The Emission Surface

What the Emission Surface Is

Every time someone asks the world a question — whether through the browser, the REST API, an MCP tool, or a directive — the Protocol Emission Engine (PEE) runs an inference turn. The output of that turn is not a chat message. It is a structured, layered document called the Emission Surface.

The Emission Surface carries everything a consumer needs to present, analyze, or build on what the world just thought: the prose response, every entity and setting the reasoning touched, resolved imagery for each, atmospheric signals from the settings involved, how dense or sparse the territory is, what claims were made and how confident the system is, and session context that carries forward into the next turn.

This is what makes Starholder's output a media substrate rather than a text response. A documentary agent reads the prose and media layers. A podcast agent reads the prose and epistemic layers. An interactive experience reads the refs and topology layers. A research tool reads the claims and contradictions. They all consume the same document — they just use different parts of it.

Schema identifier: starholder.emission.v1

---

How to Get the Emission Surface

Three transport options, same payload:

Layer Projection

You don't always need all eight layers. Add ?layers=text,refs,context to any endpoint that returns a surface and you'll get only those layers (plus identity, which is always included). This reduces payload size when you only care about specific parts of the output.

---

The Eight Layers

---

Identity Layer

Provenance for the turn. Every field you need to know who triggered it, when, and under what conditions.

Directive block (when present):

---

Text Layer

The persona's prose output in multiple parallel forms.

Mention fields:

Wire note: In the compacted wire format, mentions are not sent separately. Instead, each ref in the refs map carries a spans array of [start, end] pairs. Reconstruct the mentions array from these spans if you need it.

---

Refs Layer

Everything the system considered, used, or rejected during retrieval and reasoning.

Catalog

The full retrieval universe — every entity, setting, story, and prior reasoning packet the system found relevant during graph traversal. This is the "considered" set, not just what made it into the output.

Accepted

Refs the persona actually referenced in its output — confirmed through validation.

Anomalies

Refs the persona tried to use but were rejected — either not found in retrieval results or forbidden by policy.

Wire note: In the compacted format, refs are a single Record<string, WireSurfaceRef> keyed by ref string. Anomalies have status: 'anomaly' and a reason field. Everything else is accepted. The score field carries the relevance score.

---

Media Layer

Resolved imagery for every entity and setting referenced, with per-mention placement context.

Bundles

One bundle per referenced entity or setting, containing the best available images ranked and categorized.

Placements

Maps specific text mentions to specific images. If Dr. Chen is mentioned three times in different contexts, each mention gets its own best-match image based on the surrounding sentence.

---

Context Layer

The richest layer for immersive applications. Contains atmospheric signals, environment selection hints, music, and semantic scene profiling. This layer has several independent components that serve different purposes.

Shader / Environment Selection

Used by renderers to select visual environments, color palettes, and mood.

Structured Atmosphere

Sensory signals extracted from the dominant setting — the most relevant setting the persona referenced during reasoning. These are parsed from structured descriptors stored on each setting record.

When no settings are referenced in the turn, all atmosphere fields are null. This is correct — there is no setting data to surface. Shader selection still works because contextText is always populated from the output text and entity labels.

Canonical Setting

The dominant setting that atmosphere was extracted from, with confidence scoring.

Semantic Scene Profile

A structured analysis of the turn's semantic content, categorized into thematic phrase buckets. Useful for advanced rendering that adapts to the semantic character of the turn.

Derivation Mode

How the context layer was assembled — useful for debugging and quality assessment.

Music

Server-side music selection based on phrase extraction from the turn's output.

Primary Subject

The dominant entity or setting for this turn — the "main character" of this particular reasoning act.

Selection priority: dominant setting (if any) > highest-relevance entity > null.

---

Topology Layer

Retrieval quality metrics. Tells you how dense or sparse the territory is, what drove the retrieval, and how much prior reasoning the persona had to build on.

For builders: Use sparsity to frame output confidence. If it's void, warn users they're in uncharted territory. If inferentialCount is high, the output is well-grounded in prior analysis. Entertainment products can narrate this: "This connection is well-documented across seventeen prior explorations" vs "We're venturing into thin territory here — the historical record is sparse."

---

Epistemic Layer

What the persona claims to know and how confidently it knows it.

Claims

Specific factual assertions the persona made, with the evidence backing them.

Contradictions

When the persona's reasoning encounters conflicting information in the knowledge graph.

For builders: Contradictions are narrative texture, not errors. A podcast frames reconcilable contradictions as "competing accounts." A documentary presents both sides. An interactive experience lets users investigate. Only irreconcilable contradictions indicate actual data quality issues.

---

Session Layer

Accumulated state across turns. This is what gives multi-turn conversations coherence.

What resets between turns: identity, text, refs, media, context, topology, epistemic — all reset to empty when a new turn starts.

What persists: The session layer carries forward. The system remembers the conversation's accumulated knowledge, momentum, and open threads.

surface_complete is authoritative: When received, it replaces everything — including the session layer — with the server's canonical snapshot.

---

SSE Event Lifecycle

When consuming via the streaming endpoint, events arrive in this order during a single turn:

PEE Inner Events

These events arrive wrapped in a pee_event envelope on the hub stream. They provide additional operational visibility:

Hub-Only Events

These appear only on the browser hub stream (POST /api/world-program/hub), not on the external agent stream:

External-Only Events

---

Quick Reference: Event to Layer Mapping