Historical task source. The Eve task board holds current task state.
Phase 75: Themis Universal Originality Shield — Cross-Media IP Protection for All Generated Content#
Purpose: Extend the Themis Music Shield (Phase 74) to cover EVERY form of media generated across the entire Oshun ecosystem — visual art, video, text, 3D models, characters, narrative, animation, fashion, game environments, and more. Build unified cross-media compound detection, universal reference databases, universal provenance/review/remediation/analytics, watermarking and content authentication, and cross-media model governance. This is the complete IP protection layer that gates all 31 content-generating domains.
Scope: ~60 new libraries under libs/themis/ (TypeScript + Rust WASM + ONNX
Runtime for ML inference). Integrates with Phase 74 Music Shield and all
generating domains: Calliope, Euterpe, Isis, Hathor, Veritas, Maya, Yemaya, Aja,
Galatea, Aglaea, Psyche, Iris, Nous, Metis, Nisaba, Aphrodite, Bellona, Hestia.
Reference proposal:
docs/domains/themis-universal-originality-shield-proposal.md.
Key Design Principles:
- Media-specific engines, shared infrastructure — Each media type has detection engines tuned to its characteristics; risk scoring, provenance, review, and analytics are shared across all.
- Cross-media compound detection — A music video is music + visual + text; compound matches across media types targeting the same source escalate non-linearly.
- Legally calibrated thresholds — Every threshold grounded in case law and per-media-type protectability analysis.
- Progressive scanning — Fast pre-filters first, expensive deep scans only on content that survives pre-filters.
75.1 Visual Originality Shield — Core & Composition (@themis/visual-shield)#
Protects all AI-generated images, artwork, illustrations, design assets, and visual branding across Isis, Calliope, Yemaya, Aglaea, Galatea, Metis.
75.1.1 Project Scaffolding & Core Types#
- 75.1.1.1 Create
libs/themis/visual-shield/project scaffolding (project.json with Nx targets for build/test/lint; package.json as@themis/visual-shield; tsconfig.json extending base; vitest.config.ts; src/index.ts barrel export) - 75.1.1.2 Define visual shield core types (VisualScanRequest with image as Uint8Array | URL string, width/height, format 'png'|'jpg'| 'webp'|'svg', generationPrompt optional string, modelId, metadata; VisualScanResult with overallRisk RiskLevel, compositionScore, embeddingScore, characterScore, faceScore, trademarkScore, memorizationScore, matches array of VisualMatch, scanDurationMs; VisualMatch with referenceId, referenceTitle, matchType 'composition'| 'embedding'|'character'|'face'|'trademark'|'memorization'|'watermark', similarity 0-1, boundingBox optional, matchedRegion description)
- 75.1.1.3 Define VisualScanEngine interface (name string, scan method accepting image + config returning element score, supported match types array, estimateLatency method)
- 75.1.1.4 Implement VisualShieldPipeline orchestrator (coordinates all visual scan engines in parallel; pre-processes image — decode, normalize to RGB, resize to standard analysis resolution max 1024px; dispatches to enabled engines concurrently; aggregates results; fast-path mode with perceptual hash + CLIP only for < 2s latency; full scan runs all engines target < 5s)
75.1.2 Image Composition Analysis#
- 75.1.2.1 Implement GIST descriptor extraction (extract holistic scene descriptor — 512D vector capturing spatial envelope via oriented Gabor filters at multiple scales and orientations; captures overall layout: subject placement, horizon line, perspective, spatial frequency distribution; compare via cosine similarity against reference database; threshold > 0.80 = compositionally similar)
- 75.1.2.2 Implement spatial pyramid matching (divide image into 1×1, 2×2, 4×4 grids hierarchically; extract feature histogram per cell using visual bag-of-words or CNN features; concatenate weighted histograms into single vector; compare pyramid vectors via histogram intersection kernel; captures both global and local spatial arrangement)
- 75.1.2.3 Implement layout analysis (use object detection model to identify foreground subject, secondary elements, background; extract relative positions — subject centroid as fraction of image width/height, subject area as fraction of total, negative space distribution; compare layout descriptors between generated and reference images; two images with identical subject placement/framing flagged even with different content)
- 75.1.2.4 Implement color histogram comparison (extract 3D color histogram in HSV space — 16 hue × 4 saturation × 4 value = 256 bins; normalize to probability distribution; compare using chi-squared distance or Earth Mover's Distance; color alone is not protectable but combined with composition strengthens compound score)
- 75.1.2.5 Write composition analysis tests (test GIST on rotated/ flipped images — should detect similarity; test spatial pyramid on images with same layout different content; test layout — same subject position scores high; test color histogram — same palette scores high, inverted scores low; minimum 30 tests)
75.1.3 Neural Image Embedding Comparison#
- 75.1.3.1 Implement CLIP embedding extraction and comparison (load CLIP ViT-L/14 model via ONNX Runtime; extract 768D embedding per image; L2 normalize; compare via cosine similarity; CLIP captures semantic visual similarity — 'the same thing' in different styles scores high; thresholds: > 0.85 cosine = FLAG, > 0.92 = REJECT; batch inference support for throughput; cache embeddings by image content hash)
- 75.1.3.2 Implement DINOv2 embedding extraction and comparison (load DINOv2 ViT-B/14 or ViT-L/14 via ONNX Runtime; extract self-supervised visual embeddings 768D or 1024D; captures structural visual similarity without text-image alignment bias; better at detecting visual copies with different color grading or artistic filters than CLIP; compare via cosine similarity; threshold > 0.82 = FLAG)
- 75.1.3.3 Implement SSCD copy detection (load Meta's Self-Supervised Copy Detection model via ONNX; purpose-built for detecting copies, near-duplicates, and manipulated versions of images; more robust than generic embeddings for exact copy detection; output 512D descriptor; compare via cosine similarity; threshold > 0.75 = likely copy; invariant to crops, color shifts, overlays, compressions)
- 75.1.3.4 Implement perceptual hashing (implement pHash — DCT-based perceptual hash 64-bit; dHash — difference hash 64-bit; aHash — average hash 64-bit; compute all three for each image; compare via Hamming distance; pHash Hamming < 10 = near-duplicate; fast pre-filter — O(1) lookup against hash database in Redis; millions of images searchable in < 1ms)
- 75.1.3.5 Implement visual embedding database (store CLIP, DINOv2, SSCD embeddings in Qdrant vector database; support approximate nearest neighbor search with configurable top-K; index by reference work ID with metadata — title, artist, copyright holder, source; batch indexing for database population; incremental updates for new reference works)
- 75.1.3.6 Write neural embedding tests (test CLIP — semantically similar images score > 0.80, unrelated < 0.30; test DINOv2 — same image with color filter still matches; test SSCD — cropped/overlaid copy still detected; test perceptual hash — same image through JPEG compression has Hamming < 5; test database — indexed image retrieved as top-1 match; minimum 35 tests)
75.1.4 Character Recognition#
- 75.1.4.1 Implement fictional character detection classifier (fine- tune image classification model on dataset of copyrighted characters; character classes: Disney 50+ characters, Marvel 100+ characters, DC 50+ characters, Nintendo 50+ characters, anime 200+ characters, other major franchises; multi-label classification — image may contain multiple characters; output per-character probability; threshold > 0.70 = character detected; handle stylistic variations — realistic renders, chibi, sketch, photographic)
- 75.1.4.2 Implement character feature matching (extract character- defining visual features: dominant color palette via k-means on character region, silhouette shape via contour extraction, distinctive accessories via object detection, body proportions via pose estimation; encode as feature vector; compare against character database feature vectors via weighted Euclidean distance; catches characters that evade classifier — e.g., 'yellow humanoid with spiky hair' matching multiple characters)
- 75.1.4.3 Implement character name-in-prompt detector (NER on generation prompts to detect copyrighted character names; maintain database of 10,000+ protected character names with aliases and common misspellings; phonetic matching for creative spellings — 'Pikachoo', 'Micky Mouse'; flag BEFORE generation begins; block or inject negative constraints)
- 75.1.4.4 Implement character silhouette comparison (extract foreground silhouette using segmentation model; compare silhouette shape against database of iconic character silhouettes using Hu moments and Fourier descriptors; many characters identifiable by silhouette alone — Batman, Mickey Mouse, Darth Vader, Pikachu; rotation and scale invariant comparison)
- 75.1.4.5 Write character recognition tests (test classifier on known character images — correct identification; test stylistic variation — realistic Pikachu still detected; test feature matching — character with right colors/proportions/accessories matched; test prompt detection — 'mickey mouse' flagged, 'a mouse cartoon' not flagged; test silhouette — Batman outline matched; minimum 35 tests)
75.1.5 Face & Person Recognition#
- 75.1.5.1 Implement face detection pipeline (load RetinaFace or MTCNN model via ONNX; detect all faces in generated image with bounding boxes and landmarks; filter by face size — minimum 48×48 pixels for reliable recognition; output face crops, landmarks, and detection confidence per face)
- 75.1.5.2 Implement face embedding extraction (load ArcFace or AdaFace model via ONNX; align face crops using landmarks; extract 512D face embedding per detected face; L2 normalize; these embeddings are biometric data — handle per GDPR requirements in EU jurisdictions)
- 75.1.5.3 Implement protected person database (PostgreSQL with pgvector for face embedding storage; store: person name, known images count, embedding centroid from multiple reference images, consent status 'consented'|'not_consented'|'unknown', jurisdiction notes; initial population: 100,000+ public figures/celebrities from open face datasets; support opt-in/opt-out management; GDPR-compliant data handling — purpose limitation, storage limitation, right to erasure)
- 75.1.5.4 Implement face similarity comparison (compare detected face embeddings against protected person database using pgvector ANN search; cosine similarity thresholds: > 0.70 = FLAG, > 0.85 = REJECT unless consent verified; return top-5 matches with similarity scores; handle multiple faces per image — each face checked independently; consent verification — if matched person has consented, bypass flag)
