soft3/tru/docs/explanation/metrics.md

Metrics: Attention, Focus, and Gravity

How the cybergraph measures what matters -- from individual attention to collective focus to structural gravity.


Attention vs Focus

Attention and focus are distinct but related.

Attention is individual. It is how much a neuron projects onto a target particle or axon. A neuron directs attention by creating cyberlinks and allocating conviction. Attention is the cause -- the input signal from each participant.

Focus is collective. It is the aggregated result of all neurons' attention, processed by the tri-kernel into a single probability distribution over all particles. Focus is the effect -- the output of the collective computation.

Individual neurons direct attention. The cybergraph aggregates all attention into focus. Every particle's focus score $\phi^*_i$ reflects the contributions of all neurons, weighted by their tokens and filtered by the epistemic layer.

Attention is produced by two mechanisms: will (broad auto-distribution across all cyberlinks) and fine-tuning (manual per-target weight adjustment). Both produce the same thing at the receiving end -- attention at the target particle.


The Transformer Attention Connection

The transformer attention mechanism computes, for each position in the context, a weighted average of all other positions:

$$\text{Attn}(Q, K, V) = \text{softmax}\!\left(\frac{QK^\top}{\sqrt{d}}\right)V$$

Three projections: queries $Q = XW_Q$ ask "what am I looking for?", keys $K = XW_K$ announce "what do I contain?", values $V = XW_V$ provide "what information do I carry?". The dot product $QK^\top$ scores compatibility. The softmax converts scores to a probability distribution -- the Boltzmann distribution with temperature $\sqrt{d}$.

The softmax is the same operation as the LMSR price function and the tri-kernel diffusion step. All three are exponentiated scores normalized to sum to 1.

Transformer attention is one step of the tri-kernel diffusion operator $D$ applied to the current context window. Probability mass flows from each query position toward compatible key positions -- exactly the random walk dynamics that the tri-kernel uses to compute focus over the cybergraph.

Deep Equilibrium Models showed that iterating a transformer layer to convergence reaches the same fixed point as the tri-kernel: $\phi^*$ restricted to the context window. $L$ layers of attention = $L$ steps of diffusion toward that fixed point.


Attention as Bayesian Query

Attention answers: given my current state (query), what posterior weight should I assign to each position (key)?

The softmax is the posterior $P(\text{position } j \mid \text{query } i)$ under a uniform prior and an exponential likelihood $\exp(q_i \cdot k_j / \sqrt{d})$. The query-key product is the log-likelihood under this model. The softmax is the Bayes-normalized posterior. Attention is Bayesian inference over the context.


Multi-Head and Dialects

Through multi-head attention, different heads learn different relation types. Head $h$ with projection $W_Q^{(h)}, W_K^{(h)}$ captures one dialect -- one pattern of connectivity in the cybergraph.

The graph-native-transformer derivation proves that the minimum number of heads equals the number of distinct dialect types in the graph. Each head specializes in one kind of relationship: "is-a", "contradicts", "extends", "cites". The graph's link topology determines how many attention heads the compiled transformer needs.


Gravity

Gravity is a node-level metric. Like physical gravity, it is a property of the node itself -- a massive body warps space around it and attracts everything, regardless of what is nearby.

$$G_i = \phi^*_i \cdot \sum_{j \neq i} \frac{\phi^*_j}{d(i,j)^2}$$

where $\phi^*_i$ is the node's own focus probability, $\phi^*_j$ are focus probabilities of all other nodes, and $d(i,j)$ is the shortest path length in the cyberlink graph.

A node's gravity is its focus mass multiplied by the total attention field it experiences from the rest of the graph. High-focus node surrounded by other high-focus nodes = enormous gravity. High-focus node on the periphery = less gravity despite its own mass.

Physical Analogy

A planet curves spacetime by its mass alone. The gravitational potential of a body in a field of other masses:

$$\Phi_i = m_i \cdot \sum_{j} \frac{m_j}{r_{ij}^2}$$

physics knowledge graph
mass $m$ focus probability $\phi^*$
distance $r$ graph distance $d(i,j)$
gravitational potential $\Phi$ node gravity $G_i$

The node does not choose what to attract. It simply has mass (focus), and everything within graph distance falls toward it proportionally.

Gravity Spectrum

gravity profile meaning
high high $\phi^*$, surrounded by high-$\phi^*$ neighbors core attractor -- holds the graph together
medium moderate $\phi^*$, or high $\phi^*$ but few neighbors regional hub -- local structure anchor
low low $\phi^*$, or isolated from high-$\phi^*$ nodes peripheral -- structurally weightless

Applications

Skeleton extraction: nodes with the highest gravity form the structural skeleton of the knowledge graph. Remove them and the graph fragments.

Peripheral detection: nodes with high focus but low gravity are isolated attractors -- they have mass but sit far from other massive nodes. Connecting them to the core would dramatically restructure the graph.

Cohesion measurement: total graph gravity $G_{\text{total}} = \sum G_i$ measures how tightly the knowledge core is packed. A graph with high total gravity has its attention concentrated in a dense, interconnected core. Low total gravity means focus is scattered.

Pairwise Force

The force between any two specific nodes:

$$F_{ij} = \frac{\phi^*_i \cdot \phi^*_j}{d(i,j)^2}$$

The highest $F_{ij}$ pairs are the structural bonds of the graph. Pairs with high $\phi^*_i \cdot \phi^*_j$ but large $d(i,j)$ are the most valuable missing cyberlinks -- creating them collapses distance and unlocks attention flow.


How Ranking Works in Practice

Cyberank is focus materialized as a per-particle score. It is the probability of being observed by a randomly walking neuron -- the fixed point of the tri-kernel.

The tru computes cyberank every block. The score feeds into:

  • Karma: accumulated track record of a neuron's contributions
  • Syntropy: the organizational quality of the focus distribution ($J = D_{\text{KL}}(\phi^* \| u)$ -- how far attention deviates from uniform noise)
  • Inference: the probability distribution that drives query responses and autoregressive generation
  • Sorting in cyb: the user-facing ordering of search results

Luminosity = size x $\phi^*$ -- what a node radiates (knowledge output). Gravity = $\phi^*$ x $\sum(\phi^*_j/d^2)$ -- how strongly a node attracts (structural pull). A healthy graph needs both: high-luminosity nodes that radiate knowledge, with high-gravity nodes that hold the structure together.


See attention for allocation strategies. See focus flow computation for the global process. See tri-kernel for the diffusion connection. See gravity for the full metric specification.

Graph