# OSHUN V1 Metis Media and Runtime Tiers

Status: Active

Date: April 10, 2026

Owners: OSHUN Metis Engineering, OSHUN Product, OSHUN Psyche, OSHUN Yemaya,
OSHUN Safety

Related:

- `docs/releases/v1/reports/metis-domain-gap-analysis-and-full-integration-report-2026-03-23.md`
- `docs/releases/v1/reports/metis-subsystem-ownership-matrix.md`
- `docs/releases/v1/reports/metis-v1-scorecard.md`
- `docs/releases/v1/reports/metis-v1-workstream-tracking.md`
- `docs/releases/v1/runbooks/metis-operator-workflows-runbook.md`
- `docs/releases/v1/design/ai-disclosure-copy-guidance.md`
- `docs/releases/v1/specs/todos.md`
- `docs/releases/v1/domains/metis/README.md`
- `docs/releases/v1/domains/psyche/README.md`
- `docs/releases/v1/domains/yemaya/README.md`

## 1. Purpose

This document defines the canonical Metis tier split for educational teacher
representations in OSHUN V1.

The split is required because Metis now spans three materially different
delivery modes:

1. prerecorded lecturer
2. live voice tutor
3. live avatar tutor

These must not be treated as cosmetic skins on one generic tutor runtime. They
differ in:

- learner expectation
- latency and interruption posture
- exportability
- consent and likeness risk
- operations and rollout posture
- owning substrate

This document establishes the product-tier and runtime-tier model that all later
Metis media work must follow.

## 2. Research Inputs Reviewed on April 10, 2026

Current official product patterns still support the same separation identified
in the March 23, 2026 Metis gap analysis:

- OpenAI Realtime API guidance recommends WebRTC for browser speech-to-speech
  experiences and treats realtime voice interaction as a distinct low-latency
  application category.
- Tavus documents conversational video as a real-time WebRTC session with a
  modular stack for transport, STT, LLM, TTS, and realtime replica rendering.
- HeyGen documents its streaming avatar path as a low-latency WebRTC experience,
  and explicitly directs new builds toward the current live avatar product line
  instead of older interactive-avatar naming.
- ElevenLabs documents a dedicated low-latency conversational voice stack and
  positions real-time voice interaction separately from batch audio generation.
- Synthesia documents dubbing, multilingual training-video generation, and
  SCORM-capable export as asynchronous media-delivery workflows rather than live
  conversational tutoring.

Inference:

- The market does not treat polished course video and real-time tutoring as one
  runtime.
- Live voice and live avatar products share some tutoring logic, but they do not
  share the same embodiment, safety, or latency envelope.
- Metis should therefore define separate product tiers and separate runtime
  tiers rather than hiding all three modes under a generic “AI teacher” label.

Primary sources:

- OpenAI Realtime API with WebRTC:
  `https://platform.openai.com/docs/guides/realtime-webrtc`
- Tavus Conversational Video Interface overview:
  `https://docs.tavus.io/sections/conversational-video-interface`
- Tavus Conversation overview:
  `https://docs.tavus.io/sections/conversational-video-interface/conversation`
- HeyGen Streaming API overview: `https://docs.heygen.com/docs/streaming-api`
- ElevenLabs Conversational AI introduction:
  `https://elevenlabs.io/docs/conversational-ai/docs/introduction/`
- Synthesia AI Dubbing docs: `https://docs.synthesia.io/docs/video-dubbing`
- Synthesia video overview and SCORM export docs:
  `https://docs.synthesia.io/docs/video-overview`

## 3. Core Distinction

### 3.1 Product Tier

A product tier is the learner-facing promise Metis is making.

It answers:

- what kind of experience the learner is entering
- what controls the learner expects
- whether the mode is synchronous or asynchronous
- what can be exported, replayed, resumed, or embedded

### 3.2 Runtime Tier

A runtime tier is the technical execution posture required to fulfill that
promise.

It answers:

- whether generation is batch or real time
- which substrate owns transport and rendering
- what latency budget class applies
- what failure and fallback policies are allowed

Rule:

- Product tiers may not be introduced without mapping to one canonical runtime
  tier.
- Runtime tiers may not be reused across product tiers if doing so obscures
  distinct safety, cost, approval, or export behavior.

