Add Content Standards Protocol #621
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Introduces the Content Standards Protocol with four tasks: - list_content_features: Discover available content safety features - get_content_standards: Retrieve content safety policies - check_content: Evaluate content context against safety policies - validate_content_delivery: Batch validate delivery records Features use a generic type system (binary, quantitative, categorical) aligned with property governance patterns. Policies use prompt-based evaluation (like Scope3) rather than keyword lists. Standards can be scoped by country, brand, channel, and product. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Remove sentiment analysis (subset of suitability) - Clarify brand safety vs suitability distinction: - Safety = safe for ANY brand (universal thresholds) - Suitability = safe for MY brand (brand-specific) - Replace three separate prompts with single policy + examples: - Single policy prompt for natural language guidelines - Examples object with acceptable/unacceptable URLs as training set - Update schema and documentation to match 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Remove links to Property Governance and Creative Standards which don't exist in this branch yet. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Address review comments: 1. Examples now use typed identifiers with language tags: - Uses identifier schema pattern (type: domain, apple_podcast_id, rss_url, etc.) - Added language field (BCP 47 tags like 'en', 'en-US') - Created content-example.json schema 2. Scope fields aligned with property protocol: - brands → brand_ids (references Brand Manifest identifiers) - countries → countries_all (must apply in ALL listed countries) - channels → channels_any (applies to ANY listed channel) - Removed products field (unclear purpose) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
The buyer selects the appropriate standards_id when creating a media buy. The seller receives a reference to resolved standards - they don't need to do scope matching themselves. Scope fields are metadata for the buyer's internal organization, not for seller-side resolution. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
1. Scope is targeting, not metadata:
- Clarified scope defines where standards apply
- Removed "metadata" language
2. Add CRUD tasks for standards management:
- Added list_content_standards, create_content_standards,
update_content_standards, delete_content_standards
- Organized tasks into Discovery, Management, Evaluation sections
- Created task documentation pages
3. Clarify evaluation results pattern:
- Always returns pass/fail verdict first
- Features array is optional breakdown for debugging
- Explains use cases: optimization, safety vs suitability,
debugging third-party standards (GARM, Scope3)
- Updated check-content-response and validate-content-delivery-response
schemas to use verdict + optional features pattern
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Address review comments:
1. Rename 'examples' to 'calibration':
- Better describes the purpose (training/test set for AI interpretation)
- Add support for text snippets, image/video/audio URLs
2. Create unified content-context.json schema:
- Used for both calibration examples and check_content requests
- Supports: domain, text, image_url, video_url, audio_url, rss_url,
apple_podcast_id, spotify_show_id, podcast_guid
- Optional metadata: language, title, description, categories, text_content
3. Flatten scope structure:
- Move brand_ids, countries_all, channels_any to top level
- Aligns with property protocol pattern
4. Remove fixed response time claims:
- check_content may be async with webhook callback
- Let implementations define their own SLAs
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Integrates comprehensive content-context schema improvements: 1. Signals array (replaces simple categories): - Structured signals with key, value, confidence - Supports version, source, and reasons - Examples: iab_category, violence_score, mpaa_rating 2. Structured text_content (replaces simple string): - Array of typed blocks: paragraph, heading, image, video, audio - Images include alt_text, caption, dimensions, type - Video/audio include duration, transcript, transcript_source 3. Rich metadata object: - Open Graph, Twitter Card, JSON-LD - Author, canonical URL, meta tags 4. Temporal fields: - published_time, last_update_time (ISO 8601) 5. artifact_id for cross-reference tracking 6. additionalProperties: true for extensibility Credit: sit@ca367da 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
bokelley
changed the title
🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Rename task to emphasize dialogue-based alignment vs runtime evaluation - Add context_id and feedback parameters for multi-turn conversations - Include verbose explanations with confidence scores and policy alignment - Add workflow diagram showing Setup → Activation → Runtime phases - Update schemas with calibrate-content-request/response - Remove old check-content schemas - Update related task links across all task docs Key insight: calibrate_content is low-volume, verbose, dialogue-based; runtime decisioning happens locally at seller for scale. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Remove context_id and feedback from calibrate_content task - dialogue is handled at the protocol layer via A2A contextId or MCP context_id. Updated documentation to show how multi-turn calibration conversations work using the existing protocol mechanisms for conversation management. