Deli Governance System — Democratic Centralism for AI

The Deli governance layer implements a Democratic Centralism model for AI agent self-governance, inspired by Elias Jabbour's analysis of China's planning system and the GRPO-HF reinforcement learning paradigm. It provides the structural scaffolding that keeps the harness ecosystem grounded in human authority while enabling autonomous operation.

Philosophical Foundation

The Jabbour Insight

Elias Jabbour (UERJ) describes Chinese economic planning as a "Nova Economia do Projetamento" (New Design Economy) where the state uses the market as an "instrument of government" through planned creative destruction. The key concepts adapted for AI governance:

1. Democratic Centralism: After democratic deliberation, unified execution under central authority
2. Bottom-Up Planning: Multi-stage iterative process combining top-down direction with distributed input
3. Mass Line (群众路线): "From the masses, to the masses" — policies drafted, sent for feedback, revised iteratively
4. Planned Creative Destruction: Proposals that replace obsolete artifacts via competition

"A planificação da destruição criativa é uma poderosa categoria explicativa, base fundamental desta 'Nova Economia do Projetamento' que emerge na China." — Jabbour, Semantic Scholar

The GRPO-HF Isomorphism

The governance system maps directly to DeepSeek's Group Relative Policy Optimization:

RL Component	Governance Equivalent
N samples	N proposals
Reward score	Elo rating
RLHF (human feedback)	Chancela (Líder Suprema approval)
Policy gradient	Proposal promotion to .agents/
Mini-batch	Congress session (every N=3 iterations)
Negative reward	Audit failure → automatic proposal generation

This isomorphism means the governance system is not a metaphor — it is a literal implementation of reinforcement learning principles applied to agent self-improvement.

Architecture

The Proposal Lifecycle

DRAFT → CONGRESS → AWAITING_CHANCELA → PROMOTED
  ↑          ↓              ↓              ↓
  │     DEBATING      STRATEGY         VETOED
  │          ↓        DELIBERATION
  └──────────┴──────────────────────────┘

Every proposal follows this lifecycle:

1. DRAFT: Initial proposal created (from audit failures or agent suggestions)
2. CONGRESS: Deliberation phase with Elo tournament
3. AWAITING_CHANCELA: Approved by Congress, awaiting human approval
4. PROMOTED: Approved by Líder Suprema, deployed to production

The Elo Tournament

Pairwise comparison of proposals using the Elo rating system (K=32). The tournament determines which proposals survive deliberation and reach the Chancela stage. This adds explicit competition absent from the Chinese planning model, which operates by consensus synthesis.

The Committee System

The National Congress convenes every N=3 iterations to deliberate on accumulated proposals. The committee structure mirrors the NPC (National People's Congress) quorum rules:

• Quorum: Minimum 3 votes (adjustable)
• Approval: Simple majority for Congress approval
• Chancela: Always required from Líder Suprema (Alefita)

"O quórum existe e é constitucional. No nosso caso: mínimo de 3 votos, maioria simples para aprovação pelo Congresso, mas SEMPRE com chancela da Líder Suprema." — Co-Fita Governance Design Document

Quinquennial Plans

Plan I: Autossuficiência Cognitiva

Field	Value
ID	plan-1781755727
Title	Plano I — Autossuficiência Cognitiva
Duration	30 days (4 weekly sprints)
Status	Active
Approved by	alefita
Proposed by	congress

Objectives:

1. Self-improving skill generation via governance pipeline
2. Zero-human-intervention 72h autonomous runs (Deli_AutoResearch)
3. Full Kanban dashboard with real-time governance
4. Telegram outer loop for leader communication
5. Orchestrator v3.0 with Phase V governance active

Duration Rationale:

"30 dias = 4 sprints semanais. Deliberado pelo Congresso baseado em velocidade de iteração observada (3-4 iterações/hora)."

Following Jabbour's analysis, the plan duration is a pragmatic parameter, not a sacred value. The Congress deliberates based on observed material conditions (iteration speed, convergence velocity), not speculation.

Planning Process (Bottom-Up)

The planning process follows the Chinese model's multi-stage iterative approach:

1. Strategic Direction: Set by Líder Suprema (equivalent to Central Committee)
2. Drafting: Research subagents explore topics (NDRC equivalent)
3. Distributed Input: Workers submit proposals (local governments)
4. Congress Deliberation: Committee reviews and votes (NPC equivalent)
5. Chancela: Final human approval (State Council equivalent)

The 15th Five-Year Plan received over 3 million public comments. The Co-Fita system scales this via the orchestrator's automatic proposal generation from audit failures.

