| Session | Turns | Tokens | Cost |
|---|
| Session | Status | Turns | Actions |
|---|
Session Provider/Model sets the default LLM for the session. Narrator and Character models can override it per-role.
Quantified trait vectors, dynamic psychological state, and action taxonomy for character AI.
Description
Describe what characters you want generated for this session. The AI will use the session description and location as context.
| Attribute | Description |
|---|
No character templates available.
Create character templates first using the "Characters" tab.
| # | Gender | Role | Profession |
|---|
Choose what happens to each character when this scenario is used.
The Behavioral System replaces freeform prose-only character prompts with a structured, quantified approach. Characters gain numerical trait vectors, dynamic psychological state, and classified action intent — enabling more consistent, psychologically grounded behavior across turns.
Each character is scored 0–10 on: Warmth, Assertiveness, Composure, Honesty, Curiosity, Empathy, Impulsivity, Confidence, Deference, Resilience, Manipulativeness, Protectiveness.
Traits are derived automatically by the Character Sheet Digestor when characters are generated. Warmth is anchored to Disposition and Honesty to Moral Alignment for consistency.
Valence, Arousal, Anxiety, Defensiveness, Trust, Fatigue, Focus, Cognitive Bandwidth, Motivation — all 0–10, updated after every turn.
Updates are bounded: LLM-proposed deltas are clamped to ±2.0 per turn. The engine applies deterministic action-class effects, drains cognitive bandwidth under stress, and decays 10% toward baseline each turn to prevent runaway drift.
Every character response is classified into a strategic intent: disclose, conceal, appease, confront, deflect, withdraw, comply, resist, seek reassurance, moralize, intellectualize. This classification drives deterministic state effects and enables the two-call strategy to constrain prose generation.
| Mode | LLM Calls | Description |
|---|---|---|
| 1-Call | 1 | Action + action class + state delta in one JSON response. Fastest. |
| Hybrid | 1–2 | Starts as 1-Call; auto-escalates to 2-Call when anxiety > 7.5, defensiveness > 7.0, or cognitive bandwidth < 3.0. |
| 2-Call | 2 | Call 1 selects action class + state delta. Call 2 generates prose constrained by the selected class. Maximum control. |
Prompts are assembled in three layers optimized for LLM prefix caching:
Relationships between characters are scored on Trust, Affection, Respect, Fear, and Dependence (0–10) with a stance classification (neutral, ally, rival, etc.). The Relationship Resolver scores and selects the top-N most relevant relationships per turn based on character presence, narrative mentions, and extreme values.
The system is gated behind Enable Behavioral System (default off). When disabled, the existing monolithic prompt path is used unchanged. Enabling it activates trait vectors, dynamic state tracking, action taxonomy, and the layered prompt builder for all new sessions.