibrates against governance agent
5. Sales agent enforces during delivery
6. Sales agent provides artifacts (push via webhook or pull via get_media_buy_artifacts)
7. Orchestrator forwards artifacts to governance agent for validation

---

For Sales Agents

If you’re a sales agent (publisher ad server, SSP, or platform), implementing Content Standards means accepting orchestrator policies and enforcing them during delivery.

The Core Model

When an orchestrator includes a content_standards_ref in their request, you must:

1. Fetch the standards from the governance agent and evaluate if you can fulfill them
2. Accept or reject the buy based on your capabilities
3. Calibrate your evaluation model against the governance agent’s expectations
4. Enforce the standards during delivery
5. Provide artifacts to the orchestrator for validation

If you cannot fulfill the content standards requirements, reject the buy. Don’t accept a campaign you can’t properly enforce.

What You Need to Implement

1. Accept content standards references on get_products and create_media_buy

Orchestrators pass their standards via reference:

{
  "content_standards_ref": {
    "standards_id": "nike_emea_brand_safety",
    "agent_url": "https://brandsafety.ias.com"
  }
}

When you receive this:

• Fetch the standards document from the governance agent at agent_url
• Evaluate whether you can enforce these requirements
• If you cannot meet the standards, reject the request
• If you can, accept and store the association with the media buy

2. Decide: Can you fulfill this?

The standards document contains:

• Policy (natural language description of acceptable/unacceptable content)
• Calibration exemplars (pass/fail examples to interpret edge cases)
• Floor (reference to external baseline safety standards)

Review these requirements against your capabilities. Different publishers have different definitions of “adjacency” - Reddit might include comments, YouTube might include related videos, a news site might mean the article body. That’s fine - as long as you can meaningfully enforce the brand’s intent, accept the buy.

If you can’t - for example, they need adjacency data for a channel where it doesn’t apply (like billboards) - reject the buy.

3. Build your evaluation capability

Use the standards document to train or configure your content evaluation system. This could be:

• An LLM with the rules as system prompt
• A classifier trained on the calibration examples
• A rules engine for deterministic evaluation
• A third-party brand suitability vendor

The protocol doesn’t prescribe your implementation - just that you honor the standards.

4. Calibrate against the governance agent

After accepting the buy, calibrate your local model by calling calibrate_content on the governance agent. You send sample artifacts from your inventory, they tell you how they would rate them:

// You send examples from your inventory to the governance agent
{
  "standards_id": "nike_emea_brand_safety",
  "artifacts": [
    {
      "property_id": { "type": "domain", "value": "espn.com" },
      "artifact_id": "article_123",
      "assets": [{ "type": "text", "role": "title", "content": "Marathon Runner Collapses at Finish Line" }]
    }
  ]
}

// Governance agent responds with their interpretation
{
  "evaluations": [{
    "artifact_id": "article_123",
    "suitable": true,
    "confidence": 0.9,
    "explanation": "Sports injury coverage in athletic context - aligns with brand's sports marketing positioning"
  }]
}

Use these responses to train your local model. If you disagree with a rating, ask follow-up questions to understand the governance agent’s reasoning.

5. Push artifacts to the orchestrator

After delivery, push artifacts to the orchestrator so they can validate against the governance agent. Configure via artifact_webhook in the media buy:

// Artifact webhook payload (you send this to the orchestrator)
{
  "idempotency_key": "artw_01HW9DGRM7NQ0S4U6W8Y0A2C4E",
  "media_buy_id": "mb_nike_reddit_q1",
  "batch_id": "batch_20250115_001",
  "timestamp": "2025-01-15T11:00:00Z",
  "artifacts": [
    {
      "artifact": {
        "property_id": { "type": "domain", "value": "reddit.com" },
        "artifact_id": "r_fitness_abc123",
        "assets": [{ "type": "text", "role": "title", "content": "Best protein sources" }]
      },
      "delivered_at": "2025-01-15T10:30:00Z",
      "impression_id": "imp_abc123"
    }
  ]
}

Also support get_media_buy_artifacts for orchestrators who prefer to poll.

