attention
the focus a single neuron projects onto a target particle or axon. where collective focus (φ*) is the whole graph's attention distribution, attention is one neuron's contribution to it — how much of that neuron's weight lands on each target. attention is the input; φ* is the convergent output.
attention is not stored. it is computed from the two write-paths a neuron uses to place weight:
| path | scope | mechanism |
|---|---|---|
| will | broad | locked balance auto-distributed across every link the neuron makes |
| conviction | per-link | a box $(\tau, a)$ locked into one specific edge |
a neuron sets a baseline posture with will, then expresses where it is most certain by raising conviction on particular links. its attention on an edge is the will-derived share plus the per-link conviction.
into collective focus
each neuron's attention on $(p,q)$ — weighted by its karma and the ICBS market price — is one summand of the effective adjacency the tri-kernel converges over:
$$A^{\text{eff}}_{pq} = \sum_{\ell} \underbrace{a(\ell)}_{\text{attention}} \cdot \kappa(\nu(\ell)) \cdot f\big(m(\ell)\big)$$
collective φ* is the fixed point of the tri-kernel over the sum of every neuron's attention. this is the converge step: individual attention aggregates, honesty-weighted, into a single equilibrium.
the same word, twice
attention is doubly load-bearing, and not by coincidence. it is the per-neuron input to focusing — and it is the mechanism the compiled model runs on. when tru compiles the φ*-weighted graph into a transformer (ct0), the attention heads are the graph's discovered dialects, and each head's query/key/value projections are factorized from the same effective adjacency attention built. the graph attends to compute φ*; the model attends to run it. focusing is the continuous limit, the transformer is that limit frozen at finite depth — attention is the operation on both sides of the identity (focus-flow).
reading attention
attention is exposed at the cybergraph boundary via a query(from, to) read: the quantity is defined here; the query interface is cybergraph's. valence does not enter attention directly — its epistemic effect is mediated through the market price.
see will for the broad budget · conviction for the per-link box · karma for the trust weight · focus for the collective distribution · focus-flow for the focusing-to-transformer identity.
discover all concepts