Digital Twin

A system for building a conversational digital clone through iterative interviews. Everything learned persists in a local SQLite database and is exported as structured JSON model files. No external dependencies — just Python 3 and SQLite.

How It Works

The system operates in three modes:

1. Interview — Claude learns about you through conversation, logging observations as structured candidates that get validated and stored at session end
2. Calibration — Test the model's accuracy by generating "what would you say?" responses and providing corrective feedback
3. Impersonation — Active simulation of you in conversation, with safety controls and opt-in requirements

Each session bootstraps from the database. Context resets every session — the DB is the only source of truth.

The 7 Dimensions

Personality is modeled across seven dimensions, each with multiple facets:

Dimension	What it captures
communication_style	Sentence structure, punctuation, hedging, formality, greetings
vocabulary_language	Characteristic words, jargon, metaphors, avoidances
humor_wit	Type, timing, references, what lands vs. falls flat
values_opinions	Core values, stances, hills to die on, flexibility
knowledge_expertise	Professional domains, hobbies, cultural knowledge
emotional_relational	Expression style, stress responses, empathy style
cognitive_decision_making	Reasoning style, risk tolerance, how they change their mind

Each observation has a computed confidence score: sigmoid(evidence) × recency × consistency × depth.

Architecture

scripts/
  init_db.py          # Idempotent schema setup (12 tables, WAL mode, FK constraints)
  log_message.py      # Per-message logging with UUID dedup and fallback transcript
  session_end.py      # Token+heartbeat claim protocol for safe concurrent finalization
  export_model.py     # Deterministic JSON export with confidence scoring
  test_suite.py       # 25 automated tests

model/                # Generated JSON (personality, confidence, gaps)
db/                   # SQLite database (gitignored)
transcripts/          # Fallback transcripts and session summaries (gitignored)

.claude/
  rules/
    logging.md        # Logging protocol, directives, candidate schema
    interviewing.md   # Interview strategy by readiness level
    impersonation.md  # Safety rules, hard blocks, disclosure requirements
  settings.json       # SessionStart hook for banner

No external dependencies. No pip installs. Just python3 and the standard library.

Quick Start

# Initialize the database
python3 scripts/init_db.py

# Generate empty model files
python3 scripts/export_model.py

# Run the test suite
python3 scripts/test_suite.py

# Start a Claude Code session — CLAUDE.md handles the rest

Directives

Prefix messages with # directives to modify behavior:

Directive	Description
#meta	Off-record — message not logged, no observations captured
#dev	Dev mode for entire session — must be first message, suppresses all logging
#commands	Show available directives

Session Lifecycle

Sessions use a token+heartbeat claim protocol for safe finalization:

• Each session gets a unique ID (YYYY-MM-DD-NNN)
• Messages are logged to both DB and a length-prefixed fallback transcript
• On session end, a UUID token claims the session atomically
• Heartbeats prove liveness during processing — stale claims get reclaimed
• Orphaned sessions from crashed windows are cleaned up on next startup

Safety Model

• Impersonation requires explicit opt-in each session (resets automatically)
• Disclosure mode prepends [AI SIMULATION — not the real Gerrit] to impersonation responses
• Hard blocks: no financial, legal, or reputational harm; no unsupervised deception; no credential use
• Contradictions are data — the system models nuance rather than resolving conflicts
• Candidates, not direct writes — observations go through a validation pipeline before becoming permanent

License

MIT