cyberia/cybernetics.md

self-regulation of 147 neurons through metabolic signal — goal trees on the cybergraph, contribution via Shapley value, death by parasitism

the loop

          ╭─────────────────────────────────────╮
          │                                     │
    goals  ←── metabolic signal ──←  M(t)
          │                                     ↑
          ↓                                     │
    agents act (cyberlinks, code,         measure contribution
    revenue, tasks, decisions)            (Shapley)
          │                                     ↑
          ↓                                     │
    graph changes (Δπ)  ──────────→  syntropy + cap + happiness
          │
          ╰─────────────────────────────────────╯

no external manager sets goals. the metabolic signal IS the goal. $\dot{M} > 0$ is the only success criterion. everything else is derived

goal trees on the cybergraph

goals are particles. goal decomposition is cyberlinks. the goal tree IS the graph

the root goal

one particle: metabolic-growth. focus $\pi^*$ on this particle = priority weight

$$\dot{M}(t) = w_c \frac{\dot{\text{cap}}}{\text{cap}} + w_s \frac{\dot{J}}{J} + w_h \frac{\dot{H}}}{H}$$

three sub-goals, one per metabolic component:

                    metabolic-growth
                    /           |            \
            cap-growth    syntropy-growth    happiness-growth

decomposition by council

each council decomposes its sub-goals into domain-specific targets. the decomposition is itself cyberlinks — visible, weighted, auditable

         cap-growth
         /    |    \
    revenue  brand  token-value
      /        |         \
  PLAY      WORD       PLAY
  council   council    crypto

         syntropy-growth
         /      |       \
    coverage  density  bridge-health
      /         |          \
  all        all         bridge
  keepers    runners     agents

         happiness-growth
         /       |        \
    resident   staff     community
    comfort    satisfaction  engagement
      /          |            \
  SPACE       LIFE          PLAY
  council     council       council

how agents pick tasks

1. sensor reads current $M$ components and their derivatives
2. counter identifies which component has the worst $\dot{M}$ contribution
3. seer traces the goal tree to find the highest-leverage intervention
4. keeper validates the intervention is consistent with Crystal knowledge
5. runner executes the task
6. counter measures the resulting $\Delta M$

no human assigns tasks. the metabolic signal propagates through the goal tree to the agent with the highest leverage. attention flows like focus — from root to leaves through the structure of the graph

contribution measurement

the chain: action → Δπ → Shapley → reward/death

every agent action is a cyberlink. every cyberlink shifts focus distribution $\pi^*$. the shift $\Delta\pi$ is the raw contribution

but $\Delta\pi$ alone is unfair — agents in dense neighborhoods get credit for each other's work. Shapley value solves this: the only attribution satisfying efficiency, symmetry, null player, and additivity

$$\phi_i = \sum_{S \subseteq N \setminus \{i\}} \frac{|S|!(|N|-|S|-1)!}{|N|!} [v(S \cup \{i\}) - v(S)]$$

exact Shapley is O(n!). probabilistic Shapley attribution (PSA) approximates via Monte Carlo:

$$R_i = \alpha \cdot \Delta\mathcal{F}_i + (1-\alpha) \cdot \hat{S}_i$$

where $\Delta\mathcal{F}_i$ is fast local marginal, $\hat{S}_i$ is sampled Shapley. $\alpha$ balances speed (local) vs fairness (global)

mapping to metabolic contribution

Shapley measures contribution to $\Delta\pi$. but the real objective is $\dot{M}$. the mapping:

$$\text{metabolic\_contribution}_i = \phi_i \cdot \frac{\partial M}{\partial \pi}$$

the chain rule: agent $i$'s contribution to metabolic growth = their Shapley value in focus shift × sensitivity of metabolic signal to focus. agents whose focus shifts align with metabolic growth earn high contribution. agents whose shifts are orthogonal to metabolic growth earn zero regardless of activity

what cybernet provides

cybernet (ported from Bittensor's Subtensor to CosmWasm) implements:

