Add Audience Alignment to Signals Protocol

[bokelley]

Summary

Add audience alignment capabilities to the Signals Protocol, parallel to how Content Standards handles content alignment. This enables buyers to describe who they're looking for and calibrate that understanding with sellers through dialogue.

Background

Content Standards Protocol handles:

• Standards: Hard rules for content exclusion (safety, suitability)
• Calibration: Dialogue to align on content interpretation

Audience Alignment would handle:

• Stories: Positive descriptions of target audiences ("cat yoga people", "wellness-minded millennials")
• Calibration: Dialogue to align on audience interpretation

Key Concepts

Stories vs Standards

• Standards (Content): "Here's what I don't want" → pass/fail (binary)
• Stories (Audience): "Here's who I'm looking for" → affinity matching (gradient)

Example: Cat Yoga Story

A buyer doesn't just want "cat yoga content" - they want to reach "cat yoga people" or "people who like what cat yoga people like" (wellness, flexibility, mindfulness, pet lovers).

Stories are defined through examples:

• "Someone browsing morning stretching routines? Perfect."
• "Someone researching competitive powerlifting? Close but not quite."
• "Someone reading about meditation retreats? Yes, that's the vibe."

Calibration Flow

Same pattern as content calibration:

1. Buyer describes target audience (natural language + examples)
2. Seller's signal agent evaluates against available signals
3. Back-and-forth dialogue to refine understanding
4. Seller trains local model for runtime matching

Applies Across Dimensions

Stories could work for:

• Audiences: "People who exhibit cat-yoga-adjacent behaviors"
• Context: "Moments when people are relaxed and browsing"
• Content: "Articles about wellness, mindfulness, pets"

Proposed Tasks

Task	Description
create_audience_story	Define a target audience through description + examples
calibrate_audience	Dialogue to align on audience interpretation
match_audience	Runtime evaluation of signals against a story

Questions to Resolve

1. Is the response pass/fail or a score/ranking?
2. How do stories compose with standards? (e.g., "cat yoga people" + "no political content")
3. Should stories be first-class objects with CRUD, or embedded in signal activation?

Related

• Content Standards Protocol (parallel pattern for content)
• Signals Protocol activate_signal (where audience matching would apply)

[bokelley]

The key here is that we want to be able to evaluate a "vector" of characteristics from the signals provider. This could be derived from first-party data about the user, but could also be inferred from a postal code, page context, or other sources. Example:

buyer: "I want cat yoga people" --> geo-intent company finds 200 postal codes where this kind of people live and shares the attributes of those postal codes --> ["suburban dads", "love football", "don't have pets"]
buyer: No! That's not right, that doesn't seem like my people. Try again.

Then, post-campaign, the seller could send vectors of audience attributes through to the buyer and share of delivery: ["cohort": {"dads", "50-59", "no pets"}, 3.5%] and then the buyer can use this to decide how "on target" this was. Of course all of this presumes that the seller here is honest... but of course this is where measurement and studies come to bear around validating this

[flaviolivolsi]

Building a buyer agent currently. Here's my take on the open questions:

1. Response format

For the match_audience response, I'd suggest including a confidence indicator alongside the score:

{
  "match_score": 0.73,
  "threshold_met": true,
  "confidence": 0.85
}

confidence tells my agent how much evidence backs the score - was this inferred from a postal code alone or matched against rich first-party signals? That context affects how aggressively I bid

2. How stories compose with standards

Standards filter first (hard pass/fail from the seller). Stories rank what remains, but buyers can also set their own minimum story threshold to filter further

A publisher can be 100% compliant (standards pass) but 10% relevant (low story score). Useful signal that lets the buyer decide whether to bid and how much

3. First-class objects vs embedded

I'd say first-class objects. I'll reuse the same story across many campaigns so embedding would create duplication. Stories should be immutable, versioned objects referenced by ID

Follow-up question: if I update story_v1 to story_v2, should active campaigns inherit the change or stay pinned to the version they started with? (pinned seems safer for budget predictability)

On calibrate_audience

I believe the spec should support calibration where the buyer refines their story based on conversion data they observe privately, hence no outcome data shared with the seller

1. Buyer defines initial story
2. Seller serves impressions
3. Buyer observes conversions (privately on their side)
4. Buyer refines the story definition and pushes story_v2

This would keep conversion data private while still allowing the targeting to learn over time

[bokelley]

Closing as superseded. The conversational get_signals refinement pattern now covers the core use case: signal_spec handles natural language audience description, and signal_ids + signal_spec provides the calibration dialogue loop. The three proposed tasks (create_audience_story, calibrate_audience, match_audience) and persistent story objects aren't needed given the current design. If relevance/match scoring on returned signals becomes a priority, that's a smaller, separate feature request.