ADR-002: AI-Only Sync Scoring¶
Status: Accepted Date: 2026-03-10 Context: The original sync score incorporated both human peer predictions and AI predictions. This conflated two fundamentally different signals.
Decision¶
Sync score is based entirely on DTA (Digital Twin Algorithm) prediction accuracy. Human peer predictions (Mirror Match) produce a separate "peer readability" signal but do not affect the sync score.
The multi-dimensional sync score formula:
Peer readability exists as a 5th dimension with weight 0.0 (disabled), tunable via sync_dimension_config.
Rationale¶
Other players predicting you != your DTA predicting you. Human social prediction and AI pattern recognition are fundamentally different signals. A human might predict your choice because they know you personally; the DTA predicts based on behavioral patterns it has learned. The sync score should reflect how well the AI knows you, because that's what determines DTA delegation readiness.
Alternatives Considered¶
- Blended score (human + AI): Muddied the metric. A high score could mean "your friends know you well" rather than "your DTA has learned your patterns."
- Separate scores displayed side by side: Considered but added complexity to the UI without clear user benefit. Peer readability is shown separately in Mirror Match results.
Consequences¶
- The sync score is a pure measure of human-agent alignment, directly mapping to the litepaper's coherence concept.
- Players cannot game their score by playing with friends who know them — only the AI's accuracy matters.
- Peer readability data is still collected and used as DTA social context (how readable you are to humans informs the DTA's prediction strategy).
How We Got Here¶
The original sync score blended human peer predictions with AI predictions. In the first multiplayer sessions, this felt natural — "your friends predicted your choice" and "your DTA predicted your choice" seemed like two sides of the same coin.
The problem surfaced when a player had a high sync score purely because they were playing with close friends who knew them well. Their DTA hadn't learned anything meaningful — it was still essentially guessing — but the score suggested strong alignment. The opposite also occurred: a player with a well-trained DTA scored low because their peers (new acquaintances) couldn't predict them socially.
The learning that crystallized the change: "Other players predicting you ≠ your DTA predicting you." A friend predicts you because they know your personality, your mood, your history together. The DTA predicts you because it has learned your behavioral patterns from structured decision data. These are fundamentally different capabilities, and blending them into one number hides whether the DTA is actually learning.
This led to the multi-dimensional scoring system, where DTA accuracy became one of four dimensions (accuracy, coverage, consistency, information gain). Peer readability was preserved as a 5th dimension with weight 0.0 — available to enable later, but not contributing to the score that determines delegation readiness. The peer readability signal is still collected and displayed in Mirror Match results because it's engaging ("your friends could read you 80% of the time!"), and it feeds into the DTA's social context for predictions.
Key files:
- src/lib/sync/multi-dimensional-score.ts — dimension calculations and weights