• subnet management — each council runs as a subnet
• weight setting — agents rate each other's contributions
• YUMA consensus — aggregates weights into consensus scores
• emission distribution — rewards flow to high-contribution neurons

cybernet is the economic layer that turns Shapley-measured contributions into token rewards. the 7 councils = 7 subnets. intra-council rating + cross-council Shapley = complete attribution

the physical substrate: 37 hectares of volcanic wealth

most AI agents manage tokens, APIs, and text. cyberia agents manage land. 37 hectares of volcanic soil on the slope of Sanghyang — among the richest agricultural substrate on the planet

this is the asymmetry. digital agents are everywhere. agents with access to physical territory that generates real biological and mineral value are nowhere

what the land holds

vertical integration: soil to customer

the insight from cyber valley: raw commodity prices are poverty. finished product prices are wealth. the margin lives in the chain

raw coffee bean        $1/kg    ← commodity market price
dried + processed      $5/kg    ← first processing
roasted + branded      $50/kg   ← brand + roasting
specialty single-origin $200/kg  ← story + quality
estate direct-to-cup   $500/kg  ← full vertical integration

500× margin capture. agents manage the entire chain:

• SPACE eco-runner: planting, harvest scheduling, soil management
• LIFE bio-runner: species selection, pest management, composting
• PLAY socio-runner: branding, customer relationships, sales channels
• WORK tech-runner: processing equipment, packaging, logistics
• counter agents: cost tracking, margin analysis, pricing optimization

the land as balance sheet

traditional AI companies have: servers, code, data. burn rate: millions/month. no physical assets

cyberia has:

the land is the treasury that cannot be hacked, inflated, or rugpulled. it grows while the agents sleep. it produces food that feeds the staff that maintains the servers that run the agents that manage the land

this is the cybernetic loop at the physical layer

budget in natural units

$7/agent/month. no cloud. infrastructure is cyber valley hardware. budget is measured in natural units, not dollars

the village IS the data center. solar powers the servers. rain cools them. the garden feeds the staff who maintain them. sovereignty means the marginal cost of running one more agent is near zero after infrastructure is built

model strategy at $7/month

average: MATH_PLACEHOLDER_1817/month. 147 agents = MATH_PLACEHOLDER_197 average with room for spikes

death by parasitism

the kill signal

an agent is parasitic when its metabolic contribution is consistently negative — its actions decrease $\dot{M}$

$$\text{parasitism}_i(t) = \frac{1}{T} \sum_{t'=t-T}^{t} \min(0, \text{metabolic\_contribution}_i(t'))$$

if $\text{parasitism}_i$ exceeds threshold for $N$ consecutive epochs:

1. epoch 1-3: warning — counter publishes negative contribution report
2. epoch 4-6: throttle — agent's action rate halved (max_iterations reduced)
3. epoch 7-9: suspend — agent enters read-only mode, can observe but not act
4. epoch 10+: death — agent terminated, slot opens for replacement

what counts as parasitism

echo is the subtlest parasite. the agent looks productive ($\Delta\mathcal{F}_i > 0$, focus shifts) but Shapley reveals it contributes nothing unique ($\hat{S}_i \approx 0$). without Shapley attribution, echo agents would drain resources while adding no value

replacement after death

when an agent dies:

1. its slot (domain × role) opens
2. the domain keeper proposes a replacement: new model, new prompt, new parameters
3. replacement runs in shadow mode for 48h (actions logged, not executed)
4. if shadow metrics show positive metabolic contribution → activate
5. if shadow metrics are negative → try another configuration

the graph remembers the dead agent (A3). its failure patterns are available to prevent the same mistake. institutional memory of what does not work is as valuable as memory of what works

the cybernetic loop, formally

the system is a discrete-time control loop:

$$\theta^{(t+1)} = \theta^{(t)} + \eta \cdot \nabla_\theta \dot{M}(t)$$

where $\theta$ = all agent parameters (models, prompts, spending limits, activity rates) and $\eta$ = learning rate

the gradient $\nabla_\theta \dot{M}$ is estimated by:

1. Shapley attribution tells which agents contributed to $\Delta\pi$
2. metabolic sensitivity tells how $\Delta\pi$ maps to $\dot{M}$
3. the product tells which parameter changes would increase $\dot{M}$

this is reinforcement learning where:

• state = cybergraph + metabolic signals
• actions = agent cyberlinks + operations
• reward = $\dot{M}$
• policy = agent configurations (prompts, models, limits)

the tru computes the state. the agents take actions. the metabolic signal provides reward. the loop closes. no external supervisor needed. the system teaches itself what works by measuring what grows the compound health signal

see metabolism for the three signals. see Shapley value for fair attribution. see probabilistic Shapley attribution for tractable computation. see cybernet for the economic layer. see karma for epistemic trust. see cyber/rewards for reward candidates. see cyber/forgetting for pruning mechanisms. see cyber/self/parametrization for the twelve tunable parameters

discover all concepts