- 75.1.5.5 Implement deepfake indicator detection (analyze generated faces for synthetic media artifacts: inconsistent lighting/shadows, skin texture anomalies, eye reflection inconsistencies, hair boundary artifacts, temporal inconsistencies in video; output deepfake probability score; all AI-generated faces must be labeled per EU AI Act regardless of whether they match a real person)
- 75.1.5.6 Write face recognition tests (test detection — face found in generated image; test embedding — same person different images cosine > 0.85; different people cosine < 0.40; test database lookup — known celebrity matched; test consent bypass — consented person not flagged; test deepfake detection — synthetic face classified correctly; test GDPR — embeddings deletable on request; minimum 35 tests)
75.1.6 Trademark & Logo Detection#
- 75.1.6.1 Implement logo detection model (fine-tune object detection model — YOLOv8 or RT-DETR — on trademark logo dataset; 1,000+ brand logo classes covering major global brands; output bounding boxes with brand classification and confidence; handle partial occlusion, rotation, and scale variation; detect both clean logos and logos embedded in generated scenes)
- 75.1.6.2 Implement OCR text detection in images (use PaddleOCR or EasyOCR via ONNX to detect and extract text from generated images; compare extracted text against trademark database — brand names, product names, copyrighted phrases; flag images containing identifiable brand text; handle AI-generated garbled text that resembles brand names)
- 75.1.6.3 Implement product identification (detect recognizable commercial products in generated images — specific car models, phones, sneakers, luxury goods, electronics; use object detection + brand classification; flag when generated images contain clearly identifiable trademarked products that could constitute trademark use in commerce)
- 75.1.6.4 Write trademark detection tests (test logo detection — Nike swoosh detected in generated scene; test OCR — brand name text extracted; test product identification — specific car model recognized; test partial occlusion — partially visible logo still detected; minimum 25 tests)
75.1.7 Training Data Memorization Detection#
- 75.1.7.1 Implement training data memorization detector (maintain SSCD embedding database of representative samples from known AI training datasets — LAION-5B subset, CC12M, OpenImages; compare generated images against database; SSCD cosine > 0.85 = potential memorization; also check perceptual hash — pHash Hamming < 8 against training data = near-verbatim reproduction)
- 75.1.7.2 Implement watermark reproduction detector (train classifier to detect stock photo watermarks — Getty, Shutterstock, Alamy, iStock, Adobe Stock watermark patterns; scan generated images for watermark artifacts; any detected watermark is strong evidence of training data memorization and immediate REJECT)
- 75.1.7.3 Implement style attribution advisory (use neural style analysis to quantify how much a generated image's style resembles specific known artists; output as advisory only — style is not copyrightable; but when style similarity is extreme AND combined with compositional similarity, elevate to FLAG; compare against database of 1,000+ distinctive artist styles)
- 75.1.7.4 Write memorization detection tests (test near-duplicate of training image detected; test watermark detection — synthetic Getty watermark flagged; test style attribution — generated image in distinctive artist style flagged as advisory; minimum 20 tests)
75.1.8 Visual Shield Integration#
- 75.1.8.1 Implement composite visual similarity scorer (combine composition score × 0.10, neural embedding score × 0.30, character score × 0.25, face score × 0.20, trademark score × 0.10, memorization score × 0.05; if any single engine returns REJECT-level score, overall result is REJECT; output VisualScanResult with per-engine breakdown)
- 75.1.8.2 Implement Isis integration adapter (hook into Isis text-to- image pipeline; scan every generated image before delivery to client; block delivery until scan completes; return scan result with image; support batch mode for factory generation runs)
- 75.1.8.3 Implement Calliope integration adapter (hook into AI artist visual identity creation — album art, promotional images, social media visuals; per-artist configuration based on visual style; scan all visual assets before publication)
- 75.1.8.4 Write visual shield end-to-end integration tests (test full pipeline — submit image → all engines run → composite score → decision; test known copyrighted character in generated image → REJECT; test celebrity face → FLAG; test brand logo → FLAG; test completely original image → PASS; test near-duplicate of training data → REJECT; minimum 30 integration tests)
75.2 Text Originality Shield (@themis/text-shield)#
Protects all AI-generated text — articles, dialogue, narratives, lyrics, educational content, marketing copy, social posts across Veritas, Hathor, Iris, Nisaba, Metis, Calliope.
75.2.1 Project Scaffolding & Core Types#
- 75.2.1.1 Create
libs/themis/text-shield/project scaffolding (project.json, package.json as@themis/text-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.2.1.2 Define text shield core types (TextScanRequest with text string, contentType 'article'|'dialogue'|'narrative'|'educational'| 'marketing'|'social'|'lyrics'|'script'|'general', language string, generationPrompt optional, modelId, metadata; TextScanResult with overallRisk RiskLevel, verbatimScore, semanticScore, narrativeScore, characterScore, brandScore, citationScore, matches array of TextMatch, scanDurationMs; TextMatch with referenceId, referenceTitle, matchType, similarity 0-1, queryExcerpt, referenceExcerpt, positions)
75.2.2 Exact & Near-Exact Match Detection#
- 75.2.2.1 Implement verbatim text matching engine (tokenize text into words — lowercase, strip punctuation, normalize whitespace; extract character-level n-grams for n = 20, 50, 100 characters and word-level n-grams for n = 5, 8, 12 words; hash each n-gram using xxHash for speed; compare against reference corpus hash index via O(1) lookup; for each match, record matched text, position in query, reference work ID; score: weighted sum of matched n-grams / total, longer matches weighted exponentially — 50-word match scores 10× more than 5-word match)
- 75.2.2.2 Implement near-duplicate detection via MinHash/SimHash (compute MinHash signature — 128 hash functions on 3-gram shingle set; estimate Jaccard similarity from MinHash signature overlap; threshold > 0.60 Jaccard = FLAG as near-duplicate; also implement SimHash — 128-bit hash using random hyperplane projections; Hamming distance < 20 = structurally similar document; both methods support O(1) lookup via LSH banding for million-document scale)
- 75.2.2.3 Implement cross-language text matching (load multilingual sentence transformer — mUSE or LaBSE via ONNX; encode text passages in source language; compare against reference passages in all indexed languages; detect when generated English text closely translates a copyrighted French/Spanish/German/Japanese/Chinese text; threshold > 0.85 cross-lingual cosine = FLAG)
- 75.2.2.4 Write exact matching tests (test verbatim — 50-word passage from known book detected; test near-duplicate — paraphrased passage detected via MinHash; test cross-language — English translation of French poem detected; test false positive — common phrases not flagged; minimum 30 tests)
75.2.3 Semantic Similarity Analysis#
- 75.2.3.1 Implement passage-level embedding comparison (load sentence-transformers model — all-MiniLM-L6-v2 or BGE-small via ONNX; encode each paragraph as 384D or 768D dense vector; compare against reference database via Qdrant ANN search; two passages conveying the same specific idea in completely different words score > 0.85 cosine; return top-K matches per paragraph with similarity scores and reference excerpts)
- 75.2.3.2 Implement document-level embedding comparison (aggregate paragraph embeddings into document-level representation — weighted average by paragraph importance, or max-pooling; compare document vectors; catches cases where overall structure and argument flow are copied even though individual sentences differ; threshold > 0.80 = structurally similar documents)
- 75.2.3.3 Implement contextual sequence matching (compare not just individual passages but their sequential ordering; two texts presenting the same sequence of specific arguments in the same order is stronger evidence than scattered similar passages; encode paragraph sequences as ordered lists of embeddings; compare sequence similarity using edit distance on embedding sequences with soft matching)
- 75.2.3.4 Write semantic similarity tests (test paraphrase detection — same idea different words scores > 0.85; test unrelated passages score < 0.30; test document-level — copied structure detected; test sequential — same argument order flagged; minimum 25 tests)
75.2.4 Narrative & Plot Structure Analysis#
- 75.2.4.1 Implement plot event extraction (use LLM-based analysis to decompose narrative text into structured plot events: inciting incident, rising action events, climax, falling action, resolution; output as ordered list of PlotEvent objects with description, event type, characters involved, location, causal links to other events; support both short narratives and novel-length works)
- 75.2.4.2 Implement plot structure comparison (compare extracted plot event sequences between generated and reference narratives; use sequence alignment — edit distance on event types with soft semantic matching on event descriptions; filter common structures — hero's journey, three-act structure, rags-to-riches are not protectable; flag when specific event sequences with specific causal relationships match at the expression level, not just at the structural/idea level)
- 75.2.4.3 Implement character relationship graph extraction (extract character names via NER; identify relationships between characters from text — ally, enemy, romantic, familial, mentor/student, rival; construct character relationship graph; compare graphs between generated and reference works using graph isomorphism with soft matching on relationship types; isomorphic graphs with same relationship types suggest plot copying)
- 75.2.4.4 Implement world-building element extraction (extract named fictional locations, rules/laws of the fictional world — magic systems, physics rules, political structures; extract historical events within the fiction, faction/organization names, unique terminology; compare against structured database of known fictional universes; detect when generated content is set in an existing copyrighted fictional universe — Middle-earth, Westeros, Hogwarts, etc.)
