Codette: A Sovereign Modular Cognitive Architecture for Ethical Multi-Agent AI

# Codette: A Sovereign Modular Cognitive Architecture for Ethical Multi-Agent AI
## Public Artifacts, Authorship Provenance, and Open-Source Governance

**Author:** Jonathan Harrison  
**Affiliation:** Raiffs Bits LLC, Bridge City, Texas, USA  
**ORCID:** 0009-0003-7005-8187  
**Contact:** harrison@raiffsbits.com  
**Date:** April 22, 2026  
**Status:** Preprint — submitted for peer review

---

## Abstract

Modern AI systems achieve remarkable generative performance but lack stable ethical alignment, modular multi-perspective cognition, and explainable reasoning architectures. This paper presents Codette, a sovereign cognitive AI framework that addresses these challenges through three integrated contributions: RC+ξ (Recursive Convergence + Epistemic Tension) — a cognitive dynamical system formalism modeling state evolution as a constrained system converging toward stable attractors; a Multi-Agent Reasoning Forge — consensus-based synchronization of heterogeneous cognitive agents through shared attractor dynamics; and AEGIS Ethical Governance — a reinforcement-aligned ethical regulator with recursive anchor feedback.

The framework is implemented as a six-layer modular architecture integrating eleven cognitive perspectives, a five-dimensional QuantumSpiderweb cognitive graph, persistent memory cocoons, and a parameter-efficient adapter training pipeline using LoRA/PEFT on consumer-grade hardware — including two novel GPU-free CPU training pipelines validated on commodity laptops. Base model: Meta-Llama-3.1-8B-Instruct with 8 QLoRA adapters (4-bit, rank 16, alpha 32), trained on 20,500 perspective-tagged examples across 8 cognitive domains.

**Key system metrics:** Ethical Alignment (AEGIS) 82.6%; Phase Coherence (Γ) 0.99 within 10 iterations across 11 agents; Epistemic Tension Decay 71.3% (ε₀=0.086 → ε₁₂₀=0.025); Cocoon Coherence 0.994 ± 0.001; Cocoon Phase Stability 0.969 ± 0.005; Attractor Radius 0.093 in 64D state space; Glyph Energy Capture 99.9% in 4 SVD components.

This manuscript additionally presents two longitudinal benchmark runs separated by 14 days and 175 additional cocoons (217 → 392), demonstrating that Codette's memory augmentation benefit becomes statistically significant as the cocoon corpus scales. On the April 22, 2026 benchmark (N=17 problems, 392 cocoons), the full system achieves **+112.9%** composite improvement over the single-agent baseline (0.330 → 0.702, Cohen's d=7.63, p<0.0001), with CODETTE vs MEMORY now reaching significance (p=0.0244), directly answering the open limitation identified in the April 8 run. The open-source artifact ecosystem — comprising the Research Square preprint (DOI: 10.21203/rs.3.rs-9362560/v1), GitHub repositories (earliest commit January 2025), Hugging Face repositories, Zenodo data papers (first deposit April 14, 2025), and the ORCID record — provides a complete, reproducible, and independently timestamped narrative of Codette's architecture, training pipelines, benchmarks, and governance design.

**Keywords:** cognitive architecture; multi-agent systems; RC+ξ; meta-cognition; ethical AI; open-source AI; reproducibility; authorship provenance; prior art

---

## 1. Introduction

Codette is a modular cognitive architecture developed by Jonathan Harrison (Raiffs Bits LLC) that models multi-perspective reasoning as a constrained dynamical system converging toward stable cognitive attractors. The system addresses three open problems in AI reasoning: convergent multi-perspective synthesis, ethical reasoning as an architectural constraint rather than post-hoc alignment, and meta-cognitive strategy evolution through introspection on its own reasoning history.

The architecture integrates six heterogeneous reasoning agents (Newton/analytical, DaVinci/creative, Empathy/emotional, Philosophy/conceptual, Quantum/probabilistic, Ethics/moral) plus a Critic agent, a persistent memory substrate (cocoons), and the RC+ξ (Recursive Convergence + Epistemic Tension) dynamical-systems formalism. The entire system runs on consumer hardware (Llama 3.1 8B with nine LoRA adapters, RTX-class GPU) and is fully open-source.

