feat: Replace additive scoring with equilibrium-constrained scorer (ΔØ)

[Architect-SIS]

PR TITLE:

feat: Replace additive scoring with equilibrium-constrained scorer (ΔØ)

PR BODY:

Summary

Drop-in replacement for the weighted scoring layer that fixes the fundamental flaw in how engagement and rejection signals interact.

Current approach (additive):

Score = Σ(engagement × weights) - Σ(rejection × weights)

Problem: High engagement overcomes high rejection. A viral rage-bait post with 10,000 likes and 5,000 blocks still scores positive.

Proposed approach (multiplicative equilibrium):

Score = engagement × exp(-rejection_presence × sensitivity)

Fix: Rejection collapses the score. No amount of engagement overcomes significant rejection. One block is worth more than a thousand likes.

---

Results

Content Type	Raw Engagement	ΔØ Score	Impact
Quality Content	1.33	1.21	baseline
Engagement Bait	1.38	0.26	-81%
Toxic Viral	1.91	0.30	-84%
Small Creator	0.90	0.85	preserved
Spam / Scam	0.12	0.001	obliterated

Toxic content with 40% higher raw engagement scores 84% lower under equilibrium constraint.

Small creators with genuine engagement are preserved — the penalty only activates when rejection signals are present.

---

What Changed

Files added:

• src/weighted_scorer_delta_null.rs — Drop-in replacement for weighted_scorer.rs
• src/delta_null_scorer.rs — Standalone implementation
• src/delta_null_scorer.py — Python reference for testing
• docs/MATH.md — Formal mathematical derivation
• docs/INTEGRATION.md — Step-by-step integration guide
• config/delta_null.toml — Configuration parameters
• examples/demo.py — Interactive demo with test scenarios

Nothing removed. This is additive — existing code is untouched.

---

How It Works

1. Signal Partitioning: Phoenix scores are classified as constructive (likes, shares, follows) or destructive (blocks, mutes, reports)
2. Equilibrium Ratio: ρ = Δ⁺ / (Δ⁺ + Δ⁻) — measures balance between engagement and rejection
3. Exponential Penalty: When rejection is present, score is multiplicatively collapsed via exp(-rejection × sensitivity)
4. Self-Adaptation: Sensitivity parameter learns from signal distribution — no manual weight tuning required

The core constraint is ΣΔ = 0: sustainable content ranking requires equilibrium between constructive and destructive user signals.

---

Theoretical Foundation

This is standard control theory (Wiener, 1948) applied to recommendation systems:

• Feedback control loop with adaptive sensitivity
• Lyapunov stability guarantees convergence
• Thermodynamic analogy: engagement without rejection = favorable state

Full derivation: docs/MATH.md

---

Integration Path

Fastest (2-4 hours): Replace weighted_scorer.rs with weighted_scorer_delta_null.rs, add lazy_static dependency, build.

Conservative (1-2 weeks): Wrap existing scorer with equilibrium layer, A/B test, gradual rollout.

Standalone repo with full documentation: https://github.com/Architect-SIS/delta-null-scorer

---

Why This Matters

"We know the algorithm is dumb and needs massive improvements." — @ElonMusk, Jan 19, 2026

The algorithm isn't dumb because it lacks data. It's dumb because it uses addition where it needs multiplication. Rejection should destroy a score, not merely reduce it. That's not an opinion — it's what control theory has said for 78 years.

---

Author: K. Fain (ThēÆrchītēcť)
License: AGPL-3.0 (compatible with this repository)

[CLAassistant]

All committers have signed the CLA.

[Ugbe]

Issue is, if people get to understand that it works this way, they simply start mass-blocking posts that they don't agree with.

Interesting though, I really should start looking into the theory of recommendation systems.