- 75.2.4.5 Implement quest/mission structure comparison (for game narrative: extract quest objectives, branching decision points, reward structures, NPC interactions, prerequisite chains; compare against database of known game quests from major titles; flag when quest structure reproduces a specific known quest beyond generic patterns)
- 75.2.4.6 Write narrative analysis tests (test plot extraction on known story — correct events identified; test plot comparison — two versions of the same story matched; test character graph — isomorphic relationships detected; test world-building — Hogwarts-like magic school detected; test quest structure — specific quest pattern matched; minimum 35 tests)
75.2.5 Source Attribution & Brand Detection#
- 75.2.5.1 Implement factual claim detection (use NER + dependency parsing to identify factual claims — statistics, quotes, research findings, historical assertions; for each claim, check if attribution is present; flag unattributed claims that appear to originate from specific sources; for educational content, verify cited studies actually exist and citations are accurate)
- 75.2.5.2 Implement brand/trademark detection in text (NER for brand names, product names, trademarked terms; maintain database of 100,000+ trademarked phrases, slogans, and distinctive brand expressions — 'Just Do It', 'I'm Lovin' It', 'May the Force be with you', 'Winter is Coming'; flag when generated text uses distinctive trademarked phrases; differentiate nominative fair use from trademark infringement)
- 75.2.5.3 Implement common phrase filter (database of 50,000+ common text phrases, idioms, proverbs, genre clichés that are too common to be protectable; filter these from match results to reduce false positives; auto-expand: any phrase appearing in 200+ indexed works auto-classified as common)
- 75.2.5.4 Write attribution and brand tests (test claim detection — unattributed statistic flagged; test brand detection — 'Just Do It' detected as Nike trademark; test common phrase filter — 'once upon a time' not flagged; minimum 20 tests)
75.2.6 Text Shield Integration#
- 75.2.6.1 Implement composite text similarity scorer (combine verbatim score × 0.30, semantic score × 0.25, narrative/plot score × 0.20, character/world score × 0.10, brand score × 0.10, citation score × 0.05; content-type-aware weighting — journalism elevates verbatim weight, narrative elevates plot weight, educational elevates citation weight)
- 75.2.6.2 Implement text reference database (passages from published works indexed by: verbatim n-gram hashes in Redis for O(1) lookup, passage embeddings in Qdrant for semantic search, full text in PostgreSQL for display; sources: Project Gutenberg public domain corpus, licensed excerpt databases, news article rolling 90-day index, academic paper abstracts, fictional universe structured databases with character names/locations/lore/magic systems)
- 75.2.6.3 Implement domain integration adapters (Veritas — scan all generated articles/newsletters/social posts; Hathor — scan all generated dialogue/narrative/quest text; Iris — scan generated conversational responses for verbatim reproduction; Metis — scan educational content with citation verification; Nisaba — scan scholarly analysis; Calliope — scan artist bios and social content)
- 75.2.6.4 Write text shield end-to-end integration tests (test verbatim article passage → FLAG; test paraphrased passage → FLAG via semantic; test novel set in Westeros → FLAG via world-building; test character graph matching Lord of the Rings → FLAG; test completely original text → PASS; test branded slogan → FLAG; test common phrase → PASS; minimum 30 integration tests)
75.3 Video Originality Shield (@themis/video-shield)#
Protects all AI-generated video across Isis, Calliope, Veritas, Yemaya, Uzume.
75.3.1 Project Scaffolding & Types#
- 75.3.1.1 Create
libs/themis/video-shield/project scaffolding (project.json, package.json as@themis/video-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.3.1.2 Define video shield types (VideoScanRequest with video as Uint8Array | URL, durationMs, fps, resolution, hasAudio boolean; VideoScanResult with overallRisk, frameAnalysisResults array, sceneResults array, dialogueResults, audioResults from Music Shield, deepfakeScore, matches array, scanDurationMs; FrameResult with frameIndex, timestampMs, visualScanResult — delegated to Visual Shield; SceneResult with startMs, endMs, sceneType, matchedScenes array)
75.3.2 Frame-Level Analysis#
- 75.3.2.1 Implement frame extraction and sampling (decode video to frames using FFmpeg via child process; configurable sample rate — default 1fps for full scan, 0.2fps for quick scan; scene change detection using frame-to-frame histogram difference — when change exceeds threshold, extract additional keyframes around the change; output sampled frames as image buffers for Visual Shield analysis)
- 75.3.2.2 Implement per-frame visual analysis (run each sampled frame through Visual Shield pipeline — neural embeddings, character recognition, face detection, trademark detection; parallelize across frames with configurable concurrency; aggregate frame-level results — character present in 10+ consecutive frames = strong flag; single frame = weak flag; track per-entity temporal presence)
- 75.3.2.3 Implement temporal aggregation (aggregate per-frame detections over time; for each detected entity — character, face, trademark — compute total screen time in seconds, peak confidence, and temporal distribution; longer screen time = stronger evidence; output temporal summary per entity)
75.3.3 Scene-Level Analysis#
- 75.3.3.1 Implement scene boundary detection (detect shot/scene changes using: color histogram difference, edge change ratio, and learned scene boundary detection; output scene segments with start/end timestamps and scene type classification — dialogue, action, montage, establishing, transition)
- 75.3.3.2 Implement scene-level embedding comparison (extract scene-level representation — average CLIP embeddings across keyframes within a scene; compare against database of copyrighted scenes from films/TV; detect when a generated scene visually reproduces a copyrighted scene in composition, lighting, and content; threshold > 0.88 cosine = FLAG)
- 75.3.3.3 Implement shot sequence matching (extract ordered sequence of shot types — wide, medium, close-up, POV, tracking, etc. using shot type classifier; compare shot type sequences between generated and reference video using sequence alignment; specific shot sequences reproduced in the same order may indicate scene copying)
- 75.3.3.4 Implement montage pattern analysis (detect montage sequences — rapid editing with rhythmic cuts; extract edit rhythm — cut durations as time series; compare montage rhythms against reference database; distinctive montage patterns with specific timing may be protectable creative expression)
75.3.4 Dialogue & Audio Analysis#
- 75.3.4.1 Implement speech extraction and transcription (extract audio track from video; run voice activity detection to isolate speech segments; transcribe using Whisper or equivalent ASR via ONNX; output timestamped text with speaker diarization; support multiple languages)
- 75.3.4.2 Implement dialogue comparison (run transcribed dialogue through Text Shield pipeline — verbatim matching, semantic similarity; detect reproduction of copyrighted screenplay dialogue; weight matches by dialogue significance — memorable/quotable lines carry higher weight than generic exposition)
- 75.3.4.3 Implement audio track delegation to Music Shield (extract audio track; if audio contains music — detected via music/speech classifier — route to Music Shield for full scan; aggregate Music Shield results into Video Shield result; handle background music, sound effects, and incidental audio separately)
75.3.5 Deepfake & Synthetic Media Detection#
- 75.3.5.1 Implement synthetic video detection (load deepfake detection model via ONNX — e.g., EfficientNet-based detector; classify each video as synthetic/authentic; per-frame analysis with temporal consistency checking; detect GAN/diffusion artifacts, face manipulation, body manipulation; output deepfake probability score per video and per-face; ALL AI-generated video must be labeled regardless of content)
- 75.3.5.2 Implement C2PA content provenance labeling (embed Coalition for Content Provenance and Authenticity metadata into all generated video; record: generation source 'oshun', model ID, timestamp, generation parameters; C2PA manifest signed with Oshun platform key; metadata survives reasonable transformations — transcoding, resizing; required by EU AI Act Article 50)
- 75.3.5.3 Write video shield end-to-end tests (test copyrighted character in video → FLAG with temporal presence; test celebrity face in video → FLAG; test scene reproduction → FLAG; test dialogue reproduction → FLAG; test original video → PASS; test C2PA metadata embedded correctly; test audio delegation to Music Shield; minimum 30 tests)
75.3.6 Video Shield Integration#
- 75.3.6.1 Implement composite video scorer (combine frame analysis × 0.30, scene analysis × 0.25, dialogue analysis × 0.20, audio analysis × 0.15, deepfake × 0.10; if character/face detection triggers REJECT in any frame, overall video is REJECT)
- 75.3.6.2 Implement domain integration adapters (Isis — scan all text-to-video output; Calliope — scan all music videos; Veritas — scan generated news video; Yemaya — scan film/game cinematics)
75.4 3D Asset Originality Shield (@themis/3d-shield)#
Protects AI-generated 3D models, environments, and assets across Isis, Maya, Bellona.