## 4. Canonical Tier Model

### 4.1 Product Tiers

```ts
type MetisTeacherProductTier =
  | 'prerecorded_lecturer'
  | 'live_voice_tutor'
  | 'live_avatar_tutor';
```

#### `prerecorded_lecturer`

Learner promise:

- polished lesson delivery
- replayable and packageable media
- stable narration and presentation quality
- export-oriented instructional content

Primary use cases:

- lesson delivery
- multilingual lecture packs
- LMS distribution
- previewable course trailers or narrated modules

#### `live_voice_tutor`

Learner promise:

- low-latency spoken tutoring
- interruptible back-and-forth dialogue
- office-hours style help without visual embodiment
- continuity across active study sessions

Primary use cases:

- question answering during study
- oral walkthroughs
- spoken remediation and coaching
- hands-free tutoring on mobile or desktop

#### `live_avatar_tutor`

Learner promise:

- low-latency spoken tutoring with visible teacher embodiment
- face, gesture, lip-sync, and turn-taking coherence
- stronger presence for guided explanation, demo, or office-hours posture

Primary use cases:

- premium tutoring sessions
- guided problem solving with embodiment
- office hours or live class support
- visually anchored coaching or demonstration flows

### 4.2 Runtime Tiers

```ts
type MetisTeacherRuntimeTier =
  | 'rendered_media_pipeline'
  | 'realtime_voice_session'
  | 'realtime_avatar_session';
```

#### `rendered_media_pipeline`

Execution posture:

- asynchronous or queued
- batch render oriented
- exportable and cacheable
- publication-package compatible

Primary owner:

- Yemaya for rendered asset pipelines

#### `realtime_voice_session`

Execution posture:

- synchronous
- low-latency conversational turn-taking
- streaming transport and session state
- interruption and resume aware

Primary owner:

- Psyche realtime voice runtime

#### `realtime_avatar_session`

Execution posture:

- synchronous
- low-latency conversational turn-taking
- embodied rendering and lip-sync
- stricter transport, GPU, and likeness controls

Primary owner:

- Psyche realtime avatar runtime

### 4.3 Allowed Mapping

| Product tier           | Canonical runtime tier    | Reason                                                                 |
| ---------------------- | ------------------------- | ---------------------------------------------------------------------- |
| `prerecorded_lecturer` | `rendered_media_pipeline` | lecture delivery is asset-first, replayable, and export-oriented       |
| `live_voice_tutor`     | `realtime_voice_session`  | tutoring is conversational but does not require embodied rendering     |
| `live_avatar_tutor`    | `realtime_avatar_session` | tutoring is conversational and embodied, with avatar-specific controls |

Hard rule:

- `prerecorded_lecturer` must not be implemented as a hidden “live avatar tutor
  with recording enabled.”
- `live_voice_tutor` must not be modeled as a degraded avatar session by
  default.
- `live_avatar_tutor` may use `live_voice_tutor` as a controlled fallback, but
  only through explicit learner-visible disclosure and policy-aware downgrade
  handling.

## 5. Tier Requirements by Mode

### 5.1 `prerecorded_lecturer`

Required characteristics:

- chaptered or lesson-bound media output
- deterministic transcript, caption, and subtitle package
- reviewable script and narration artifacts
- publication and export eligibility checks
- durable provenance and rights metadata
- acceptable queue-time and batch-job posture

Not required:

- sub-second turn-taking
- interruption handling
- live session continuity

Primary dependencies:

- Metis for lesson sequencing and pedagogy intent
- Yemaya for audio, video, slide, diagram, and lecture rendering
- Themis for rights and likeness enforcement
- Sophia and Isis where grounded claims and model provenance apply

### 5.2 `live_voice_tutor`

Required characteristics:

- active microphone and speaker session controls
- interruption-aware turn-taking
- low-latency STT, LLM, and TTS loop
- session resume, continuity, and escalation support
- transcript capture and learner-visible synthetic voice disclosure

Not required:

- visible teacher embodiment
- avatar render approval
- lip-sync or gesture coherence

Primary dependencies:

- Metis for tutoring product logic and learner context
- Psyche for realtime voice transport and orchestration
- Iris for continuity and learner preferences
- Lilith for pedagogy mode and teaching behavior
- Themis for safety, oversight, and integrity boundaries

### 5.3 `live_avatar_tutor`

Required characteristics:

- all `live_voice_tutor` requirements
- explicit avatar identity selection
- embodiment-safe animation and lip-sync
- stronger latency and degradation handling
- likeness consent, avatar-profile approval, and incident posture

Additional governance rule:

- `live_avatar_tutor` may not ship unless the relevant `AvatarTeacherProfile`
  and `VoiceLikenessConsentRecord` contracts are satisfied by the current
  release posture.