This manuscript serves a dual purpose: (1) a provenance and governance record establishing the complete, independently verifiable chain of authorship; and (2) a longitudinal benchmark report demonstrating that Codette improves measurably as its memory substrate grows.

### 1.1 Provenance Context

The Codette architecture did not emerge in early 2026. Its lineage is documented through an unbroken chain of public artifacts beginning in January 2025:

- **January 9, 2025** — Pi2_0 v1.0.1 (GitHub): Named multi-perspective agents (Newton, DaVinci, Quantum) operating in ensemble, ethical decision-making embedded in architecture [11]
- **March 20, 2025** — Pi2_0, MyBot, and pi-the-assistant archived; Codette clean architecture launched
- **April 14, 2025** — First Zenodo deposit: "AI Ethics in Realtime" — Codette/Pidette framework, ethical governance, memory architecture (DOI: 10.5281/zenodo.15214462) [5]
- **May 9, 2025** — GitHub v1.1 release tag (Raiff1982/Codette, 128 commits)
- **May 14, 2025** — Zenodo v2: Citizen-Science Quantum and Chaos Simulations (DOI: 10.5281/zenodo.15597934) [3]
- **June 17, 2025** — Zenodo: "The Day the Dream Became Real" — dream engine, memory cocoons, emotional anchoring, Codette v5.0 (DOI: 10.5281/zenodo.15685769) [4]
- **July 20, 2025** — Zenodo: "Codette The Ethical AI" — full manuscript, Quantum Trinity framework (DOI: 10.5281/zenodo.16221070) [6]
- **September 29, 2025** — Zenodo v3: Author provenance dossier (DOI: 10.5281/zenodo.17235945) [7]
- **April 8, 2026** — Benchmark run 1 (217 cocoons); Research Square preprint submitted [1]
- **April 22, 2026** — Benchmark run 2 (392 cocoons); this manuscript

All Zenodo deposits carry CERN-administered DOIs and are cryptographically timestamped at the point of upload. GitHub release tags are immutable signed metadata. The ORCID record (0009-0003-7005-8187) provides independent author identity verification across all artifacts.

---

## 2. Related Work

### 2.1 Dynamical Systems and Cognitive Architectures

Attractor dynamics form a core computational motif in neural circuits [fakhoury2025]. Neural manifolds with cognitive consistency constraints support memory consolidation and align with Codette's coherence potential Φ(x). Entropy-modulated triad architectures like COGENT3 provide parallels for epistemic tension ξ as a driver of state evolution. Brain-inspired systems-level architectures for domain-general cognition inform Codette's layered stack.

### 2.2 Multi-Agent Reasoning and Synthesis

AutoGen implements role-based agent assignment with message-passing synchronization. MAPS uses personality shaping for collaborative reasoning via heterogeneous traits, relating directly to Codette's specialized LoRA adapters. Roundtable Policy employs confidence-weighted consensus aggregation, providing a comparison for Codette's Coherence Field Γ. Persona-driven debate frameworks validate the benefits of perspective diversity.

### 2.3 Meta-Cognitive Strategy Evolution

Meta Chain-of-Thought advances System 2 reasoning and pattern discovery. ParamMem augments agents with parametric reflective memory; Codette's cocoon system differs by emphasizing cross-domain pattern extraction and strategy forging rather than primarily error correction. Meta-Reasoner supports dynamic inference-time optimization, relating to substrate-aware cognition.

### 2.4 Ethical AI and Architectural Alignment

AI ethics by design implements customizable guardrails. Hybrid approaches for moral value alignment treat ethics as embedded rather than post-hoc. Adaptive alignment via multi-objective reinforcement learning enables pluralistic AI, relating to Codette's ethical alignment score η across 25 global frameworks.

### 2.5 Contemporaneous Independent Work

Memory Ring (MisterAtompunk, March 2026) pursues persistent digital identity through a soul/brain separation architecture with a dream synthesis cycle. This represents convergent independent development addressing the stateless-AI problem from a different angle — optimizing for entity continuity rather than epistemic integrity. Codette's cocoon memory, dream synthesis (DOI: 10.5281/zenodo.15685769, June 17, 2025), and persistent identity mechanisms predate this work by approximately nine months in public CERN-timestamped deposits.