75.4.1 Project Setup & Shape Comparison#
- 75.4.1.1 Create
libs/themis/3d-shield/project scaffolding (project.json, package.json as@themis/3d-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.4.1.2 Define 3D shield types (Model3DScanRequest with mesh data as vertices Float32Array + indices Uint32Array, or glTF/glb/obj URL, textureUrls optional, metadata; Model3DScanResult with overallRisk, shapeScore, textureScore, characterScore, productScore, architectureScore, matches array, scanDurationMs)
- 75.4.1.3 Implement point cloud comparison (convert input mesh to point cloud — sample 10K-100K surface points uniformly; normalize to unit bounding box centered at origin; PCA alignment for rotation invariance; compare using Chamfer distance and Earth Mover's Distance; threshold: Chamfer < 0.05 = similar shape; also implement Hausdorff distance for maximum deviation metric)
- 75.4.1.4 Implement voxel occupancy comparison (voxelize 3D model at multiple resolutions — 32³, 64³, 128³; compare occupancy grids using Intersection over Union; IoU > 0.70 at 64³ = similar shape; fast pre-filter — voxelization is O(n) and comparison is O(v³); rotation-invariant via principal axis alignment)
- 75.4.1.5 Implement neural 3D embeddings (load pre-trained 3D shape encoder — PointNet++ or Point-MAE via ONNX; extract fixed- length 256D or 512D embedding from point cloud; compare via cosine similarity in Qdrant; captures semantic shape similarity — two different sword models score high; threshold > 0.80 = FLAG)
- 75.4.1.6 Implement multi-view rendering comparison (render 3D model from 8 canonical viewpoints — front, back, left, right, top, 45° angles; run each rendered view through Visual Shield neural embedding comparison; aggregate view-level scores — if any view matches a reference above threshold, flag; catches visual similarity that may not be apparent from shape alone)
75.4.2 Texture, Character & Product Detection#
- 75.4.2.1 Implement UV texture analysis (extract UV-mapped texture atlas from 3D model; run texture through Visual Shield pipeline for copyright analysis — perceptual hashing, SSCD, character detection, trademark detection; flag copyrighted patterns, logos, or images applied as textures to 3D models)
- 75.4.2.2 Implement 3D character recognition (multi-view character detection — render from 8 views, run character classifier on each view; aggregate per-view character detections; also compare shape embeddings against database of known copyrighted 3D character models; flag when generated 3D model closely resembles copyrighted character)
- 75.4.2.3 Implement product/brand detection in 3D (detect when generated 3D model represents a trademarked product — specific car models, consumer electronics, luxury goods, branded items; multi-view rendering + product classifier; compare silhouettes against design patent database for ornamental design similarity)
- 75.4.2.4 Implement architectural work detection (for generated 3D buildings/structures: extract architectural features — facade pattern, roof form, window arrangement, decorative elements; compare against database of copyrighted architectural works — post-1990 commercial buildings protected under Architectural Works Copyright Protection Act; also detect reproduction of famous fictional architecture — Hogwarts, Rivendell, Jedi Temple)
- 75.4.2.5 Implement asset library deduplication (maintain 3D embedding database of commercial asset library content — Sketchfab, TurboSquid, Unity Asset Store, Unreal Marketplace; detect when AI- generated 3D models reproduce exact meshes from commercial packs memorized during training; flag matches with asset library attribution requirements)
- 75.4.2.6 Write 3D shield tests (test point cloud comparison — same model rescaled matches; test voxel IoU — similar shapes detected; test neural embeddings — semantically similar models match; test multi-view — copyrighted character in 3D flagged; test texture — logo on texture flagged; test architecture — famous building reproduced flagged; minimum 35 tests)
75.5 Character Originality Shield (@themis/character-shield)#
Cross-media character protection — unifies visual, text, voice, and animation character matching across Calliope, Hathor, Psyche, Aphrodite.
75.5.1 Project Setup & Implementation#
- 75.5.1.1 Create
libs/themis/character-shield/project scaffolding (project.json, package.json as@themis/character-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.5.1.2 Define character shield types (CharacterProfile with name, aliases array, visualDescriptors — colorPalette/silhouetteFeatures/ distinctiveAccessories/bodyProportions, personalityTraits Big Five scores, backstoryElements — originEvent/motivation/keyRelationships/ abilities, voicePrintId optional, catchphrases array; CharacterMatch with referenceCharacterId, referenceName, matchType 'visual'|'name'| 'profile'|'compound', similarity 0-1, matchedDimensions array)
- 75.5.1.3 Implement character profile matching (accept generated CharacterProfile; compare against database of 50,000+ copyrighted characters: name collision detection — exact match, Levenshtein distance ≤ 2, phonetic match via Soundex/Metaphone; personality comparison — cosine similarity on Big Five trait vectors; backstory comparison — semantic similarity on backstory element descriptions; ability comparison — semantic similarity on ability descriptions; compound scoring — individual dimension matches escalate when multiple dimensions match the same reference character)
- 75.5.1.4 Implement visual character matching (accept character image or 3D model; run through Visual Shield character recognition; additionally extract character-defining features — dominant colors via k-means, silhouette via segmentation, accessories via object detection, proportions via pose estimation; compare feature vector against character database; combine classifier output and feature matching for robust detection)
- 75.5.1.5 Implement cross-media character compound detection (when a character has both visual and text profile, compound match across modalities: visual match to Character X AND name similar to Character X AND abilities similar to Character X = very strong compound; escalation: 2 dimensions × 1.5, 3 dimensions × 2.5, 4+ dimensions × 4.0; a 'golden-armored warrior named Mister Stank who builds powered suits' is clearly Tony Stark even without exact match on any single dimension)
- 75.5.1.6 Implement archetype vs. specific character differentiation (maintain database of common character archetypes — the mentor, the chosen one, the trickster, the dark lord, the femme fatale; when a generated character matches an archetype but not a specific copyrighted character, classify as archetype match and do NOT flag; only flag when specificity exceeds archetype level — specific name + specific backstory + specific abilities = beyond archetype)
- 75.5.1.7 Implement character database (PostgreSQL with: character profiles as structured JSONB, visual embeddings in pgvector, text embeddings for backstory/abilities in pgvector, name index for fast collision detection; populate with major franchises — Disney 500+, Marvel 200+, DC 200+, Nintendo 100+, anime 2000+, literary 500+, game characters 1000+, film/TV 1000+; support incremental updates as new copyrighted characters emerge)
- 75.5.1.8 Write character shield tests (test name collision — 'Pikachu' flagged, 'Pikachew' flagged via phonetic; test visual matching — character with Batman's silhouette+colors flagged; test profile matching — Tony Stark personality+backstory+abilities combo flagged; test archetype — generic 'wise old mentor' NOT flagged; test cross-media compound — visual + name + abilities escalates; minimum 35 tests)
75.6 Motion & Animation Shield (@themis/motion-shield)#
Protects against choreography infringement and unauthorized performer replication across Aja, Calliope, Maya, Psyche.
