# Themis Universal Shield Reviewer Guide

This guide is for reviewers who clear, modify, reject, or escalate Universal
Shield review items across audio, music, voice, image, text, video, 3D, motion,
character, fashion, and environment content.

Human review is required when a Universal Shield scan returns `review`,
`remediate`, or `reject`, or when a domain policy requires review before
publication. Reviewers should not approve or reject from the automated risk
level alone. The reviewer task is to inspect the evidence, use the correct
comparison tools, document a rationale, and choose a defensible action.

---

## Reviewer Roles

Universal review work is routed by media type and risk:

| Expertise              | Primary scope                                                 |
| ---------------------- | ------------------------------------------------------------- |
| `musicologist`         | music, audio samples, melody, harmony, rhythm, structure      |
| `voice_rights`         | voice likeness, consent evidence, biometric or persona claims |
| `visual_arts`          | image composition, style, layout, visual trade dress          |
| `text_editorial`       | prose, lyrics, dialogue, translation, paraphrase              |
| `video_editorial`      | shots, sequencing, edits, storyboards, audiovisual compounds  |
| `three_d_asset`        | 3D topology, silhouette, materials, rig compatibility         |
| `motion_specialist`    | mocap, keyframes, choreography, signature movement            |
| `character_specialist` | silhouette, costume, backstory, franchise traits              |
| `fashion_specialist`   | garment silhouette, patterns, tech packs, trade dress         |
| `environment_designer` | scene layout, landmarks, level design, architectural motifs   |
| `legal`                | high-risk, reject, uncertain precedent, major release impact  |

High-risk and reject items route to legal or senior review first. Compound
matches route to the most specialized reviewer plus legal when the same
reference work appears across multiple media types.

---

## Queue Workflow

Review items move through these statuses:

- `pending`: item exists and awaits assignment.
- `assigned`: item has an assigned reviewer.
- `in_review`: reviewer has locked the item.
- `decided`: reviewer decision has been recorded.
- `escalated`: reviewer needs additional expertise or evidence.

Priority runs from `1` to `5`; `1` is highest. Promote priority when:

- `overallRiskLevel` is `high_risk` or `reject`.
- A compound match spans more than one media type.
- The release is public, paid, major-account, or time-sensitive.
- The reference work is commercially sensitive or heavily enforced.
- Voice, biometric, likeness, child-safety, or jurisdiction metadata is missing.
- SLA is near breach.

Default active review SLA is 4 hours. Items enter alert state at 75 percent of
SLA and breach at 100 percent. Breached items should be reassigned or escalated.

### Start A Review

1. Confirm the `reviewId`, `scanId`, `generationId`, domain, and release gate.
2. Read `overallRiskLevel`, `actionRequired`, `detectedMediaTypes`, and
   `routedShieldIds`.
3. Open every `componentResult` before writing a decision.
4. Identify matched `referenceWorkId` values and whether one reference appears
   across multiple components.
5. Open the comparison tool for each affected media type.
6. Check provenance, consent, license, and reference database context.
7. Decide, request modification, reject, or escalate with a rationale.

Do not clear a review item when the evidence packet is incomplete.

---

## Evidence Package

Every all-media review should include:

- Universal scan response with `overallRiskLevel`, `actionRequired`,
  `detectedMediaTypes`, `routedShieldIds`, and `compoundMatches`.
- Component results with `componentId`, media type, shield id, status, score,
  risk level, findings, and reference work id.
- Candidate artifact or preview URL for every reviewed component.
- Matched reference work metadata: title, creator, rightsholder, catalog,
  public-domain status, license status, and protectability notes.
- Domain request metadata: prompt, model id, model version, seed, workflow,
  user/account id where allowed, release gate, and jurisdiction.
- Provenance: generation id, input hash, normalized artifact hash, scan engines,
  cache status, and audit level.
- Prior decisions, reviewer rationales, remediation attempts, and rescan links.
- Consent or rights evidence for voice, likeness, biometric, and licensed source
  material.

If any required evidence is missing, choose `escalate` or request a database or
integration correction.

---

## Comparison Tools

Use the tool for the component under review. A compound review may require
several tools before the final decision.

| Media type  | Primary comparison tool                           | Use                                             |
| ----------- | ------------------------------------------------- | ----------------------------------------------- |
| audio/music | A/B segment player, element isolation, heatmap    | Compare melody, rhythm, sample, vocal, mix      |
| voice       | voice-likeness panel, consent ledger, transcript  | Confirm likeness signal and consent status      |
| image       | side-by-side viewer, overlay, crop grid, heatmap  | Compare composition, subject placement, palette |
| text        | aligned diff, paraphrase map, phrase heatmap      | Compare expressive sequence and wording         |
| video       | shot-by-shot viewer, timeline heatmap, keyframes  | Compare shot order, blocking, edits, visuals    |
| model_3d    | mesh overlay, silhouette viewer, topology report  | Compare topology, proportions, materials        |
| motion      | skeleton overlay, curve viewer, tempo alignment   | Compare key poses, timing, transitions          |
| character   | turnaround comparison, trait matrix, lore checker | Compare silhouette, costume, identity signals   |
| fashion     | garment overlay, pattern diff, tech-pack compare  | Compare silhouette, cut, pattern, ornament      |
| environment | map overlay, landmark matrix, path analysis       | Compare level layout, landmarks, scene grammar  |