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Addresses feedback: 1. Rename "content-context" to "content" - too many "context" terms 2. Restructure to use assets array pattern like creative assets Content is now represented as a collection of assets (text, images, video, audio) plus metadata and signals. This aligns with the creative asset model and avoids overloading "context" terminology. Key changes: - content.json replaces content-context.json - assets array with typed blocks (text, image, video, audio) - identifiers object for platform-specific IDs (apple_podcast_id, etc.) - url is now the primary required field - Updated all docs and schema references 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Content adjacent to ads is now modeled as an "artifact" - identified by property_id (using existing identifier types) plus artifact_id (string defined by property owner). This avoids overloading "content" and enables identification of artifacts that don't have URLs (Instagram, podcasts, TV scenes). Changes: - Rename content.json to artifact.json with new required fields - property_id uses existing identifier type schema (type + value) - artifact_id is a string - property owner defines the scheme - format_id optional - can reference format registry (like creative formats) - url now optional since not all artifacts have URLs - Update all schema references and documentation examples 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Address review feedback: 1. Remove signals from artifact schema and examples - the verification agent evaluates content directly without pre-computed classifications. This simplifies the model: send content, get responses. 2. Add secured URL access pattern for private assets - supports bearer tokens, service accounts, and pre-signed URLs for content that isn't publicly accessible (AI-generated images, private conversations, paywalled content). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Restructure Content Standards docs around the key strategic questions: 1. What content? → New artifacts.mdx page with full schema details 2. How much adjacency? → Define in products, negotiate in media buys 3. What sampling rate? → Negotiate coverage vs cost tradeoffs 4. How to calibrate? → Dialogue-based alignment process Changes: - Add standalone artifacts.mdx with asset types and secured access - Replace ASCII workflow with mermaid sequence diagram - Add adjacency and sampling_rate sections to overview - Simplify policy examples (remove verbose calibration details) - Move secured URL access documentation to artifacts page 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
…t_id - Make assets array required (artifacts without content are meaningless) - Remove title, language, description top-level fields (they're text assets) - Add variant_id for A/B tests, translations, and temporal versions - Add language field to text assets for mixed-language content - Update all documentation examples to use new pattern - Fix schema validation test to handle internal $defs references 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Rename content_standards to content_standards_adjacency_definition - Add Adjacency Units table defining: posts, scenes, segments, seconds, viewports, articles - Clarify human-in-the-loop aspect of calibration process 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Address review comments: - Remove suggestion field from features (not appropriate for calibration) - Remove policy_alignment (redundant with features breakdown) - Make confidence explicitly optional in docs - Simplify A2A examples to use data-only parts (no text when calling skill) - Update response field table to show required vs optional 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Address review comment about who calls validate_content_delivery: - Delivery records flow from seller → buyer → verification agent - Buyer aggregates delivery data and knows which standards_id applies - Verification agent works on behalf of the buyer Changes: - Add Data Flow section with mermaid diagram to validate_content_delivery.mdx - Add media_buy_id field to request schema and delivery records - Update overview mermaid diagram to show correct flow - Remove suggestion field from response examples 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Content artifacts (text, images, video) are separate from delivery metrics. This new task allows buyers to request content samples from sellers for validation, keeping the data flow clean: - Buyer calls get_media_buy_artifacts to get content samples - Buyer validates samples with verification agent - Delivery metrics stay in get_media_buy_delivery Updates validate_content_delivery flow to reference new task. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
1. Remove redundant top-level media_buy_id - keep at record level since records can come from multiple media buys 2. Add country and channel fields to delivery records for targeting context validation 3. Replace creative_id with brand_context placeholder object - the governance agent needs brand/SKU info, not opaque creative IDs. Schema marked as TBD pending brand identifier standardization. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
sit
reviewed
Jan 8, 2026
|
|
||
| **Schema**: [artifact.json](https://adcontextprotocol.org/schemas/v2/content-standards/artifact.json) | ||
|
|
||
| An artifact represents content adjacent to an ad placement - identified by `property_id` + `artifact_id` and represented as a collection of assets: |
Contributor
There was a problem hiding this comment.
Suggested change
| An artifact represents content adjacent to an ad placement - identified by `property_id` + `artifact_id` and represented as a collection of assets: | |
| An artifact represents content context where ad placements occur - identified by `property_id` + `artifact_id` and represented as a collection of assets: |
Contributor
Author
There was a problem hiding this comment.