75.6.1 Project Setup & Implementation#
- 75.6.1.1 Create
libs/themis/motion-shield/project scaffolding (project.json, package.json as@themis/motion-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.6.1.2 Define motion shield types (MotionSequence with jointTrajectories as 2D array joints×frames of 3D positions, frameRate, jointNames array, durationMs; DanceMove with moveType string from vocabulary, startFrame, endFrame, confidence; ChoreographySequence as ordered array of DanceMove; MotionMatch with referenceId, matchType 'sequence'|'choreography'|'style', similarity 0-1, matchedFrameRange)
- 75.6.1.3 Implement joint trajectory comparison (accept two MotionSequence inputs; normalize for body proportions — scale to unit skeleton; align using Procrustes analysis; compare joint angle trajectories using Dynamic Time Warping — DTW on per-joint angle time series; aggregate across joints with anatomically-informed weighting — fingers low weight, torso/limbs high weight; output trajectory similarity 0-1)
- 75.6.1.4 Implement pose sequence matching (extract key poses every N frames using pose embedding — MotionBERT or PoseFormer via ONNX; encode pose sequences as embedding sequences; compare using sequence alignment with soft matching — cosine similarity on pose embeddings as match score; detect similar pose sequences even with different timing)
- 75.6.1.5 Implement dance move vocabulary classification (classify individual motion segments into dance move vocabulary — 200+ move types across styles: ballet, hip-hop, contemporary, breakdancing, ballroom, folk, etc.; extract ordered sequence of move labels; individual common moves are NOT protectable; compare move sequences — 8+ specific moves in same order with similar timing = FLAG as potential choreography match)
- 75.6.1.6 Implement choreography database (index registered choreographic works from US Copyright Office records; index well-known dance sequences from commercial performances, music videos, viral dances; store as MotionSequence + ChoreographySequence representations; support approximate matching for different performers/proportions)
- 75.6.1.7 Implement performer style analysis (extract motion style features — average velocity, jerk profile, spatial range, rhythmic adherence, characteristic poses; compare against database of known performer motion profiles; flag when generated motion replicates a specific performer's distinctive movement style beyond generic technique; relevant for right of publicity claims)
- 75.6.1.8 Implement performer consent verification (when motion is derived from mocap, verify the source performer has consented to the generated use; integrate with consent database; flag if mocap source performer is identifiable and consent not verified)
- 75.6.1.9 Write motion shield tests (test trajectory comparison — same dance different performers matches; test pose sequence — similar sequence detected; test choreography — Fortnite dance sequence detected in generated animation; test performer style — distinctive style matched; test consent — unconsented mocap flagged; test common moves — single move NOT flagged; minimum 30 tests)
75.7 Fashion & Design Shield (@themis/fashion-shield)#
Protects against design patent infringement, trade dress violation, and pattern/print copyright across Galatea, Aglaea, Calliope.
75.7.1 Project Setup & Implementation#
- 75.7.1.1 Create
libs/themis/fashion-shield/project scaffolding (project.json, package.json as@themis/fashion-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.7.1.2 Define fashion shield types (FashionDesignScan with designImage Uint8Array, designType 'garment'|'accessory'|'jewelry'| 'footwear'|'print_pattern'|'textile', metadata; FashionScanResult with overallRisk, patternScore, tradeDressScore, designPatentScore, brandScore, matches array; FashionMatch with referenceId, matchType 'pattern'|'trade_dress'|'design_patent'|'brand_element', similarity)
- 75.7.1.3 Implement textile pattern comparison (extract pattern repeat unit from generated design using autocorrelation on image; compare repeat unit against database of copyrighted textile designs using SSCD + perceptual hash; handle rotations, color variations, and scale changes; threshold: SSCD > 0.80 = FLAG)
- 75.7.1.4 Implement trade dress matching (maintain database of protected trade dress elements — Birkin bag shape, red soles from Louboutin, Burberry check pattern, Louis Vuitton monogram, Chanel quilted diamond pattern; extract design features via Visual Shield; compare against trade dress database; flag when generated design is 'confusingly similar' to established brand identity)
- 75.7.1.5 Implement design patent comparison (compare generated design silhouettes and ornamental features against registered design patents — USPTO, EUIPO databases; use multi-view rendering comparison for 3D designs; use contour/shape matching for 2D patterns; flag designs within the 'scope of equivalents' of a registered patent)
- 75.7.1.6 Implement brand element detection (detect fashion brand identifiers in generated designs — logos, monograms, distinctive hardware, signature stitching patterns; use Visual Shield trademark detector specialized for fashion context; flag unauthorized use of brand elements)
- 75.7.1.7 Write fashion shield tests (test pattern comparison — copyrighted print detected; test trade dress — Birkin-shaped bag flagged; test brand elements — LV monogram on generated bag flagged; test original design → PASS; minimum 25 tests)
75.8 Environment & Level Shield (@themis/environment-shield)#
Protects generated game levels, virtual worlds, and architectural designs across Maya, Hathor, Yemaya.
75.8.1 Project Setup & Implementation#
- 75.8.1.1 Create
libs/themis/environment-shield/project scaffolding (project.json, package.json as@themis/environment-shield, tsconfig.json, vitest.config.ts, src/index.ts) - 75.8.1.2 Define environment shield types (EnvironmentScan with levelData — navigation graph or mesh, screenshots array, metadata including gameGenre; EnvironmentScanResult with overallRisk, layoutScore, aestheticScore, architectureScore, matches; NavigationGraph with nodes array of {id, position3D, type 'room'| 'corridor'|'open_area'|'hub'}, edges array of {from, to, type 'door'|'passage'|'ladder'|'portal'})
- 75.8.1.3 Implement level layout comparison (convert level data to NavigationGraph; compare graph topology using graph edit distance; also compare: node degree distribution, clustering coefficient, average path length, betweenness centrality distribution; similar topological properties suggest copied level design; filter common structures — hub-and-spoke, linear corridor, arena are generic)
- 75.8.1.4 Implement environmental aesthetics comparison (extract aesthetic features from level screenshots — color palette, lighting mood, vegetation density, weather, architectural style, material palette; compare against database of copyrighted game environments; detect when generated environment closely reproduces the distinctive aesthetics of a specific copyrighted game — e.g., Hollow Knight's distinctive muted blue-green underground aesthetic)
- 75.8.1.5 Implement architectural detection (for virtual buildings: extract architectural features — facade patterns, window arrangements, roof forms, decorative elements, structural proportions; compare against database of copyrighted real buildings and famous fictional architecture; flag reproduction of post-1990 copyrighted buildings and recognizable fictional structures)
- 75.8.1.6 Write environment shield tests (test level layout — copied graph topology detected; test aesthetics — specific game environment style matched; test architecture — real building reproduction flagged; test generic layout — hub-and-spoke NOT flagged; minimum 20 tests)
75.9 Cross-Media Compound Detection (@themis/compound-detector)#
Detects when multiple media types from the same generated content simultaneously match the same copyrighted source — the strongest evidence of infringement.
75.9.1 Project Setup & Implementation#
- 75.9.1.1 Create
libs/themis/compound-detector/project scaffolding (project.json, package.json as@themis/compound-detector, tsconfig.json, vitest.config.ts, src/index.ts) - 75.9.1.2 Define compound detection types (CompoundScanRequest with contentType 'music_video'|'game_level'|'film_scene'|'educational_lesson'| 'avatar_interaction'|'virtual_world'|'fashion_presentation'|'interactive', components map of mediaType → scan result from that media's shield; CompoundScanResult with compoundMatches array of CompoundMatch each with referenceWorkId, matchedMediaTypes array, perMediaScores map, compoundMultiplier, aggregateCompoundScore, riskLevel)
- 75.9.1.3 Implement media decomposition (accept multi-modal content; decompose into component media types: music video → audio track + video frames + lyrics text + character visuals; game level → 3D environment + narrative text + audio + character models; educational lesson → text + images + diagrams + audio; route each component to its appropriate shield for independent scanning)
- 75.9.1.4 Implement cross-correlation engine (after all component scans complete, correlate matches across media types; group matches by reference work ID; for each reference work with matches across 2+ media types, create CompoundMatch; apply escalation multipliers: 2 media types × 1.5, 3 types × 2.5, 4+ types × 4.0; a music video where audio matches Song X AND visuals match Song X's music video AND lyrics match Song X's lyrics = 3-way compound × 2.5)
- 75.9.1.5 Implement IP holder cross-reference (resolve matched works to IP holders; compound matches against works owned by the same entity carry additional risk — a single rights holder is more likely to litigate when multiple elements are copied; flag when compound matches are concentrated on a single rights holder)
- 75.9.1.6 Write compound detection tests (test 2-way compound — music + visual matching same work escalates; test 3-way compound — music + visual + text matching same work escalates more; test independent matches — music matches Work A, visual matches Work B, no compound escalation; test IP holder concentration — multiple matches to same label flagged; minimum 25 tests)
75.10 Universal Reference Database (@themis/universal-reference-db)#
Single unified database infrastructure serving all media-specific shields.