### Comparison Principles

- Inspect the generated artifact alone before viewing the reference.
- Inspect the reference alone before side-by-side comparison.
- Compare the matched region or segment first, then the whole artifact.
- Separate protectable expression from ideas, genre conventions, functional
  constraints, and public-domain material.
- Record what changed after remediation; do not rely on the prior review.
- When multiple components point to the same reference work, assess cumulative
  similarity.

---

## Interpreting Similarity Scores

Scores are evidence, not decisions. A high score means the adapter found a
strong signal under its method; it does not automatically mean infringement.

| Signal                     | Review meaning                                             |
| -------------------------- | ---------------------------------------------------------- |
| High component score       | A media-specific shield found strong similarity            |
| Multiple medium scores     | Cumulative risk may be higher than any component alone     |
| Same `referenceWorkId`     | Possible compound match across media or components         |
| `compoundMatches` present  | Gateway found a cross-component originality concern        |
| Low confidence or failures | Evidence may be incomplete and should be escalated         |
| Public-domain reference    | Composition may be clear, but recording or edition may not |
| Consent metadata missing   | Voice, likeness, or biometric review cannot be cleared     |

Common false-positive patterns:

- Genre-standard chord progressions, poses, camera moves, garment cuts, or level
  archetypes.
- Short phrases, titles, idioms, or common prompt vocabulary.
- Generic character archetypes without distinctive trade dress.
- Public-domain material with no protected arrangement or recording copied.
- Functional 3D topology dictated by object class or manufacturing constraints.
- Similar mood, theme, style, or color palette without protected expression.

High-risk patterns:

- Distinctive hook, lyric sequence, composition, shot, pose, topology, or layout
  copied with minimal transformation.
- Same reference work matched in several components.
- Recognizable voice or persona match without verified consent.
- Character, fashion, or environment trade dress that creates source
  association.
- Near-whole-work copy, distinctive-part copy, or repeated matched segments.
- Missing license or rightsholder metadata for a high-scoring reference.

---

## Media-Specific Review Techniques

### Audio And Music

- Listen to the candidate and reference independently, then A/B by matched
  segment.
- Use element isolation for melody, harmony, rhythm, bass, vocal, and sample
  signals.
- Treat common loops and chord progressions as lower risk unless combined with a
  distinctive melody, hook, lyric, or sample.
- Check whether pitch, tempo, or timbre changes still leave the protected
  expression recognizable.

### Voice

- Confirm whether the scan is about performance similarity, identity likeness,
  or explicit voice-clone consent.
- Open the consent ledger and verify scope: speaker, model, use case, territory,
  duration, and revocation status.
- Compare transcript content separately from voice identity.
- Escalate missing consent, celebrity/public-figure matches, minor voice
  signals, or biometric retention uncertainty.

### Image

- Compare composition, subject placement, pose, lighting, palette, and
  distinctive visual elements.
- Use crop grids and overlays to distinguish general style from copied layout.
- Check whether the reference is a protected work, public-domain work,
  trademarked character, licensed source, or training-only reference.
- Escalate when the match depends on a living artist's recognizable style and
  policy requires legal or creator-relations review.

### Text

- Use aligned diff and paraphrase maps; exact overlap is not required for risk.
- Focus on expressive sequence, character-specific phrasing, dialogue cadence,
  plot beats, and translated paraphrase.
- Treat short phrases, facts, recipes, procedures, and generic instructions as
  lower risk unless selection and arrangement are distinctive.
- Preserve paragraph order in the rationale when sequence drives the match.

### Video

- Compare shot ordering, blocking, camera movement, edits, and audiovisual
  synchronization.
- Use keyframe grids for fast triage, then inspect the matched timeline range.
- Check whether video risk comes from visuals, music, text, character, or
  compound evidence.
- Escalate famous scene recreations, near-shot-for-shot matches, and music-video
  compound matches.

### 3D Models

- Compare silhouette, topology, mesh density, UV layout, materials, rigging, and
  distinctive proportions.
- Functional objects can share broad shape; focus on nonfunctional expressive
  details and asset-pack identifiers.
- Check whether the match is to a licensed base mesh, marketplace asset, scanned
  object, or generated derivative.
- Require rescan after decimation, retopology, or material replacement.

### Motion

- Compare skeleton overlays, timing curves, key poses, transitions, and loop
  structure.
- Distinguish common athletic, dance, or yoga movement from signature
  choreography or mocap capture.
- Check whether the motion is bound to a protected character, performer, or
  licensed mocap pack.
- Escalate celebrity performance matches and high-scoring choreography
  sequences.

