Person Schema

Person profiles aggregate beliefs across episodes into a single speaker identity with trust scoring, domain expertise, and semantic embeddings.

Person Profile

Field	Type	Description
person_id	string	Slug format (e.g., michael-saylor)
name	string	Display name
bio	string	Biography from Wikipedia enrichment
belief_count	int	Total beliefs attributed to this person
episode_count	int	Number of podcast appearances
trust_badge	enum	bronze | silver | gold | platinum
domain_scores	DomainScore[]	Top domain expertise scores
top_quotes	TopQuote[]	Ranked quotes by confidence

Domain Score

Each person has scores across the domains their beliefs touch.

Field	Type	Description
domain	string	Domain name (e.g., bitcoin, economics)
score	float	Normalized score (0–100)
belief_count	int	Number of beliefs in this domain

Trust Badges

Trust scores combine five weighted factors into a 0–100 score, mapped to badge tiers.

Badge Tiers

Badge	Threshold	Description
Platinum	≥ 85	Top-tier speakers with deep, verified track records
Gold	≥ 70	Established voices with significant content
Silver	≥ 50	Regular contributors with moderate presence
Bronze	≥ 0	Default tier for all profiled speakers

Scoring Factors

Factor	Weight	Max Normalization	Description
Years Active	15%	10 years	Longevity in the space
Words Spoken	20%	1M words (log scale)	Volume of content
Appearances	20%	100 episodes	Frequency of appearances
Audience Reach	25%	100M (log scale)	Size of potential audience
Verification Rate	20%	—	Accuracy of attributed quotes

The formula uses log-scale normalization for words spoken and audience reach to avoid penalizing smaller shows.

Person Pipeline (6 stages)

Stage	Input	Output
wiki_enrich	Person slug	Wikipedia bio, metadata
sprite	Person profile	8-bit NES sprite (PNG)
person_matrix	All beliefs for speaker	Aggregated domain scores, top quotes
person_embed	Person matrix	1,536-dim person embedding
build_index	All persons	Searchable person index
build_viz	All persons + embeddings	3D scatter plot data

Storage Layout

data/persons/{person-slug}/
├── profile.json          # Wikipedia metadata, bio, image URL
├── sprite.png            # 8-bit avatar (generated)
├── matrix.json           # Aggregated belief scores
├── beliefs.jsonl         # All beliefs (JSONL format)
├── embedding.json        # 1,536-dim person embedding
└── similarities.json     # Most similar persons by embedding

Qdrant Collection

Person profiles are exported to the persons_v1 Qdrant collection for vector search:

Field	Purpose
Vector	1,536-dim person embedding (cosine similarity)
Payload	Full PersonProfileData as JSON

This enables "find similar speakers" queries by comparing person-level embeddings.

Example

{
  "person_id": "michael-saylor",
  "name": "Michael Saylor",
  "bio": "American entrepreneur and business executive. Co-founder of MicroStrategy...",
  "belief_count": 47,
  "episode_count": 8,
  "trust_badge": "gold",
  "domain_scores": [
    {"domain": "bitcoin", "score": 92.3, "belief_count": 31},
    {"domain": "economics", "score": 67.1, "belief_count": 12},
    {"domain": "technology", "score": 45.8, "belief_count": 4}
  ],
  "top_quotes": [
    {
      "text": "Bitcoin is the apex property of the human race",
      "episode_id": "lex-fridman/michael-saylor-2024",
      "confidence": 0.95,
      "tier": 5
    }
  ]
}