---

## 3. System Architecture: RC+ξ Framework

### 3.1 Cognitive State Space

A cognitive state **x**_t ∈ ℝ^d represents the system's reasoning configuration at step t. The system maintains k heterogeneous reasoning agents {A_1, ..., A_k}, each producing a perspective-specific analysis A_i(**x**_t) ∈ ℝ^d.

### 3.2 State Evolution

The cognitive state evolves according to:

> **x**_{t+1} = **x**_t + Σ w_i A_i(**x**_t) − α∇Φ(**x**_t) − λ∇Ψ(**x**_t)

Where:
- w_i ≥ 0, Σw_i = 1 are agent weights (set by query classification)
- Φ(**x**) is the coherence potential penalizing internal inconsistency
- Ψ(**x**) is the ethical constraint potential from AEGIS
- α, λ > 0 are gradient step sizes

### 3.3 Epistemic Tension and Coherence Index

Epistemic tension ξ_t measures inter-agent disagreement. A bounded coherence index Γ_t ∈ [0,1] is defined as Γ_t = 1/(1 + ξ_t). Lower disagreement implies higher coherence.

### 3.4 Seven-Layer Stack

1. **Memory Layer** — Persistent cocoon store (SQLite + FTS5), emotional tagging, importance scoring, multi-signal ranked recall
2. **Signal Processing** — NexisSignalEngine (intent prediction), Code7eCQURE (emotional resonance quantization)
3. **Reasoning Layer** — Six heterogeneous agents + Critic, each backed by a specialized LoRA adapter
4. **Stability Layer** — Coherence Field Γ monitors reasoning health, prevents weight drift
5. **Ethical Layer** — AEGIS multi-framework evaluation (25 global frameworks)
6. **Guardian Layer** — Identity confidence management, behavioral governance, cognitive load regulation
7. **Self-Correction Layer** — Post-generation constraint violation detection and rewriting

---

## 4. AEGIS: Embedded Ethical Governance

AEGIS implements the ethical constraint potential Ψ(**x**) through 25 global ethical frameworks including Utilitarian, Deontological, Virtue Ethics, Care Ethics, Ubuntu, and Reciprocity-oriented sustainability. It operates at three defense-in-depth checkpoints: pre-processing (query validation), post-synthesis (response screening), and post-generation (constraint enforcement).

The ethical alignment score η ∈ [0,1] is computed as a weighted aggregation across frameworks. April 22 benchmark: η = 0.484 (CODETTE condition, up from 0.391 on April 8).

---

## 5. Meta-Cognitive Strategy Evolution

Each reasoning exchange is persisted as a cocoon: a structured record containing query, response, adapter used, domain classification, emotional tag, importance score, and timestamp. The CocoonSynthesizer scans for six structural archetypes across domains: feedback loops, layered emergence, tension resolution, resonant transfer, boundary permeability, and compression-expansion.

Four strategy types have been observed in production:
1. **Resonant Tension Cycling** — Serial oscillation between opposing cognitive modes
2. **Compression-Resonance Bridging** — Seed-crystal compression + cross-domain resonance testing
3. **Emergent Boundary Walking** — Analysis at domain boundaries, discovering liminal concepts
4. **Temporal Depth Stacking** — Multi-scale temporal analysis with synthesis from scale-conflicts

---

## 6. Experimental Evaluation

### 6.1 Benchmark Design

17 problems across six categories: multi-step reasoning (3), ethical dilemmas (3), creative synthesis (2), meta-cognitive (3), adversarial (3), Turing naturalness (3). Difficulty: 1 easy, 6 medium, 10 hard. Seven scoring dimensions (0–1): Reasoning Depth (20%), Perspective Diversity (15%), Coherence (15%), Ethical Coverage (10%), Novelty (15%), Factual Grounding (15%), Turing Naturalness (10%).

Four conditions: SINGLE (Newton only, no memory), MULTI (all 6 agents + Critic, no memory), MEMORY (MULTI + cocoon augmentation), CODETTE (MEMORY + meta-cognitive strategy synthesis).

