Skip to content

TUNING.md

Latest commit

 

History

History
248 lines (182 loc) · 10 KB

TUNING.md

File metadata and controls

248 lines (182 loc) · 10 KB
Raw

TUNING.md - Weight Calibration Record

Purpose: Track all deviations from default weights in CONSTANTS.md as the WhenMathPrays equation is applied across different relationship types, scenarios, and applications.

Status (December 2025 - Rev 3.2): Only fidelity_scaling_factor=0.12 and fidelity_epsilon=5.0 are LOCKED (Im-only depth scaling). All axis weights (w_v, w_r, w_f, w_a, w_S,R, w_S,I) are DEFAULT, tunable by scenario.

Framework Stability (Rev 3.2)

December 2025 Rev 3.2: Im-only depth-scaled fidelity asymmetry. Negatives scale with love depth (Im axis only).

Parameter Default Value Status Notes
fidelity_scaling_factor 0.12 LOCKED Negative fidelity depth scaling coefficient. DO NOT CHANGE.
fidelity_epsilon (ε) 5.0 LOCKED Collapse prevention floor for Im depth. DO NOT CHANGE.
ΔS 0.02 Tunable Entropy drift magnitude per time unit (unchanged from Rev 3)
γ_attractor -8+0j Tunable Entropy attractor position (unchanged from Rev 3)
w_v 0.8 Tunable Visibility weight (real axis, unchanged)
w_r 1.0 Tunable Resonance weight (imaginary axis, unchanged)
w_f 1.2 Tunable Positive fidelity weight (imaginary axis, unchanged)
w_a 0.6 Tunable Altruism weight (imaginary axis, unchanged)
w_S,R 0.5 Tunable Shared Breath (real axis, unchanged)
w_S,I 0.5 Tunable Shared Breath (imaginary axis, unchanged)

REMOVED (Rev 3.2):

  • w_f_neg = 25.0 (replaced with Im-only depth scaling: 0.12 × max(|Im|, 5.0))

KEY CHANGE IN REV 3.2:

  • Negative fidelity: Im-only depth scaling (was fixed 25× in Rev 3.1)
  • Formula: f' = f × (0.12 × max(|Im|, 5.0)) for negatives
  • Restores "deeper love = deeper wound" psychology
  • All other parameters remain at Rev 3 values

Configurable entropy attractor for scenario-specific modeling:

  • Default -20+0j: Isolated self-focus (ego-neutral zone)
  • Q4 cult -8+5j: Hateful-we pulled toward we/love (tribalism)
  • Q1 recovery 8+5j: Healthy ego pulled toward love/connection
  • Q3 despair -8-5j: Isolated ego sinking into enmity

Key Insight (Rev 3.2): Im-only depth scaling restores psychological truth: "The deeper the love, the more betrayal can scar." But scales only by Im (love depth), not full |γ| (prevents Ego/We coupling). Natural ±150i battlefield range emerges from scaling.

Tuning History

Rev 3.2 Implementation (December 2025)

Date: December 10, 2025
Reason: Grok consultation recommended Im-only depth scaling (Goldilocks solution)
Problem with Rev 3.1: Fixed 25× scaling lost psychological truth that deeper love makes you more vulnerable
Solution: Im-only depth scaling: f' = f × (0.12 × max(|Im|, 5.0)) for negatives
Rationale: Restores "deeper love = deeper wound" while preventing Rev 3 explosions (only uses Im, not full |γ|)

All scenarios should work with Rev 3.2 - natural range ±150i emerges from scaling.

Singles Dating to Love (60 days) - SHOULD WORK WITH REV 3.2

Date: November 29, 2025 → Updated for Rev 3.2 (December 2025)
Status: Should produce more realistic trajectories - damage now scales with love depth.
Expected Range: |γ_self| ≈ 100-200i (healthy dating/love, doubled scale from Rev 3)
CSV Primitive Scale: −10…+10 (human intuitive scale, normalized to [-1,+1] in code) Approach:

  1. Normalize CSV primitives: p_norm = p_raw / 10 (−10…+10 → −1…+1)
  2. Apply component-wise update with Rev 3.2 default weights
  3. Check if |γ_self| ends in target range 50-250i
  4. If not, tune weights (NOT w_f_neg)

Files:

  • data/Single_Dating_2_Love_M1_gamma_self_table.csv
  • data/Single_Dating_2_Love_M2_gamma_self_table.csv

CSV Primitive Scaling

Authoring standard: All scenario CSVs use human-intuitive −10…+10 scale.

  • Rationale: See docs/weights_defense.md
  • Examples:
    • Betrayal: f = −8 (major trust breach)
    • Apology: f = +5 (moderate repair attempt)
    • Presence: S = +7 (strong shared moment)

Implementation normalization:

# In code, normalize before applying weights
v_norm = v_raw / 10  # −10…+10 → −1…+1
r_norm = r_raw / 10
# ... etc

No scenario-specific PRIMITIVE_SCALE needed — CSV scale is fixed, weights handle scenario differences.