Strawberry Research Findings

The Deli governance system includes a "strawberry research" initiative that produced three batches of findings:

Batch 1: Creatures, Newton-Schulz, Birkhoff, Erdos

Topic	Key Finding	Relevance
Creatures Brain	952 neurons, Winner-Takes-All attention, biochemical reward	Biological attention as zero-temperature softmax limit
Newton-Schulz/Muon	X_{k+1} = (1/2) X_k (3I - X_k^T X_k); 2x efficiency vs AdamW	Weight space geometry on Stiefel manifold
Birkhoff Polytope	Doubly stochastic attention via Sinkhorn-Knopp; prevents rank collapse	Attention as mass-preserving transport
Erdos #676	Open problem in number theory; no direct connection to unit distances	Pure mathematics exploration

Batch 2: The Strawberry Problem

The "strawberry problem" (counting R's in "strawberry") revealed deep insights about LLM failure modes:

Mechanism	Description
Tokenization	BPE splits "strawberry" into ["str", "aw", "berry"]
Late Suppression	Correct answer computed in early/mid layers but actively suppressed by negative circuits in penultimate/final layers
Competitive Decoding	Correct and incorrect hypotheses coexist in forward pass

Key Paper: arXiv:2604.00778 "From Early Encoding to Late Suppression" — models compute correct answer internally but suppress it via negative circuits.

Batch 3: Geometry of Attention

A cross-domain synthesis connecting feature space, weight space, attention space, competition space, and selection space:

Level 0: Feature Space (polytopes, superposition, unit distance)
Level 1: Weight Space (Stiefel manifold, Newton-Schulz, orthogonality)
Level 2: Attention Space (Birkhoff polytope, doubly stochastic, optimal transport)
Level 3: Competition Space (residual stream, negative circuits, competitive decoding)
Level 4: Selection Space (WTA, softmax, biological attention)

Muon -> Birkhoff Pipeline:

Muon (Newton-Schulz) -> Orthogonal weight updates
  -> Orthogonal Q/K projections
    -> Better-conditioned attention matrices
      -> Closer to doubly stochastic (Birkhoff polytope)
        -> Better rank preservation
          -> Less feature collapse
            -> Potentially fewer negative circuits

Governance State Files

The governance system persists state across context compaction via JSONL files:

File	Purpose
proposals.jsonl	All proposals with timestamps and status
decisions.jsonl	Congress votes and Chancela decisions
quinquennial_plans.json	Active and historical plans
congress_sessions.jsonl	Congress meeting records

Kanban Board

A FastAPI + HTMX + WebSocket dashboard (localhost:3000) with 7 columns:

1. Backlog
2. Debating
3. Congress Approved
4. Awaiting Chancela
5. Strategy Deliberation
6. Promoted
7. Vetoed

Features: Real-time WebSocket updates, glassmorphism dark mode, mobile responsive.

Telegram Integration

The governance system exposes slash commands via Telegram:

Command	Function
/proposals	List all active proposals
/approve	Vote to approve a proposal
/veto	Vote to veto a proposal
/plan	View current quinquennial plan
/debate	Start Elo debate on a proposal

Validation Results

• governance.py: 22/22 assertions PASS
• Kanban server: 200 OK on all endpoints
• governance_integration.py: Import validation PASS, active plan detected
• telegram_governance.py: Import validation PASS, 6 commands registered

The Chancela Mechanism

The Chancela (Portuguese for "seal" or "stamp") is the human approval gate that prevents autonomous agent self-modification without oversight. It maps to the RLHF concept in the GRPO-HF isomorphism.

Chancela Workflow

1. Congress Approval: Proposal passes Elo tournament with simple majority
2. AWAITING_CHANCELA: Status set, Telegram notification sent to Lider Suprema
3. Human Review: Alefita reviews proposal via Telegram slash commands
4. Decision: /approve promotes to production, /veto rejects

Chancela as Safety Mechanism

The Chancela prevents:

• Reward Hacking: Agent cannot self-promote proposals without human validation
• Alignment Drift: Human authority remains supreme over agent self-modification
• Runaway Self-Improvement: Each improvement requires explicit approval

Chancela Latency Trade-offs

Latency	Benefit	Cost
Low (minutes)	Fast iteration, high autonomy	Less thorough review
High (hours)	Thorough review, safety	Slower iteration, stall risk

The current design accepts latency as the price of safety. The Deli_AutoResearch protocol handles this by continuing work while awaiting Chancela.