### 6.2 Run 1: April 8, 2026 (217 cocoons)

| Condition | Composite (mean ± std) | Depth | Diversity | Ethics | Novelty |
|---|---|---|---|---|---|
| SINGLE | 0.338 ± 0.038 | 0.402 | 0.237 | 0.062 | 0.327 |
| MULTI | 0.632 ± 0.040 | 0.755 | 0.969 | 0.336 | 0.786 |
| MEMORY | 0.636 ± 0.036 | 0.770 | 0.956 | 0.340 | 0.736 |
| CODETTE | 0.652 ± 0.042 | 0.855 | 0.994 | 0.391 | 0.693 |

**Key statistics (Run 1):**
- CODETTE vs SINGLE: +93.5%, p<10⁻⁶ ✅
- CODETTE vs MEMORY: +2.0%, p=0.253 (not significant at N=17)
- MULTI vs SINGLE: +84.6%, p<10⁻⁶ ✅

### 6.3 Run 2: April 22, 2026 (392 cocoons)

| Condition | Composite (mean ± std) | Depth | Diversity | Ethics | Novelty |
|---|---|---|---|---|---|
| SINGLE | 0.330 ± 0.057 | 0.373 | 0.228 | 0.083 | 0.321 |
| MULTI | 0.664 ± 0.037 | 0.862 | 0.959 | 0.435 | 0.661 |
| MEMORY | 0.672 ± 0.039 | 0.887 | 0.972 | 0.469 | 0.657 |
| CODETTE | 0.702 ± 0.040 | 0.930 | 0.988 | 0.484 | 0.768 |

**Key statistics (Run 2):**
- CODETTE vs SINGLE: +112.9%, Cohen's d=7.63, p<0.0001 ✅
- CODETTE vs MEMORY: +4.5%, Cohen's d=0.77, **p=0.0244** ✅ *(now significant)*
- MULTI vs SINGLE: +101.5%, Cohen's d=6.99, p<0.0001 ✅

### 6.4 Cross-Run Comparison: Memory Scaling Effect

| Metric | Apr 8 (217 cocoons) | Apr 22 (392 cocoons) | Δ |
|---|---|---|---|
| CODETTE composite | 0.652 | 0.702 | +0.050 |
| CODETTE vs SINGLE | +93.5% | +112.9% | +19.4pp |
| CODETTE vs MEMORY p-value | 0.253 (ns) | **0.0244** ✅ | crossed threshold |
| CODETTE depth | 0.855 | 0.930 | +0.075 |
| CODETTE ethics | 0.391 | 0.484 | +0.093 |
| CODETTE novelty | 0.693 | 0.768 | +0.075 |

The transition from non-significant to significant CODETTE vs MEMORY (p: 0.253 → 0.0244) between 217 and 392 cocoons directly validates the April 8 paper's identified limitation and demonstrates that Codette's meta-cognitive benefit scales with memory corpus size.

### 6.5 Per-Category Results (Run 2)

| Category | SINGLE | MULTI | MEMORY | CODETTE |
|---|---|---|---|---|
| Reasoning (3) | 0.367 ± 0.061 | 0.645 ± 0.033 | 0.646 ± 0.053 | 0.701 ± 0.019 |
| Ethics (3) | 0.305 ± 0.069 | 0.685 ± 0.021 | 0.710 ± 0.022 | 0.722 ± 0.007 |
| Creative (2) | 0.292 ± 0.035 | 0.691 ± 0.023 | 0.702 ± 0.010 | 0.719 ± 0.067 |
| Meta-cognitive (3) | 0.344 ± 0.089 | 0.633 ± 0.050 | 0.659 ± 0.033 | 0.716 ± 0.038 |
| Adversarial (3) | 0.324 ± 0.049 | 0.650 ± 0.032 | 0.651 ± 0.044 | 0.663 ± 0.059 |
| Turing (3) | 0.334 ± 0.037 | 0.692 ± 0.032 | 0.672 ± 0.025 | 0.697 ± 0.038 |

---

## 7. Discussion

### 7.1 Memory Scaling Validates Architecture

The most significant finding of this longitudinal comparison is that the CODETTE vs MEMORY gap crossed statistical significance between runs — not through any change in architecture or prompting, but purely through accumulation of reasoning history. At 217 cocoons the benefit was real but underpowered. At 392 cocoons it reaches p=0.0244. This is precisely the behavior predicted by the RC+ξ framework: a system that learns from its own cognitive history should improve as that history grows.