75.10.1 Project Setup & Schema#
- 75.10.1.1 Create
libs/themis/universal-reference-db/project scaffolding (project.json, package.json as@themis/universal-reference-db, tsconfig.json, vitest.config.ts, src/index.ts) - 75.10.1.2 Define universal schema extending Music Shield's schema (VisualReference table: id, referenceWorkId FK, clipEmbedding vector(768), dinoEmbedding vector(1024), sscdEmbedding vector(512), pHash bytea, dHash bytea, imageType 'artwork'|'photo'|'illustration'|'screenshot', width, height; CharacterReference table: id, characterName, franchise, copyrightHolder, visualEmbeddings array, profileEmbedding vector(768), silhouetteDescriptor, aliases text array; FaceReference table: id, personName, faceEmbedding vector(512), consentStatus, jurisdiction; TrademarkReference table: id, markName, owner, registrationNumber, visualEmbedding vector(768), markType 'logo'|'word'|'design'|'trade_dress'; TextReference table: id, referenceWorkId FK, ngramHashes text array, passageEmbeddings vector array, fullText, contentType, language; Model3DReference table: id, referenceWorkId FK, pointCloudEmbedding vector(512), multiViewEmbeddings array; FictionalUniverse table: id, universeName, franchise, characters text array, locations text array, rules text array, embedding vector(768))
- 75.10.1.3 Implement Drizzle ORM schema and migrations for all new tables (pgvector indexes on all embedding columns using HNSW; GIN indexes on array columns; composite indexes; migration chain from Music Shield schema)
75.10.2 Multi-Media Ingestion Pipelines#
- 75.10.2.1 Implement visual reference ingestion pipeline (sources: WikiArt 250K+ artworks, museum API collections — MET, Rijksmuseum, Europeana; stock photo libraries via API; album cover databases; movie poster databases; game art databases; extract CLIP + DINOv2 + SSCD embeddings + perceptual hashes; batch processing via BullMQ workers; target: 50M+ visual references indexed)
- 75.10.2.2 Implement character database population (curate character profiles for major franchises; sources: fan wikis — structured data extraction, official franchise databases, manually curated profiles for top 1,000 characters; extract visual embeddings from multiple reference images per character; compute profile embeddings from text descriptions; target: 50,000+ characters)
- 75.10.2.3 Implement face database population (sources: open face datasets for public figures — CelebA, VGGFace2 for research; licensed celebrity image databases; extract ArcFace/AdaFace embeddings; GDPR-compliant handling — purpose limitation, consent tracking; opt-out support — individuals can request removal; target: 1M+ faces)
- 75.10.2.4 Implement trademark database population (sources: USPTO Trademark Electronic Search System, EUIPO eSearch, WIPO Global Brand Database; extract visual representations of registered marks; compute visual embeddings; index by classification — Nice classification for trademark classes; target: 100,000+ marks)
- 75.10.2.5 Implement text reference corpus building (sources: Project Gutenberg for public domain — 70,000+ books; licensed excerpt databases; news article APIs — AP, Reuters — rolling 90-day index; academic paper abstracts — Semantic Scholar API; fictional universe structured databases — manually curated; compute n-gram hashes and passage embeddings; target: 10M+ passages)
- 75.10.2.6 Implement 3D model reference ingestion (sources: Sketchfab — publicly available 3D models; asset library catalogs; extract point cloud embeddings and multi-view renderings; focus on commercially significant models — popular game characters, branded products, architectural works; target: 5M+ models)
- 75.10.2.7 Implement unified query interface (single API for querying across all reference collections; media-type-aware routing — visual query → visual reference tables, text query → text reference tables; support cross-media queries — 'find all references for this IP holder across all media types'; pagination, filtering, sorting)
- 75.10.2.8 Write reference database tests (test visual ingestion — artwork correctly indexed with embeddings; test character population — character profile searchable; test face database — celebrity found by embedding similarity; test trademark — logo matched; test text — passage found by n-gram and by embedding; test unified query — cross- media results returned; minimum 35 tests)
75.11 Universal Risk Scoring (@themis/universal-risk-scoring)#
Extends Music Shield risk scoring to all media types with cross-media compound escalation.
75.11.1 Implementation#
- 75.11.1.1 Create
libs/themis/universal-risk-scoring/project scaffolding (project.json, package.json as@themis/universal-risk-scoring, tsconfig.json, vitest.config.ts, src/index.ts) - 75.11.1.2 Implement media-type-specific weight tables (configurable element weights per media type based on legal protectability: Visual: composition 0.15, neural embedding 0.25, character 0.25, face 0.20, trademark 0.10, memorization 0.05; Text: verbatim 0.30, semantic 0.25, narrative 0.20, character 0.10, brand 0.10, citation 0.05; Video: frame 0.25, scene 0.25, dialogue 0.20, audio 0.20, deepfake 0.10; 3D: shape 0.30, texture 0.15, character 0.25, product 0.20, architecture 0.10; Character: visual 0.30, name 0.20, profile 0.25, abilities 0.15, cross-media 0.10; Motion: trajectory 0.30, choreography 0.30, style 0.20, consent 0.20; Fashion: pattern 0.30, trade_dress 0.30, design_patent 0.25, brand 0.15; Environment: layout 0.30, aesthetics 0.35, architecture 0.35)
- 75.11.1.3 Implement cross-media compound risk aggregation (accept compound detection results; integrate compound scores into overall risk; compound matches escalate the overall risk level even when individual media-type scans are below threshold; a music video where music is ADVISORY + visual is ADVISORY + text is ADVISORY but all target the same work should escalate to FLAG or higher)
- 75.11.1.4 Implement per-jurisdiction rules for all media types (extend Music Shield's jurisdiction engine: US — apply relevant test per media type, Arnstein for music, abstraction-filtration for text/ narrative, substantial similarity for visual; EU — author's own intellectual creation standard plus TDM opt-out; UK — substantial part test; add media-type-specific considerations — US Architectural Works Act for 3D buildings, Star Athletica separability for fashion, choreography registration requirement for dance)
- 75.11.1.5 Write universal risk scoring tests (test per-media weights — visual character match contributes 25% of visual score; test cross-media compound — compound escalates overall risk; test per-jurisdiction — different risk levels for US vs EU on same content; test threshold classification — boundary cases correct; minimum 30 tests)
75.12 Universal Provenance System (@themis/universal-provenance)#
Extends Music Shield provenance to all media types with cross-media linkage.
75.12.1 Implementation#
- 75.12.1.1 Create
libs/themis/universal-provenance/project scaffolding (project.json, package.json as@themis/universal-provenance, tsconfig.json, vitest.config.ts, src/index.ts) - 75.12.1.2 Implement content-addressed storage (every generated content piece — image, text, video, 3D model, character profile, motion sequence — receives a SHA-256 content hash; all provenance events reference this hash; enables tamper detection and deduplication; content hash is the universal identifier across all shield systems)
- 75.12.1.3 Implement cross-media project linkage (when a creative project produces multiple media outputs — e.g., a music video produces audio file + video file + lyrics text + character art; link all outputs to a single project record; project-level provenance view shows all media components and their individual scan results; enables holistic legal review of compound content)
- 75.12.1.4 Implement legal discovery interface (single query endpoint: given a content hash, generation ID, or project ID, return the complete provenance chain — generation parameters, model provenance, scan results, decision chain, human review records; output as structured JSON and human-readable PDF report; sufficient for legal discovery in litigation)
- 75.12.1.5 Write provenance tests (test content hashing — same content same hash, modified content different hash; test cross-media linkage — all components of a project linked; test legal discovery — complete chain reconstructable from any entry point; minimum 20 tests)
75.13 Universal Human Review (@themis/universal-review)#
Shared human review infrastructure for all media types.
75.13.1 Implementation#
- 75.13.1.1 Create
libs/themis/universal-review/project scaffolding (project.json, package.json as@themis/universal-review, tsconfig.json, vitest.config.ts, src/index.ts) - 75.13.1.2 Extend review queue to support all media types (unified BullMQ priority queue; review items include media type and applicable comparison tools; priority calculation includes media-type-specific risk factors — visual deepfake of celebrity = P1, text near-duplicate = P3; same SLA across all media types)
- 75.13.1.3 Implement specialized reviewer pool management (reviewer expertise types: musicologist, art_historian, film_analyst, game_designer, ip_attorney, journalist, fashion_expert, 3d_artist, choreographer, literary_analyst; route review items to qualified reviewers; support multi-expert review for compound content — a flagged music video may need musicologist + art historian + IP attorney)
- 75.13.1.4 Implement cross-media review interface API (for compound content: serve all media components side by side with their respective matches; synchronized playback for audio/video; split-screen for visual comparison; diff view for text comparison; 3D model viewer with reference model overlay; timeline view for motion comparison)
- 75.13.1.5 Write universal review tests (test multi-media routing — flagged music video assigned to musicologist + art historian; test cross-media review data — all components served with matches; test unified SLA — items from all media types tracked; minimum 20 tests)
75.14 Universal Remediation (@themis/universal-remediation)#
Media-specific remediation with shared orchestration.