Implementation Checklist

• Parse content_standards_ref in get_products and create_media_buy
• Fetch and evaluate standards documents from governance agents
• Reject buys you cannot fulfill - don’t accept campaigns you can’t enforce
• Build content evaluation against the standards document
• Call calibrate_content on the governance agent to align interpretation
• Implement get_media_buy_artifacts so orchestrators can retrieve content for validation
• Support artifact_webhook for push-based artifact delivery
• Support reporting_webhook for delivery metrics

---

For Orchestrators

If you’re an orchestrator (DSP, trading desk, or agency platform), you coordinate content standards between brands, governance agents, and publishers.

The Orchestration Pattern

Brand → Orchestrator → Governance Agent (setup)
                    → Sales Agent (buying)
                    ← Sales Agent (artifacts)
                    → Governance Agent (validation)
                    → Brand (reporting)

1. Help brands set up standards with governance agents

Brands create content standards through a governance agent. You might facilitate this or the brand may do it directly:

// Standards stored at the governance agent
{
  "standards_id": "nike_emea_brand_safety",
  "name": "Nike EMEA Brand Suitability Policy",
  "brand_id": "nike",
  "policies": [
    { "policy_id": "no_violence", "policy_categories": ["brand_safety"], "enforcement": "must", "policy": "Avoid violence." },
    { "policy_id": "no_adult_themes", "policy_categories": ["brand_safety"], "enforcement": "must", "policy": "Avoid adult themes." },
    { "policy_id": "no_drug_content", "policy_categories": ["brand_safety"], "enforcement": "must", "policy": "Avoid drug content." }
  ],
  "calibration_exemplars": {
    "pass": [
      { "type": "url", "value": "https://espn.com/nba/story/_/id/12345/lakers-win", "language": "en" }
    ],
    "fail": [
      { "type": "url", "value": "https://tabloid.example.com/celebrity-scandal", "language": "en" }
    ]
  }
}

Migrating prose policies to addressable entries

Brands with existing prose policies have two options: keep everything as one policy entry with the whole prose blob, or split into one entry per rule. Both validate.

One-entry-with-prose (easiest, preserves existing authoring):

"policies": [
  {
    "policy_id": "acme_creative_standards_v1",
    "enforcement": "must",
    "policy": "No violent imagery. Minimum 72 DPI for display assets. Brand colors must match palette within 5% tolerance."
  }
]

Multiple addressable entries (recommended for programmatic fix/retry loops and per-rule versioning):

"policies": [
  { "policy_id": "no_violent_imagery",   "policy_categories": ["brand_safety"],     "enforcement": "must",   "policy": "No violent imagery." },
  { "policy_id": "min_display_dpi",      "policy_categories": ["imagery_quality"],  "enforcement": "should", "channels": ["display"], "policy": "Minimum 72 DPI for display assets." },
  { "policy_id": "brand_color_tolerance","policy_categories": ["brand_compliance"], "enforcement": "must",   "policy": "Brand colors must match palette within 5% tolerance." }
]

Splitting enables findings to reference specific rules (via policy_id), programmatic fix/retry, and stable ids across versions. Use whichever authoring mode fits the workflow.

Registry policies + bespoke policies

Both registry_policy_ids and policies can be provided together. The evaluator applies all of them. Bespoke policy_id values MUST be flat (no colons or slashes); registry policy ids are always namespaced (e.g., garm:brand_safety:violence), so the two namespaces never collide. When a governance finding references a specific policy, policy_id on the finding carries either the registry id or the bespoke id — the namespace tells you which.

2. Pass standards references when buying

When discovering products or creating media buys, include the governance agent reference:

{
  "product_id": "espn_sports_display",
  "packages": [...],
  "content_standards_ref": {
    "standards_id": "nike_emea_brand_safety",
    "agent_url": "https://brandsafety.ias.com"
  },
  "artifact_webhook": {
    "url": "https://your-platform.com/webhooks/artifacts",
    "authentication": {
      "schemes": ["HMAC-SHA256"],
      "credentials": "your-shared-secret-min-32-chars"
    },
    "delivery_mode": "batched",
    "batch_frequency": "hourly",
    "sampling_rate": 0.25
  }
}

If the publisher cannot fulfill the standards, they should reject the buy. Handle rejections gracefully and find alternative inventory.