### Character

- Compare silhouette, costume, color blocking, props, abilities, backstory,
  naming, and catchphrases as a combined identity signal.
- Generic archetypes are not enough; distinctive trade dress or source
  association raises risk.
- Check whether the output uses a permitted parody, licensed franchise asset, or
  internal character bible.
- Escalate child-directed releases, high-value franchise matches, and policy
  override requests.

### Fashion

- Compare garment silhouette, cut, pattern placement, textile print, hardware,
  ornament, and lookbook context.
- Functional garment shapes can be common; distinctive trade dress or protected
  pattern art is higher risk.
- Open the tech pack and pattern diff when available.
- Escalate runway-look copies, brand-identifying trade dress, and unresolved
  license claims.

### Environment

- Compare map topology, player path, landmark placement, room sequence,
  architecture, lighting plan, and prop clusters.
- Common genre rooms and gameplay affordances are lower risk without distinctive
  arrangement.
- Use path analysis to see whether the traversal experience copies the
  reference.
- Escalate near-identical level layouts, iconic scenes, and licensed venue
  recreations.

---

## Decision Types

Allowed decisions are:

| Decision   | Use when                                                       |
| ---------- | -------------------------------------------------------------- |
| `approve`  | Evidence does not support blocking or modification             |
| `modify`   | Publication may proceed only after listed changes and rescan   |
| `reject`   | Publication should be blocked                                  |
| `escalate` | Additional expertise, legal review, or missing evidence needed |

Decision rules:

- Every decision requires `reviewerId`, expertise, timestamp, and rationale.
- Rationale must describe evidence, tool usage, and why the outcome follows.
- `modify` requires concrete changes and a rescan condition.
- `reject` requires the release-blocking evidence and reference work ids.
- `escalate` requires a target reviewer group and what evidence is missing.
- Reviewer decisions must never delete the original automated scan response.

Example decision:

```json
{
  "decision": {
    "reviewId": "universal-review-123",
    "decision": "modify",
    "rationale": "The character sheet and costume component both preserve the same protected franchise silhouette and color blocking. Modify the costume, change the prop language, and rescan the full character bundle.",
    "conditions": [
      "Replace protected color blocking on costume component",
      "Remove franchise-specific prop silhouette",
      "Run a new Universal Shield scan on the complete character bundle"
    ],
    "reviewerExpertise": "character_specialist",
    "reviewerId": "reviewer-42"
  }
}
```

---

## Writing Rationale

A good rationale answers:

- What component matched?
- Which reference work matched?
- Which comparison tools were used?
- What is protected or not protected about the overlap?
- Why is the result approval, modification, rejection, or escalation?
- What must happen before publication?

Good:

> The generated environment uses the same landmark order, central room shape,
> and traversal loop as reference-environment-layout-1. Individual sci-fi wall
> panels are generic, but the map overlay and path analysis show a distinctive
> sequence. Modify the level layout and rescan before publication.

Insufficient:

> Looks different enough.

Do not speculate about intent. Review the artifact, reference, and evidence.

---

## Escalation

Escalate when:

- Required media expertise is unavailable.
- Legal context is unclear or high impact.
- Consent, license, public-domain, or rightsholder metadata is missing.
- A compound match spans multiple media types and at least one component is
  `flag` or higher.
- A famous person, franchise, brand, venue, or major artist is implicated.
- Reviewer assessment and automated evidence materially disagree.
- The scan adapter failed, skipped unexpectedly, or returned incomplete
  findings.
- The case could change threshold calibration or future policy.

Escalation targets:

- `senior_reviewer`: complex media-specific analysis or reviewer disagreement.
- `legal`: high-risk, reject, jurisdiction, claim, or license issues.
- `reference_database_admin`: missing, duplicate, stale, or disputed reference
  metadata.
- `consent_operations`: voice, likeness, biometric, or persona consent gaps.
- `external_expert_panel`: major release or uncertain expert evidence.

Escalated rationale should state what question the next reviewer must answer.

---

## Reviewer Quality Controls

Review quality is tracked through:

- false positive and false negative counts from later claims and appeals;
- reviewer agreement rate on double-reviewed items;
- escalation correctness;
- SLA adherence without rushed rationale;
- modification success rate after rescan;
- per-media calibration notes.

Reviewers should complete monthly calibration against the continuous validation
corpus and known original controls. A reviewer who repeatedly clears missing
evidence packets or writes insufficient rationales should return to supervised
review until calibration is restored.

---

## Do Not Clear Checklist

Do not approve when any of the following are true:

- The candidate artifact or reference artifact is unavailable.
- The matched reference work lacks rights or protectability metadata.
- Voice, likeness, biometric, or persona consent is missing or out of scope.
- The scan has a failed component without an operational waiver.
- A compound match is present but only one component was reviewed.
- The requested publication territory lacks required jurisdiction context.
- Remediation was requested but no rescan result is attached.
- The rationale would not explain the decision to another reviewer, counsel, or
  auditor.