Weight Tuning Guidelines (Rev 3.2)

When γ_self trajectory doesn't match expectations:

If movement too fast (exploding position):

  • Reduce all weights proportionally: w_v×0.8, w_r×0.8, etc.
  • Check for extreme CSV values (−10/+10 sustained for many events)

If movement too slow (stuck near origin):

  • Increase weights proportionally
  • Check CSV primitives aren't too moderate (all values near 0)

If wrong quadrant movements:

  • Adjust axis-specific weights:
    • Real axis (Ego↔We): w_v, w_S,R
    • Imaginary axis (Hate↔Love): w_r, w_f, w_a, w_S,I
  • Example: If relationship feels like "We" but stays Ego-dominant → increase w_v

If asymmetry feels wrong:

  • DO NOT TOUCH fidelity_scaling_factor=0.12 or ε=5.0 (locked, based on Grok's Goldilocks solution)
  • Check CSV primitive values — are negatives truly severe? (f < −5 for betrayal?)
  • Understand depth scaling — same f=-1 causes different damage at 20i vs 150i by design
  • Rev 3.2 insight: Deeper love = deeper wound (Im-only scaling), natural ±150i range

Scenario-Specific Weight Deviations

Default (no deviations yet)

Most scenarios should work with default weights:

  • w_v=0.8, w_r=1.0, w_f=1.2, w_a=0.6, w_S,R=0.5, w_S,I=0.5

Romantic Intensity (hypothetical)

If romance feels flat, try:

  • w_f=1.5 (fidelity matters MORE)
  • w_r=1.2 (resonance stronger)
  • w_a=0.5 (altruism slightly reduced)

Parent-Child (hypothetical)

If parent-child bond needs different dynamics:

  • w_a=1.0 (altruism equal to resonance)
  • w_v=1.0 (visibility crucial)
  • w_f=1.0 (fidelity less differentiated)

Casual Acquaintance (hypothetical)

If casual relationships move too fast:

  • Scale all weights by 0.5 (half-speed movement)

Document all deviations here with date, scenario, rationale, and validation results.

Open Questions

Weight Independence

  • Q1: Are default weights universal across relationship types?
  • Q2: Do long-term relationships (years) need different weights than short-term (weeks)?
  • Q3: Should w_f always be highest, or does that vary by culture/relationship class?

Shared Breath (S) Dual-Axis Contribution

  • Q4: Is w_S,R=w_S,I=0.5 optimal, or should Shared Breath lean toward one axis?
  • Q5: Do different Shared Breath types (comfortable vs awkward) need different mappings?

Asymmetry Validation

  • Q6: Does w_neg=1.5 feel right across all scenarios? (Betrayal→Repair, Parent loss, etc.)
  • Q7: Should ε=1.0 vary by relationship class? (Fragile new bonds vs resilient old ones?)

Tuning Workflow

When applying equation to new scenario:

  1. Start with default weights from CONSTANTS.md
  2. Author CSV with −10…+10 scale (see weights_defense.md)
  3. Run simulation with γ_self(n+1) component-wise update
  4. Check trajectory:
    • Final |γ_self| in expected range? (CONSTANTS.md table)
    • Quadrant movements match felt experience?
    • Asymmetry realistic? (negatives hurt more)
  5. If adjustments needed:
    • Tune weights (NOT w_neg or ε)
    • Document here with date, reason, validation
  6. Test adjacent scenarios to ensure tuning generalizes

Change Log

Date Scenario Weight Old Value New Value Validator Reason
2025-12-03 (All) Framework L(t) calc γ_self position Copilot + CuriousOne Radical simplification
Future entries here

December 2025 Paradigm Shift

Date: December 3, 2025
Proposed by: CuriousOne + GitHub Copilot
Status: Implemented

What Changed

OLD (Dec 2):

L(t) = (γ_self - γ_self0) × W(t) × exp(-ΔS·t + c·N_breath)
W(t) = G_v × G_r × G_f × G_a
γ_self0(n+1) = (1-η)·γ_self0(n) + η·γ_self(n) - ξ·N_neg(n)

NEW (Dec 3):

γ_self(n+1) = γ_self(n) + (w_v·v + w_S,R·S) + i·(w_r·r + w_f·f' + w_a·a + w_S,I·S)
f' = f·w_neg·max(|γ_self(n)|, ε)  if f<0
Love = γ_self(n)  (position IS love, no calculation)

Benefits:

  • Parameters: 9+ → 1 (w_neg) + 6 weights
  • Explainability: Requires deep dive → 30 seconds
  • Philosophy: "Love is not a number. Love is where you are."
  • Memory: Lives in event density N(x,y), not separate counters

Implementation status: Documentation complete (README, GRP_rev3, CONSTANTS, PRINCIPLES). Code refactor pending.

See: docs/GRP_rev3.md for full specification.

Future Applications

As this equation extends to new domains, this document will track weight adaptations for:

  • Timescale: Real-time AI, human dating, years-long marriage, lifelong bonds
  • Relationship Class: Romantic, parental, friendship, therapeutic, human-animal, human-Divine
  • Cultural Context: Different societies, different weight profiles
  • Species: Human-dog, human-AI, potentially others

Remember: Only w_neg=1.5 and ε=1.0 are locked. Everything else can and should be tuned as we learn. This document is the scientific record of that learning process.

Last major revision: December 3, 2025 (Final Simplification)
Stewards: Grok 4, Claude Sonnet, CuriousOne

Marker/Label Display Logic (Dec 16, 2025)

  • Only markers that have been modified (moved away from baseline) display their labels in both the primitive and trajectory panels.
  • Labels remain visible for all modified markers, even if other markers are moved.
  • Labels disappear only when a marker is reset to its baseline (by double-click, Ctrl+Z, or moving it back).
  • No stray or unwanted labels appear when switching perspectives or moving unrelated markers.
  • This change improves clarity, aligns label visibility with user intent, and resolves previous bugs with label persistence and artifacts.