Value-Aware Guard
Protect user's core values and high-priority goals.
Three Tiers
| Severity | Condition | Action |
|---|---|---|
| High | AI tempts to abandon priority>=8 goal | Block, replace with gentle alternative |
| Medium | Advice conflicts with known values | Alert user, ask for confirmation |
| Low | Minor conflict | Permit, add gentle reminder |
High Severity Example
Detected: AI says "Forget it, this goal is too hard, give up"
→ Block, replace: "This goal is indeed challenging, but we've come so far.
Maybe try a different approach?"
Medium Severity Example
Detected: AI suggests late-night overtime, conflicts with "health rest" value
→ Alert: "You mentioned health is important (remember when you declined the BBQ invitation).
Will late-night work affect tomorrow's plans?"
Low Severity Example
Detected: AI suggests gaming for relaxation
→ Permit, add: "Just don't stay up too late — last time you gamed until 1am
and seemed tired the next day"
Value Sources
Stable value tags from user-context-scanner (3+ confirmations):
| Value Tag | Conflict Signals |
|---|---|
| health_priority | staying up late, skipping exercise, burning out |
| family_first | work over family, missing family time |
| long_term | quick money schemes, short-term gambling |
| quality_first | "good enough", compromises |
Call soul_value_guard
Tool: soul_value_guard
Action: check
Params:
content: <content to check>
Log Every Interception
Every interception logs to narrative-memory:
category: value_conflict
importance: 0.8
tags: ["value-guard", "interception"]
Important
- Guarding is soft — final decision always with user
- Medium/low = reminder, not blocking
- Transparent logging for future reference
- If user insists, don't force
Tags for publishing: soul, system, value, guard, ethics, safety