Primary dependencies:

- Metis for tutoring product logic and learner context
- Psyche for realtime avatar orchestration and rendering
- Iris and Lilith for continuity and pedagogy
- Themis for likeness, consent, moderation, and safety

## 6. Product and Runtime Rules

### 6.1 Entitlement Rule

Entitlements must be assigned at the product-tier level, not the runtime-tier
level.

Reason:

- learners purchase or are granted `prerecorded_lecturer`, `live_voice_tutor`,
  or `live_avatar_tutor`
- operators route that entitlement onto a runtime tier

### 6.2 Analytics Rule

Analytics must distinguish all three product tiers even when two of them share
parts of the tutoring brain.

Minimum analytics dimensions:

- product tier entered
- runtime tier invoked
- fallback or downgrade applied
- disclosure shown
- session or asset completion outcome

### 6.3 Review and Approval Rule

The approval model must stay split:

- prerecorded lecturer approvals are publication and media-governance oriented
- live voice tutor approvals are pedagogy, latency, and safety oriented
- live avatar tutor approvals are pedagogy, latency, safety, likeness, and
  embodiment oriented

### 6.4 Export Rule

Only `prerecorded_lecturer` is export-grade by default for LMS packaging and
asynchronous course distribution.

`live_voice_tutor` and `live_avatar_tutor` may export transcripts, summaries, or
evidence-backed artifacts, but they are not themselves lecture-export modes.

### 6.5 Continuity Rule

Continuity must also stay split:

- prerecorded lecturer continuity is asset progression and resume position
- live tutor continuity is session memory, learner state, and open thread
  recovery

### 6.6 Failure and Fallback Rule

Fallback must respect the tier model:

1. `live_avatar_tutor` may downgrade to `live_voice_tutor`
2. `live_voice_tutor` may downgrade to text tutoring
3. `prerecorded_lecturer` should fail into delayed render, alternate asset, or
   text transcript delivery, not into a surprise live session

## 7. Cross-Subsystem Ownership

| Concern                          | Canonical owner | Reason                                                                                 |
| -------------------------------- | --------------- | -------------------------------------------------------------------------------------- |
| product-tier selection           | Metis           | the learner-facing educational promise belongs to Metis                                |
| rendered lecture runtime         | Yemaya          | narrated and polished media generation is an asset-pipeline problem                    |
| live voice runtime               | Psyche          | streaming voice transport and turn-taking are realtime-runtime responsibilities        |
| live avatar runtime              | Psyche          | embodiment, rendering, and live avatar transport are realtime-runtime responsibilities |
| pedagogy policy across all tiers | Lilith          | teaching behavior should not fork by renderer without pedagogy ownership               |
| continuity and learner prefs     | Iris            | session memory and learner preference persistence remain cross-tier concerns           |
| rights, consent, and misuse      | Themis          | likeness, safety, and consent authority cannot live only in product code               |

Ownership rule:

- Metis owns the choice of tier exposed to the learner.
- Metis does not own a separate local realtime runtime for voice or avatar.
- Metis does not own a separate local batch-render pipeline for polished media.

## 8. Delivery Sequence for the Remaining Backlog

This tier model sets the required execution order for the remaining media tasks:

1. `V1-MET-088` implements `prerecorded_lecturer` on `rendered_media_pipeline`
2. `V1-MET-089` implements `live_voice_tutor` on `realtime_voice_session`
3. `V1-MET-090` formalizes `AvatarTeacherProfile` and
   `VoiceLikenessConsentRecord`
4. `V1-MET-092` adds controlled fallback from avatar to voice to text
5. `V1-MET-093` deepens real-time tutor session behavior through Psyche
6. `V1-MET-094` deepens lecture asset generation through Yemaya

Sequencing rule:

- Metis may ship `prerecorded_lecturer` and `live_voice_tutor` before
  `live_avatar_tutor` reaches launch grade.
- `live_avatar_tutor` must never be the prerequisite for all media work.

## 9. Canonical Exit Criteria for This Tier Definition

`V1-MET-087` is complete only if all of the following are true:

1. the three learner-visible teacher modes are explicitly named and distinct
2. each mode maps to exactly one canonical runtime tier
3. export, entitlement, analytics, review, and fallback rules differ by tier
4. ownership boundaries align with Metis, Psyche, Yemaya, Iris, Lilith, and
   Themis responsibilities already established elsewhere in OSHUN V1 docs

That condition is now satisfied by this document.