@sit why do you think "context where placements occur" is more clear? l can imagine adjacent not being descriptive but do you have an example?
| { | ||
| "errors": [ | ||
| { | ||
| "code": "STANDARDS_NOT_FOUND", |
Contributor
There was a problem hiding this comment.
Is it important to add this to the list of standard error codes?
I wonder if worth documenting an error like "STANDARD_IN_USE".
| { | ||
| "type": "object", | ||
| "description": "Success response", | ||
| "properties": { |
Contributor
There was a problem hiding this comment.
Wonder if it's worth having a standard.json to reference, similar to artifact.json. Then this and the create_content_standards could have a common schema and we could validate all the fields are the same.
Contributor
Author
There was a problem hiding this comment.
intuitively why would describing my standards have a similar schema to describing the artifact we're assessing? or do you mean "break out the object"
Comment on lines
20
to
24
| "effective_date": { | ||
| "type": "string", | ||
| "format": "date-time", | ||
| "description": "When this version became effective" | ||
| }, |
Contributor
There was a problem hiding this comment.
Can a standard be effective at multiple times, applied to different campaigns?
Would looking for a null or "" be a way to determine which standards are unused (e.g., for the purposes of deletion)? Or maybe it has to be inferred by listing media buys and seeing which ones have standards active.
Or if this intended to indicate when we should start using this standard as opposed to when we started using it ("became effective" vs "becomes effective")— in that case what happens when there are multiple effective at the same time, and should there be a "termination_date" concept.
| "type": "string", | ||
| "description": "Natural language policy describing acceptable and unacceptable content contexts" | ||
| }, | ||
| "calibration": { |
Contributor
There was a problem hiding this comment.
I could see this getting quite large and blowing up context windows easily.
| - A **scene** in a CTV show | ||
| - An **AI-generated image** in a chat conversation | ||
|
|
||
| Artifacts are identified by `property_id` + `artifact_id` - the property defines where the content lives, and the artifact_id is the property owner's identifier for that specific piece of content. |
Contributor
There was a problem hiding this comment.
FWIW, that's not what we do today — artifact_id in our current system is a domain/path combo. And we don't ask the property owner for the id of course, we take it from the URL itself.
For Meta, we use meta:<meta_id> as a distinguisher, even though we also have a source: META field.
Comment on lines
+199
to
+202
| "access": { | ||
| "method": "bearer_token", | ||
| "token": "eyJhbGciOiJIUzI1NiIs..." | ||
| } |
Contributor
There was a problem hiding this comment.
That's a lot of tokens (e.g., we've seen web pages where we extract hundreds of images... whether or not that's a good idea).
I'd consider extracting an auth mechanism that could be referenced, either within the response or outside the workflow.
| ```json | ||
| { | ||
| "type": "image", | ||
| "url": "https://cdn.openai.com/secured/img_abc123.png", |
Contributor
There was a problem hiding this comment.
We could consider distinguishing the published official URL from some internal access URL (e.g.,a GCS bucket or something).
| |-----------|------|----------|-------------| | ||
| | `brand_ids` | array | No | Filter by brand identifiers | | ||
| | `countries` | array | No | Filter by country codes | | ||
| | `channels` | array | No | Filter by channels | |
Contributor
There was a problem hiding this comment.
Could add a filter for effective date?
|
|
||
| ```json | ||
| { | ||
| "standards": [ |
Contributor
There was a problem hiding this comment.
Could reference schema here, not sure if these examples are intentionally abbreviated or not (e.g., do you want to see the calibration data in this response).
- #1/#16: Update artifact description to "content context where placements occur" - #2: Add STANDARDS_IN_USE error code to delete_content_standards - #4: Clarify effective_date semantics (null = not yet effective, multiple allowed) - #5: Add note about calibration data size concerns - #7: Document scope conflict handling in create_content_standards - #8/#9: Rename examples to calibration for consistency - #10: Increase validate_content_delivery response time to 60s - #11: Clarify artifact_id is opaque/flexible - #12/#13: Add notes about auth mechanism extraction for large artifacts - #14/#15: Add effective_date filter and abbreviated response note to list 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
Adds the Content Standards Protocol - draft governance protocol for content safety and suitability rules.
docs/governance/content-standards/static/schemas/source/content-standards/list_content_features,get_content_standards,list_content_standards,check_contentDraft for AdCP 3.0 - Part of the AAO Governance Working Group initiatives.
Test plan
mintlify devnpm test🤖 Generated with Claude Code