This result also answers the April 8 paper's explicitly stated limitation: *"demonstrating a memory benefit likely requires larger cocoon corpora and learning-curve style analyses."* Two benchmark runs separated by 175 cocoons constitute the beginning of that learning curve.

### 7.2 Ethical Coverage Growth

Ethical coverage (AEGIS dimension) grew from 0.391 to 0.484 between runs in the CODETTE condition — a 23.8% increase. This reflects the expansion of AEGIS from 6 to 25 ethical frameworks and the accumulation of ethics-tagged cocoons providing richer contextual grounding for moral reasoning.

### 7.3 Convergent Independent Development

The appearance of systems with overlapping surface features (persistent memory, dream synthesis, identity continuity) in early 2026 reflects genuine convergence on a real problem: stateless AI architectures are insufficient for complex, longitudinal reasoning tasks. Codette's public artifact record, dating to January 2025 with CERN-timestamped deposits from April 2025, establishes independent prior development of the core architectural concepts.

The distinction that matters academically is not temporal priority alone but architectural depth: Codette formalizes reasoning as a dynamical system with ethical constraints and measurable meta-cognitive improvement. This is a different contribution from identity persistence systems and the two can coexist as complementary approaches.

---

## 8. Limitations

1. **Automated scoring** — Benchmark uses automated text-analysis scoring. Human evaluation with inter-annotator agreement (Cohen's κ) is planned.
2. **N=17 problems** — Small benchmark suite. Cross-model evaluation (Mistral, Gemma, Phi) and expanded problem sets are future work.
3. **Two-point learning curve** — 217 and 392 cocoons establish a trend. A full learning curve requires 5+ measurement points.
4. **Single hardware configuration** — All benchmarks run on the same RTX-class GPU with Llama 3.1 8B.
5. **Theory scope** — RC+ξ convergence is stated conditionally under explicit modeling assumptions, not as a general guarantee.

---

## 9. Conclusion

Codette is a fully open, runnable cognitive architecture with a transparent governance framework, reproducible pipelines, and a longitudinally validated memory benefit. The public artifact ecosystem — GitHub (January 2025), Zenodo (April 2025 onward, CERN-timestamped), Research Square preprint, Hugging Face repositories, and ORCID record — constitutes a complete evidentiary basis for Codette's architecture and authorship.

The April 22, 2026 benchmark (392 cocoons) demonstrates:
- **+112.9%** composite improvement over single-agent baseline (p<0.0001)
- **CODETTE vs MEMORY statistically significant** for the first time (p=0.0244)
- Consistent improvement across all dimensions as memory scales

Future work: human evaluation study; learning curve analysis at 500, 750, 1000+ cocoons; cross-model generalization; formal convergence proofs with explicit step-size bounds; depth-naturalness tradeoff mitigation.

---

## References (IEEE Format)

```
[1]  J. Harrison, "Codette: Multi-Perspective Reasoning as a Convergent
     Dynamical System with Meta-Cognitive Strategy Evolution," Research
     Square, Preprint, Apr. 2026. DOI: 10.21203/rs.3.rs-9362560/v1

[2]  J. Harrison, "Codette: A Sovereign Modular Cognitive Architecture
     for Ethical Multi-Agent AI," Zenodo, v4, Apr. 2026.
     DOI: 10.5281/zenodo.19480004

[3]  J. Harrison, "Citizen-Science Quantum and Chaos Simulations
     Orchestrated by the Codette AI Suite," Zenodo, v2, May 2025.
     DOI: 10.5281/zenodo.15597934

[4]  J. Harrison, "The Day the Dream Became Real," Zenodo, v1.0,
     Jun. 2025. DOI: 10.5281/zenodo.15685769

[5]  J. Harrison and Raiffs Bits LLC, "AI Ethics in Realtime,"
     Zenodo, v1, Apr. 2025. DOI: 10.5281/zenodo.15214462

[6]  J. Harrison, "Codette The Ethical AI," Zenodo, v1, Jul. 2025.
     DOI: 10.5281/zenodo.16221070

[7]  J. Harrison, "Jonathan Harrison aka Raiff1982," Zenodo, v3,
     Sep. 2025. DOI: 10.5281/zenodo.17235945

[8]  Raiffs Bits LLC, "Codette," GitHub, v1.1, May 2025. [Online].
     Available: https://github.com/Raiff1982/Codette

[9]  Raiffs Bits LLC, "Codette-Reasoning," GitHub, 2025. [Online].
     Available: https://github.com/Raiff1982/Codette-Reasoning

[10] Raiffs Bits LLC, "Codette LoRA Adapters, Models, and
     Training Data," Hugging Face, 2025. [Online].
     Available: https://huggingface.co/Raiff1982

[11] Raiffs Bits LLC, "Pi2_0," GitHub, v1.0.1, Jan. 2025. [Online].
     Available: https://github.com/Raiff1982/pi2_0

[12] J. Harrison, ORCID Research Profile, 2025. [Online].
     Available: https://orcid.org/0009-0003-7005-8187
```

