relational-filter

name: relational-filter description: "Evaluates messages and actions against 23 empirically-grounded relational norms from Miller's Intimate Relationships (9th ed). Returns relational assessments for communication filtering, conflict pattern analysis, and trust erosion detection." trit: -1

Relational Filter

"Is this something one expects someone who cares about them to do?"

Version: 1.0.0 Trit: -1 (Validator — checks actions against relational norms) Bundle: verify Source: Miller, R.S. Intimate Relationships, 9th ed. (McGraw-Hill, 2021)

Overview

Relational Filter evaluates messages, actions, and behavioral patterns against 23 empirically-grounded relational norms distilled from relationship science. It takes an action + social context and returns a relational assessment: does this action align with how someone who genuinely cares would behave?

This is NOT a morality judge. It's a pattern matcher against decades of research on what sustains vs. destroys intimate bonds.

When to Use This Skill

• Evaluating whether a message draft sounds caring or inadvertently harmful
• Checking if a request to/from someone reflects healthy relational dynamics
• Filtering communication through Beeper/messages before sending
• Assessing patterns in conversation history (e.g., "is this person stonewalling?")
• Self-check: "am I about to do something that erodes trust?"
• Analyzing conflict patterns between people in project context

Database

/Users/alice/worlds/intimate_relationships.duckdb

Tables

Norm Domains (7)

How to Apply the Filter

Step 1: Identify the action and context

ACTION: What is being done/said?
CONTEXT: What is the relationship? What history exists? What was the bid?

Step 2: Query relevant norms

-- Find applicable norms for a given situation
SELECT domain, norm, violation, severity, valence
FROM relational_norms
WHERE severity = 'core'  -- start with core norms
ORDER BY domain;

For specific domains:

-- If the situation involves conflict
SELECT norm, violation FROM relational_norms 
WHERE domain IN ('respect', 'repair') AND severity = 'core';

-- If the situation involves trust/honesty
SELECT norm, violation FROM relational_norms
WHERE domain = 'trust';

-- If the situation involves someone being distant/unresponsive
SELECT norm, violation FROM relational_norms
WHERE domain IN ('responsiveness', 'care', 'commitment');

Step 3: Assess alignment

For each relevant norm, evaluate:

Step 4: Generate assessment

Report format:

RELATIONAL FILTER: [action summary]
Context: [relationship + history]

Norms checked:
  [+1] responsiveness: Attending to partner's bid ✓
  [-1] respect: Criticism of character, not behavior ✗
  [ 0] trust: Insufficient information

Assessment: [ALIGNED / MIXED / MISALIGNED]
Core violations: [list any severity='core' norms scored -1]
Note: [brief explanation grounded in Miller's research]

Quick-Reference: The 15 Core Norms

These are the relationship-defining ones. Chronic violation of ANY predicts relationship failure.

DO (things someone who cares does)

1. Respond to bids — when someone reaches out, turn toward them
2. Celebrate wins — enthusiastic response to partner's good news
3. Be honest — small lies erode trust disproportionately
4. Follow through — reliability IS caring
5. Complain about behavior, not character — "I feel X when Y" not "you always"
6. Accept influence — let your partner change your mind sometimes
7. Think "us" — cognitive interdependence, not scorekeeping
8. Soft startup — begin hard conversations gently
9. 5:1 ratio — five positive interactions per negative one
10. Satisfy the need to belong — show up regularly with warmth

DON'T (things that signal you don't care, even if you think you do)

11. Never express contempt — #1 predictor of relationship death
12. Don't stonewall — withdrawing from engagement is devastating
13. Don't betray then minimize — "it wasn't a big deal" IS the second betrayal
14. Don't use power coercively — unilateral control is abuse
15. Never use violence — no exceptions, no excuses

Integration with Beeper MCP

When filtering messages before sending or evaluating received messages:

# Pseudocode for message filter
def relational_check(message, relationship_context):
    """
    Run before sending a message or when evaluating received behavior.
    """
    norms = query_relevant_norms(message, relationship_context)
    
    scores = []
    for norm in norms:
        score = assess_alignment(message, norm)
        scores.append((norm.domain, score, norm))
    
    core_violations = [s for s in scores if s[1] == -1 and s[2].severity == 'core']
    
    if core_violations:
        return f"⚠️ PAUSE — core norm violation: {core_violations[0][2].norm}"
    
    avg = sum(s[1] for s in scores) / len(scores) if scores else 0
    if avg > 0.3:
        return "✓ aligned"
    elif avg > -0.3:
        return "~ mixed signals"
    else:
        return "✗ misaligned — reconsider"

Integration with DuckDB-IES

The relational filter can be applied to interaction history:

-- Assess communication patterns over time
WITH message_norms AS (
    SELECT 
        m.timestamp,
        m.content,
        n.domain,
        n.norm,
        CASE 
            WHEN m.content ILIKE '%you always%' OR m.content ILIKE '%you never%' 
                THEN -1  -- criticism pattern
            WHEN m.content ILIKE '%I feel%' OR m.content ILIKE '%I need%'
                THEN 1   -- healthy complaint pattern
            ELSE 0
        END as alignment
    FROM messages m
    CROSS JOIN relational_norms n
    WHERE n.domain = 'respect'
)
SELECT 
    DATE_TRUNC('week', timestamp) as week,
    AVG(alignment) as avg_respect_alignment,
    COUNT(*) FILTER (WHERE alignment = -1) as violations
FROM message_norms
GROUP BY 1
ORDER BY 1;

CatColab Connection

This skill maps to ThSignedCategory in the intimate-relationships-catcolab distillation:

• Each norm is a morphism with a sign (+1 do, -1 don't)
• Composition reveals cascades (one contempt → defensive → stonewall)
• The 5:1 ratio is the sign-balance threshold

GF(3) Triads

relational-filter (-1) ⊗ beeper-poly (0) ⊗ artifacts-builder (+1) = 0 ✓
intent-sink (-1) ⊗ relational-filter (-1) ⊗ scientific-visualization (+1) = needs +1 partner
relational-filter (-1) ⊗ statistical-analysis (0) ⊗ duckdb-ies (+1) = 0 ✓

Examples

Example 1: Checking a message before sending

Action: "You said you'd handle the cables three days ago and nothing happened." Context: barton, co-founder, physically manages GX10 hardware, often unresponsive

Filter:

• respect: This is a complaint about behavior (good), not character attack ✓
• trust: References broken commitment — legitimate concern ✓
• responsiveness: Could add a bid — "can you update me on timing?" would be better
• repair: Startup is moderately hard — softening with "hey, checking in on..." would improve

Assessment: MIXED → suggest softened version: "Hey, checking in — any update on the cables? Let me know if something's blocking you."

Example 2: Evaluating received behavior

Action: Partner hasn't responded to 3 messages over 2 days Context: Usually responsive, no stated conflict

Filter:

• responsiveness: Failing to turn toward bids [-1 if pattern, 0 if one-off]
• care: Not satisfying need to belong through regular contact [-1]
• boundary: Could be reactive jealousy on your end if you're catastrophizing [check]

Assessment: Before escalating, check — are they overwhelmed? (invisible support norm). One message: "Hey, haven't heard from you — everything ok?" respects their autonomy while expressing care.

Skill Name: relational-filter
Type: Message/Action Validator
Trit: -1 (MINUS — VALIDATOR)
GF(3): Checks relational alignment before execution
Database: /Users/alice/worlds/intimate_relationships.duckdb
Source: Miller 9th ed., Gottman, Rusbult, Baumeister & Leary