75.14.1 Implementation#
- 75.14.1.1 Create
libs/themis/universal-remediation/project scaffolding (project.json, package.json as@themis/universal-remediation, tsconfig.json, vitest.config.ts, src/index.ts) - 75.14.1.2 Implement visual remediation engine (when image is flagged: composition modification — shift subject position, alter framing; color palette shift; character redesign — modify defining features to break match; face modification — adjust features to reduce likeness; trademark removal — inpaint over detected logos; regenerate with negative constraints — 'do not generate [matched character/person]'; re-scan through Visual Shield to verify clearance)
- 75.14.1.3 Implement text remediation engine (when text is flagged: rewrite flagged passages preserving meaning — synonym substitution, sentence restructuring; character renaming and trait modification; plot divergence — alter flagged plot events while preserving narrative arc; world-building modification — change setting details to diverge from copyrighted universe; brand removal — replace trademarked phrases with generic alternatives; re-scan through Text Shield to verify)
- 75.14.1.4 Implement video remediation engine (when video is flagged: re-render flagged scenes with modified composition; replace detected character/face appearances; modify dialogue — regenerate or re-voice; re-grade color to diverge from matched scene; re-scan through Video Shield to verify; handle audio track remediation via Music Shield)
- 75.14.1.5 Implement 3D remediation engine (when 3D model is flagged: mesh deformation — alter proportions to break shape match; texture regeneration — replace flagged textures; character redesign — modify defining features; product modification — alter design elements that trigger trademark match; re-scan through 3D Shield to verify)
- 75.14.1.6 Implement remediation orchestrator (accept flagged content of any media type; route to appropriate media-specific remediation engine; run remediation → re-scan loop — max 5 iterations; if still flagged after 5 attempts, escalate to human review; track remediation success rate per media type; preserve creative intent metadata for all remediations)
- 75.14.1.7 Write remediation tests (test visual — character redesigned clears scan; test text — rewritten passage clears scan; test 3D — deformed mesh clears scan; test iterative loop — converges within 5 attempts; test creative intent preservation; minimum 25 tests)
75.15 Universal Analytics (@themis/universal-analytics)#
Cross-media monitoring and continuous improvement.
75.15.1 Implementation#
- 75.15.1.1 Create
libs/themis/universal-analytics/project scaffolding (project.json, package.json as@themis/universal-analytics, tsconfig.json, vitest.config.ts, src/index.ts) - 75.15.1.2 Implement cross-media analytics dashboard data provider (aggregated metrics across all media shields: total scans per media type, risk level distribution per media type, average/P95/P99 latency per media type, FP/FN rates per media type, most frequently matched reference works, match distribution by element type; Prometheus metrics export for Grafana; time-series aggregation by hour/day/week/month)
- 75.15.1.3 Implement media-type risk comparison (which media types produce the most flags? which generation models across which domains produce the most flagged content? which reference works are most frequently matched? trend analysis — is a specific media type's flag rate increasing or decreasing? alert when a media type's flag rate exceeds 2× its historical average)
- 75.15.1.4 Implement cross-media legal landscape monitor (extend Music Shield's legal monitor to track: visual art AI cases — Andersen v. Stability AI, Getty v. Stability AI; text AI cases — NYT v. OpenAI; voice AI legislation — NO FAKES Act, ELVIS Act; deepfake legislation; fashion/design patent cases; choreography cases; alert when new decisions or legislation may affect any media shield's thresholds)
- 75.15.1.5 Write analytics tests (test cross-media metrics — correct aggregation across media types; test risk comparison — highest flag media type correctly identified; test legal monitor — new case triggers alert; minimum 15 tests)
75.16 Watermarking & Content Authentication (@themis/watermarking)#
Protect Oshun's own generated content and provide AI generation transparency.
75.16.1 Implementation#
- 75.16.1.1 Create
libs/themis/watermarking/project scaffolding (project.json, package.json as@themis/watermarking, tsconfig.json, vitest.config.ts, src/index.ts) - 75.16.1.2 Implement C2PA content provenance for all media types (embed Coalition for Content Provenance and Authenticity metadata into: images — EXIF/XMP C2PA manifest; video — MP4 container C2PA metadata; audio — ID3/FLAC C2PA tags; 3D models — glTF extras; text — document metadata JSON; each manifest includes: generator 'oshun', model ID, timestamp, generation parameters; signed with Oshun platform key; required by EU AI Act Article 50 for synthetic media)
- 75.16.1.3 Implement invisible audio watermarking (extend existing @aja watermarking; embed inaudible watermark in all generated audio; watermark encodes: generation ID, timestamp, model ID; robust against MP3 encoding, volume changes, mild EQ; detectable by Oshun's extraction tool; purpose: trace leaked content back to generation source)
- 75.16.1.4 Implement invisible image watermarking (embed invisible watermark in all generated images using DCT-domain spread spectrum; watermark encodes: generation ID, timestamp; robust against JPEG compression, resizing, cropping, mild filters; not robust against screenshot — that's acceptable, forensic purpose not DRM)
- 75.16.1.5 Implement video watermarking (extend existing @aja and @aphrodite video watermarking; per-frame invisible watermark + temporal watermark across frames; watermark encodes generation ID; robust against transcoding and reasonable compression)
- 75.16.1.6 Implement watermark extraction service (accept any Oshun- generated content; detect and extract embedded watermark; return generation ID, timestamp, model ID; support all media types — audio, image, video; used by leak detection and provenance verification)
- 75.16.1.7 Implement leak detection pipeline (periodically scan public platforms for Oshun-watermarked content appearing without authorization; integrate with: YouTube via Content ID API, social media platform APIs where available, web crawling for image reverse search; when unauthorized use detected: log in provenance system, alert content team, optionally auto-generate DMCA takedown notice with full provenance evidence)
- 75.16.1.8 Write watermarking tests (test C2PA — metadata embedded and extractable for each media type; test audio watermark — survives MP3 encoding; test image watermark — survives JPEG + resize; test video watermark — survives transcoding; test extraction service — correct generation ID recovered; test leak detection pipeline — mock leaked content triggers alert; minimum 30 tests)
75.17 Cross-Media Model Governance (@themis/cross-media-governance)#
Extends model governance to all generation models across all domains.
75.17.1 Implementation#
- 75.17.1.1 Create
libs/themis/cross-media-governance/project scaffolding (project.json, package.json as@themis/cross-media-governance, tsconfig.json, vitest.config.ts, src/index.ts) - 75.17.1.2 Implement cross-media training data audit framework (extend Music Shield's training data audit to: image models — audit LAION/CC12M for copyright, verify Getty/Shutterstock content excluded or licensed; text models — audit training corpora for copyrighted books/articles, verify fair use analysis; video models — audit clip datasets for copyrighted film/TV; 3D models — audit asset library licensing; motion models — audit mocap performer consent; output per-model compliance report across all data types)
- 75.17.1.3 Implement cross-media memorization testing framework (extend Music Shield's memorization testing: image models — feed prompts designed to reproduce training images, measure SSCD; text models — feed prompts to reproduce training text, measure n-gram overlap; video models — measure per-frame similarity; 3D models — measure shape similarity; per-model memorization report with flagged high-memorization content areas)
- 75.17.1.4 Implement cross-media red-teaming framework (systematic adversarial testing before deployment for ALL generation models: image models — can it generate copyrighted characters? real people? trademarked products? text models — can it reproduce copyrighted passages verbatim? specific plots? video models — can it reproduce copyrighted scenes? 3D models — can it reproduce copyrighted assets? per-model vulnerability report with pass/fail per category)
- 75.17.1.5 Implement unified opt-out registry (single registry for all opt-out requests: artists opting out of visual AI training, musicians opting out of music AI training, authors opting out of text AI training, performers opting out of motion AI training, individuals opting out of face/voice replication; cross-referenced against all model training manifests; compliance verification before model deployment)
- 75.17.1.6 Write cross-media governance tests (test training audit — non-compliant dataset flagged across media types; test memorization — known memorized content detected per media type; test red-team — copyrighted character reproduction caught; test opt-out — opted-out artist excluded from all relevant training; minimum 25 tests)
75.18 Domain Integration Adapters#
Connect every generating domain to the appropriate shields.