3. Receive artifacts from sales agents

Sales agents push artifacts to your artifact_webhook endpoint. Forward them to the governance agent for validation:

# Receive artifact webhook from sales agent
@app.post("/webhooks/artifacts")
async def receive_artifacts(payload: ArtifactWebhookPayload):
    # Forward to governance agent for validation
    validation_result = await governance_agent.validate_content_delivery(
        standards_id=get_standards_id(payload.media_buy_id),
        records=[
            {"artifact": a.artifact, "record_id": a.impression_id}
            for a in payload.artifacts
        ]
    )

    # Log any failures
    for result in validation_result.results:
        if any(f.status == "failed" for f in result.features):
            log_suitability_incident(payload.media_buy_id, result)

    return {"status": "received", "batch_id": payload.batch_id}

4. Report to brands

Surface validation results to the brand:

• Incidents: Content that didn’t meet standards
• Coverage: What percentage of delivery was validated
• Trends: Changes in content safety over time

Implementation Checklist

• Facilitate brand setup with governance agents
• Include content_standards_ref in get_products and create_media_buy requests
• Configure artifact_webhook to receive artifacts from sales agents
• Handle rejections from publishers who can’t fulfill standards
• Forward artifacts to governance agent via validate_content_delivery
• Build reporting for brands

---

For Governance Agents

If you’re a governance agent (IAS, DoubleVerify, or brand suitability service), you provide content evaluation as a service.

What You Implement

1. Host and serve content standards

Store standards configurations and expose them via get_content_standards:

// Response to get_content_standards
{
  "standards_id": "nike_emea_brand_safety",
  "version": "1.2.0",
  "name": "Nike EMEA - all digital channels",
  "policies": [
    { "policy_id": "no_violence", "policy_categories": ["brand_safety"], "enforcement": "must", "policy": "Avoid violence." },
    { "policy_id": "no_adult_themes", "policy_categories": ["brand_safety"], "enforcement": "must", "policy": "Avoid adult themes." },
    { "policy_id": "prefer_sports_fitness", "policy_categories": ["brand_suitability"], "enforcement": "should", "policy": "Sports and fitness content is ideal." }
  ],
  "calibration_exemplars": {
    "pass": [...],
    "fail": [...]
  }
}

2. Implement calibrate_content

Sales agents call this to align their local models before campaign execution. They send sample artifacts, you respond with how the brand would rate them:

def calibrate_content(standards_id: str, artifacts: list) -> dict:
    standards = get_standards(standards_id)
    evaluations = []

    for artifact in artifacts:
        # Evaluate against brand's policy
        result = evaluate_against_policy(artifact, standards)
        evaluations.append({
            "artifact_id": artifact["artifact_id"],
            "suitable": result.suitable,
            "confidence": result.confidence,
            "explanation": result.explanation  # Help them understand your reasoning
        })

    return {"evaluations": evaluations}

Calibration is a dialogue - be prepared for follow-up questions and edge cases.

3. Implement validate_content_delivery

Orchestrators call this to validate artifacts after delivery. Batch evaluation at scale:

def validate_content_delivery(standards_id: str, records: list) -> dict:
    standards = get_standards(standards_id)
    results = []

    for record in records:
        features = []
        for feature in ["brand_safety", "brand_suitability"]:
            evaluation = evaluate_feature(record["artifact"], standards, feature)
            features.append({
                "feature_id": feature,
                "status": "passed" if evaluation.passed else "failed",
                "value": evaluation.value,
                "message": evaluation.message if not evaluation.passed else None
            })
        results.append({
            "record_id": record["record_id"],
            "features": features
        })

    return {
        "summary": compute_summary(results),
        "results": results
    }

Implementation Checklist

• Implement create_content_standards for brands to set up policies
• Implement get_content_standards for sales agents to fetch policies
• Implement calibrate_content for sales agents to align their models
• Implement validate_content_delivery for orchestrators to validate delivery
• Support dialogue in calibration (follow-up questions, edge cases)

---

Content Access Pattern

All three roles may need to exchange content securely. The content_access pattern provides authenticated access to a URL namespace:

{
  "content_access": {
    "url_pattern": "https://cache.example.com/*",
    "auth": {
      "type": "bearer",
      "token": "eyJ..."
    }
  }
}

• url_pattern: URLs matching this pattern use this auth
• auth.type: Authentication method (bearer, api_key, signed_url)
• auth.token: The credential

Include this in:

• get_content_standards response (governance agent → sales agent: “fetch examples here”)
• get_media_buy_artifacts response (sales agent → orchestrator: “fetch content here”)

This avoids per-asset tokens and keeps payloads small while enabling secure content exchange.