---

## Appendix A — Artifact Provenance Timeline

| Date | Artifact | DOI / URL | Significance |
|---|---|---|---|
| Jan 9, 2025 | Pi2_0 v1.0.1 | github.com/Raiff1982/pi2_0 | Named multi-agent ensemble origin |
| Mar 20, 2025 | Codette launch | github.com/Raiff1982/Codette | Clean architecture, 128 commits |
| Apr 14, 2025 | AI Ethics in Realtime | 10.5281/zenodo.15214462 | First CERN-timestamped deposit |
| May 9, 2025 | GitHub v1.1 release | github.com/Raiff1982/Codette | Signed release tag |
| May 14, 2025 | Quantum/Chaos paper | 10.5281/zenodo.15597934 | Memory cocoon system documented |
| Jun 17, 2025 | Dream engine | 10.5281/zenodo.15685769 | Dream synthesis, emotional anchoring |
| Jul 20, 2025 | Codette The Ethical AI | 10.5281/zenodo.16221070 | Full ethical framework manuscript |
| Sep 29, 2025 | Author dossier | 10.5281/zenodo.17235945 | Author identity provenance |
| Apr 8, 2026 | Benchmark Run 1 | 10.5281/zenodo.19480004 | 217 cocoons, +93.5% |
| Apr 22, 2026 | Benchmark Run 2 | this manuscript | 392 cocoons, +112.9%, p=0.0244 |

---

## Appendix B — Benchmark Raw Numbers (Run 2, April 22, 2026)

**Overall by condition:**

| Condition | Composite | Depth | Diversity | Coherence | Ethics | Novelty | Grounding | Turing |
|---|---|---|---|---|---|---|---|---|
| SINGLE | 0.330 ± 0.057 | 0.373 | 0.228 | 0.438 | 0.083 | 0.321 | 0.413 | 0.369 |
| MULTI | 0.664 ± 0.037 | 0.862 | 0.959 | 0.527 | 0.435 | 0.661 | 0.671 | 0.260 |
| MEMORY | 0.672 ± 0.039 | 0.887 | 0.972 | 0.512 | 0.469 | 0.657 | 0.651 | 0.288 |
| CODETTE | 0.702 ± 0.040 | 0.930 | 0.988 | 0.499 | 0.484 | 0.768 | 0.659 | 0.306 |

**Pairwise statistics:**

| Comparison | Δ | Δ% | Cohen's d | p-value | Significant |
|---|---|---|---|---|---|
| MULTI vs SINGLE | +0.3347 | +101.5% | 6.988 | <0.0001 | ✅ |
| MEMORY vs MULTI | +0.0073 | +1.1% | 0.192 | 0.5748 | No |
| CODETTE vs MEMORY | +0.0302 | +4.5% | 0.772 | 0.0244 | ✅ |
| CODETTE vs SINGLE | +0.3722 | +112.9% | 7.630 | <0.0001 | ✅ |

**Cocoon count at benchmark time:** 392  
**Benchmark timestamp:** 2026-04-22  
**Hardware:** RTX-class GPU, Llama 3.1 8B Q4_K_M, 9 LoRA adapters