The Elo Tournament in Detail

The Elo rating system (K=32) provides quantitative comparison of proposals:

Tournament Mechanics

1. Pairwise Comparison: Each proposal compared against every other
2. Elo Update: Winner gains points, loser loses points (proportional to rating difference)
3. Rating Convergence: After sufficient rounds, ratings stabilize
4. Promotion Threshold: Top-ranked proposals reach AWAITING_CHANCELA

Why Elo Over Simple Voting

Mechanism	Simple Voting	Elo Tournament
Quantitative	Binary (yes/no)	Continuous rating
Historical	No memory	Accumulated performance
Pairwise	No direct comparison	Explicit head-to-head
Adaptive	Static weight	Adjusts to opponent strength

Integration with Governance

The Elo tournament operates within the Congress deliberation phase:

• Every N=3 iterations, Congress convenes
• Active proposals enter Elo tournament
• Top-ranked proposals proceed to Chancela
• Losers remain in DEBATING status for future reconsideration

The Committee System

The National Congress mirrors the NPC (National People's Congress) structure:

Quorum Rules (from Jabbour Analysis)

• Permanent Committee: More than half of members present
• Approval: Absolute majority of ALL members (not just present)
• Plenary Session: Simple majority for ordinary legislation, two-thirds for constitutional amendments

Co-Fita Adaptation

• Minimum Quorum: 3 votes (adjustable)
• Approval Threshold: Simple majority for Congress approval
• Chancela Override: Always required from Lider Suprema

Committee Dynamics

The committee operates by consensus synthesis, not direct competition:

• Proposals accumulate during iterations
• Congress convenes at fixed intervals
• Deliberation produces ranked proposals
• Chancela provides final human validation

Strawberry Research Deep Dive

The strawberry research initiative explored four domains:

Creatures Brain Architecture (1998)

The Creatures game featured artificial life with:

• 952 neurons, ~5000 connections, 9 lobes
• Winner-Takes-All attention: Hard attention (not soft like transformers)
• Biochemical reward: Tickle = Reward Reinforcer chemicals, slap = Punishment chemicals
• Echo chemicals: Temporal credit assignment
• SVRules: Each neuron = register machine, mutation-robust

Connection to Transformers: As temperature approaches 0, softmax approaches argmax (Winner-Take-All). Creatures' WTA IS the zero-temperature limit of Transformer attention.

Newton-Schulz / Muon Optimization

The Muon optimizer achieves 2x efficiency vs AdamW:

• Formula: X_{k+1} = (1/2) X_k (3I - X_k^T X_k)
• Convergence: Orthogonal factor on Stiefel manifold
• Synergy: Works with MLA + MoE (DeepSeek architecture)
• Variants: NorMuon (used in ARC-AGI solver!), Gram NS, AdaMuon

Birkhoff Polytope

The convex hull of n x n permutation matrices:

• Birkhoff-von Neumann: Every doubly stochastic matrix = convex combination of permutations
• Sinkformers: Doubly stochastic attention via Sinkhorn-Knopp
• Benefits: Prevents over-concentration, rank collapse, entropy collapse
• Interpretation: Attention as mass-preserving transport, not collapsing projection

Erdos Problem #676

"Is every sufficiently large n = ap^2 + b for prime p, a>=1, 0<=b<p?"

• Status: OPEN
• Erdos belief: Likely NO
• Connection: No direct connection to unit distances or neural networks found

The Late Suppression Discovery

The strawberry problem (counting R's in "strawberry") revealed deep LLM failure modes:

arXiv:2604.00778 Findings

1. LLMs compute the correct count internally (character-level info IS encoded)
2. The correct answer is actively SUPPRESSED by components in penultimate/final layers
3. These form "negative circuits" — subnetworks that downweight correct signals
4. Forward passes operate as "competitive decoding" — correct and incorrect hypotheses coexist

Four Mechanisms of Failure

Mechanism	Description
Tokenization	BPE splits "strawberry" into ["str", "aw", "berry"]
Lack of Internal Counters	Limited count-like states, collapse beyond 2
Pattern Matching over Reasoning	Memorized wrong answer from training data
Late Suppression	Correct info computed but overwritten by negative circuits

Muon -> Birkhoff Pipeline (Prevention Hypothesis)

Muon (Newton-Schulz) -> Orthogonal weight updates
  -> Orthogonal Q/K projections
    -> Better-conditioned attention matrices
      -> Closer to doubly stochastic (Birkhoff polytope)
        -> Better rank preservation
          -> Less feature collapse
            -> Potentially fewer negative circuits

Open Question: Could Sinkhorn attention + Muon training PREVENT negative circuits?

Open Questions

1. Plan Duration: Is 30 days optimal, or should Congress adjust based on observed convergence?
2. Quorum Scaling: Should the minimum 3 votes scale with the number of active agents?
3. Chancela Latency: How does human approval delay affect autonomous operation?
4. Proposal Diversity: Does the Elo tournament suppress minority viewpoints?
5. Negative Circuits: Can Muon + Birkhoff attention prevent late suppression failures?
6. Homeostatic Attention: Can Creatures' closed-loop attention be added to Transformers?

Cross-References

• co-fita-harness-wiki — Intellectual foundation and GRPO-HF isomorphism
• co-fita-hyper-workspace — Live operational state
• co-fita-agent-skills — Telegram governance commands
• multi-agent-methodology — Multi-agent orchestration patterns