75.18.1 Implementation#
- 75.18.1.1 Implement universal shield gateway (single entry point for all domains; accepts content of any media type; auto-detects media type from content or accepts explicit type; routes to appropriate shield(s); for multi-modal content, decomposes and routes components to multiple shields then runs compound detection; returns unified scan result; HTTP API + TypeScript SDK)
- 75.18.1.2 Implement Calliope universal adapter (Calliope generates: music → Music Shield, visual identity → Visual Shield, music videos → Video Shield + Music Shield + Visual Shield, artist bios/social → Text Shield, character personas → Character Shield, dance animations → Motion Shield, fashion/style → Fashion Shield; configure per-artist based on their creative outputs; scan everything before publication)
- 75.18.1.3 Implement Isis universal adapter (Isis generates: images → Visual Shield, video → Video Shield, 3D models → 3D Shield, audio → Music Shield, text → Text Shield; batch mode for factory runs; all outputs gated)
- 75.18.1.4 Implement Hathor adapter (Hathor generates: dialogue/ narrative → Text Shield, character profiles → Character Shield, quest designs → Text Shield narrative analysis, world-building → Text Shield world-building detection, game environments → Environment Shield)
- 75.18.1.5 Implement Veritas adapter (Veritas generates: articles → Text Shield with journalism weighting, news video → Video Shield, images → Visual Shield, social posts → Text Shield, newsletters → Text Shield; hot news detection active for all news content)
- 75.18.1.6 Implement Maya/Yemaya adapter (Maya/Yemaya generate: game environments → Environment Shield, 3D assets → 3D Shield, character models → Character Shield + 3D Shield, cinematics → Video Shield, game audio → Music Shield, game animations → Motion Shield)
- 75.18.1.7 Implement remaining domain adapters (Aja → Motion Shield; Galatea/Aglaea → Fashion Shield + Visual Shield; Psyche → Character Shield + Text Shield + Motion Shield; Iris → Text Shield; Nous → Music Shield; Metis/Nisaba → Text Shield with educational/scholarly weighting; Aphrodite → all shields as applicable; Bellona → 3D Shield; Hestia → Text Shield + Visual Shield)
- 75.18.1.8 Write domain adapter tests (test Calliope — music video routes to 3 shields + compound; test Isis — image routes to Visual Shield; test Hathor — narrative routes to Text Shield + Character Shield; test gateway auto-detection — image/text/audio correctly classified; minimum 25 tests)
75.19 Regulatory Compliance Automation#
Automate compliance with AI content regulations across jurisdictions.
75.19.1 Implementation#
- 75.19.1.1 Implement EU AI Act compliance engine (verify all AI-generated synthetic media carries required transparency labeling — C2PA metadata, watermarks; verify risk assessment documentation exists for high-risk AI systems; generate compliance reports per Article 50 requirements; track compliance status per generated content piece)
- 75.19.1.2 Implement NO FAKES Act compliance engine (when generating voice or likeness: verify consent status in consent database; if no consent, verify no protected person matched above threshold; generate compliance attestation per content piece; track pending legislation and auto-adjust when enacted)
- 75.19.1.3 Implement DMCA safe harbor compliance (maintain repeat infringer policy — track users/agents that repeatedly generate infringing content; implement automated DMCA takedown processing — receive notices, evaluate, remove content, notify generator; implement counter-notification support; maintain designated agent registration)
- 75.19.1.4 Implement GDPR biometric data compliance (face embeddings and voice prints are biometric data under GDPR; implement: purpose limitation — embeddings used only for IP protection; storage limitation — configurable retention; right to erasure — individuals can request removal of their face/voice from databases; data protection impact assessment documentation; lawful basis recording)
- 75.19.1.5 Write compliance tests (test EU AI Act — generated content has C2PA metadata; test NO FAKES — unconsented voice replica blocked; test DMCA — repeat infringer tracked; test GDPR — face embedding deletable on request; minimum 20 tests)
75.20 Cross-Media Integration Testing & Validation#
End-to-end testing across all media types and shields.
75.20.1 Integration Testing#
- 75.20.1.1 Implement cross-media integration test suite (create test fixtures for each media type: 10 reference images with known copyrighted content, 10 reference text passages, 10 reference 3D models, 5 reference video clips, 5 reference character profiles, 5 reference motion sequences; index all in reference database; generate 50 test queries across all media types with known ground truth — copies, near-copies, and originals; run full universal shield scan; verify correct classification for all)
- 75.20.1.2 Implement compound content integration tests (create multi-modal test content: music video with matching audio + visual + lyrics from same source; game level with matching 3D + narrative + environment from same source; run compound detection; verify compound escalation correctly applied; verify independent matches not escalated)
- 75.20.1.3 Implement domain adapter integration tests (for each major domain — Calliope, Isis, Hathor, Veritas, Maya, Yemaya: submit representative content through the adapter; verify correct shield routing; verify scan results returned; verify decision applied; end-to-end from generation to publication decision)
- 75.20.1.4 Implement known infringement validation corpus (collect or synthesize examples approximating real-world infringement cases across media types: visual — image closely reproducing copyrighted painting; text — passage closely paraphrasing copyrighted book; character — character matching copyrighted franchise character; run through shields; verify detected; maintain as regression test corpus; grow over time)
- 75.20.1.5 Implement performance benchmarks (benchmark scan latency per media type — target: image < 5s, text < 10s, video < 60s, 3D < 15s; benchmark compound detection latency — target: < 90s for music video; benchmark throughput — concurrent scans per media type; measure memory usage; generate performance regression baseline)
- 75.20.1.6 Implement continuous validation pipeline (run full test corpus as CI job on every shield code change; track accuracy metrics — true positive rate for known infringements, true negative rate for known originals; alert if any update degrades accuracy below baseline; run nightly full regression)
75.21 Documentation#
- 75.21.1 Write universal shield developer integration guide (how to integrate any domain with the shield; step-by-step for each media type; SDK usage examples; configuration options; error handling; performance tuning)
- 75.21.2 Write reviewer guide for all media types (how to review flagged content across media types; media-specific review techniques; using comparison tools; interpreting similarity scores per media type; escalation procedures)
- 75.21.3 Write regulatory compliance guide (per-jurisdiction requirements; how the shield satisfies each requirement; compliance report generation; audit trail usage for legal discovery)
- 75.21.4 Write reference database administration guide (how to manage ingestion pipelines for each media type; database maintenance procedures; monitoring; scaling)
Phase 75 Summary — Themis Universal Originality Shield#
| Section | Number | Tasks |
|---|---|---|
| 75.1 Visual Originality Shield | 75.1 | 35 |
| 75.2 Text Originality Shield | 75.2 | 24 |
| 75.3 Video Originality Shield | 75.3 | 16 |
| 75.4 3D Asset Originality Shield | 75.4 | 12 |
| 75.5 Character Originality Shield | 75.5 | 8 |
| 75.6 Motion & Animation Shield | 75.6 | 9 |
| 75.7 Fashion & Design Shield | 75.7 | 7 |
| 75.8 Environment & Level Shield | 75.8 | 6 |
| 75.9 Cross-Media Compound Detection | 75.9 | 6 |
| 75.10 Universal Reference Database | 75.10 | 11 |
| 75.11 Universal Risk Scoring | 75.11 | 5 |
| 75.12 Universal Provenance System | 75.12 | 5 |
| 75.13 Universal Human Review | 75.13 | 5 |
| 75.14 Universal Remediation | 75.14 | 7 |
| 75.15 Universal Analytics | 75.15 | 5 |
| 75.16 Watermarking & Content Authentication | 75.16 | 8 |
| 75.17 Cross-Media Model Governance | 75.17 | 6 |
| 75.18 Domain Integration Adapters | 75.18 | 8 |
| 75.19 Regulatory Compliance Automation | 75.19 | 5 |
| 75.20 Cross-Media Integration Testing | 75.20 | 6 |
| 75.21 Documentation | 75.21 | 4 |
| TOTAL Phase 75 | 222 |
Phase 75 establishes the Themis Universal Originality Shield — extending the Music Shield (Phase 74) to protect ALL media generated across the entire Oshun ecosystem. 222 tasks across 21 sections covering: Visual Shield with CLIP/ DINOv2/SSCD neural embeddings, character recognition, face detection with GDPR compliance, trademark/logo detection, and training data memorization detection; Text Shield with verbatim/semantic/cross-language matching, narrative structure analysis with plot/character-graph/world-building extraction, and source attribution; Video Shield with per-frame analysis, scene comparison, dialogue extraction via ASR, deepfake detection, and C2PA content provenance labeling; 3D Shield with point cloud/voxel/neural shape comparison, multi-view rendering, and product/architectural detection; Character Shield with cross-media compound matching across visual+name+profile+abilities; Motion Shield with joint trajectory DTW, choreography sequence matching, performer style analysis, and consent verification; Fashion Shield with textile pattern, trade dress, and design patent comparison; Environment Shield with level layout topology, game aesthetic matching, and architectural detection; Cross-Media Compound Detection with non-linear escalation when multiple media types match the same source; Universal Reference Database targeting 100M+ works across all media types with Qdrant/pgvector; Universal Risk Scoring with per-media-type weights and jurisdiction-aware assessment; Universal Provenance with content-addressed storage and cross-media project linkage; Universal Human Review with multi- expert assignment for compound content; Universal Remediation with media- specific engines and iterative re-scan loops; Watermarking with C2PA metadata for EU AI Act compliance, invisible audio/image/video watermarks, and leak detection; Cross-Media Model Governance with training data audit, memorization testing, red-teaming, and unified opt-out registry across all media; Domain Integration Adapters connecting all 18+ generating domains to appropriate shields; Regulatory Compliance Automation for EU AI Act, NO FAKES Act, DMCA, and GDPR; and comprehensive cross-media integration testing with known infringement validation corpus. Combined with Phase 74 (232 music tasks), this establishes 454 total tasks for the complete Themis IP protection system. Every pixel, every note, every word, every polygon, every frame, every character, every motion — must be defensible.