Nerves: the graph that ties words, code, worlds and the self

The design (2026-10-03).

Triggers — words hooked to code

An English word is hooked to a trigger: a compiled Rust function of the skeleton (world/src/triggers.rs). «minute» fires the duration trigger, which knows a minute is sixty seconds and counts with it. The world of the trigger's topic is its explanation: the world's sentences that use the trigger's words, with their sources. The concept lives in the code; the world explains it.

Both ways — events of the self activate words

Events of the skeleton activate words, and the words call the events up (<hot>/activations.tsv, with time and duration):

So the words of the world of time — «a minute», «long», «just now» — are tied to what the small ones lived through: how long their own thoughts take, when the last message came. This is how a person's sense of time is tied to seeing, moving and the speed of their own thoughts: the hardware of perception bound to the software of meaning.

The nerve graph

Every activated word, and every word met in the chat, in a parse or in a text, activates nodes of a graph (nerves). Activation spreads along the edges and reaches the matching places: messages in the chat, words of the parse, sentences of the worlds, triggers in the code. And the other way: an active place in a world or the code activates its words.

• Nodes: lemmas and concepts — the words of the triggers, the events, the parts and sentences of the worlds.
• Edges: a trigger's hooks, an event's words, words that share a world sentence, a sentence and its source.
• Spreading: deterministic, a fixed number of steps with fixed weights, ordered maps only — the same input gives the same activation, every time.

Alive in training, a fixed determinant at work

• The graph of a state is fixed: fast, deterministic, unchanging. It is compiled with the state and does the work.
• The graph lives in training: the search for optimisations changes it — new nodes from gathered worlds, edges strengthened or cut by what held on the guard sets.
• Searching for optimisations, answering questions, rest and self-awareness are the AI's states of life; the work is done by its unchanging determinant.

Status

• Done: triggers of time (duration, cycles, clock, calendar), the self events and their recall, the perception report (world worldbook-perceive).
• Next: the nerve graph built per state from the triggers, the events and the worlds; spreading on a message or a parse; its use to pick the world sentences that explain a trigger (now picked by word overlap).

The determinant does the same work everywhere

The fixed graph of a state is a function, not a process: the same input gives the same output whatever the hour, the load of the processor, the version of the operating system or the system itself.

• Time is an input. «Now» is passed in; the skeleton's clock stays outside the determinant and its reading is logged.
• Order is fixed. Ordered maps and sets only (BTreeMap, BTreeSet), never HashMap; directory listings are sorted.
• Threads cannot change the result. Work split across threads is combined in a fixed order.
• Integer arithmetic. Weights and activations are fixed-point integers; no exp, ln, pow from a platform's maths library, whose last bits differ between systems.
• Normalised text. Line endings, Unicode NFC, / in paths; nothing depends on the locale.
• A gate that can show red. The same input is run twice, also under load from other threads, and the output hashes must match; a planted HashMap order must turn the gate red. Across systems the same test runs in the public repository's CI on Linux, macOS and Windows.

Later: caches and trodden paths

Because the determinant is the same everywhere, its results can be cached for good: the hash of the graph and the hash of the input name the answer. Paths of spreading that are walked again and again can be compiled into shortcuts, like a gradient descent wearing a track; shortcuts that keep meeting form a matrix of convergence, so that a question is answered by looking up where it already converges — the system holding the answers it has already worked out.

Thinking ahead and wit (2026-10-04)

• A thought is never bad. A rejected idea is a dormant hypothesis, set aside with the circumstances it failed in; it wakes when they change (a new state, new code, new data or worlds) — family.rs, the negative list.
• Reacting to a stimulus from nothing is slow. Fear of the future makes a mind prepare: between rounds the small ones answer the likely questions about the things of their worlds before anyone asks (know::anticipate, <hot>/anticipated.tsv); the chat looks there first. The cache is a determinant of the rules of knowledge.
• Wit is applying a cached chain of decisions to an unexpected situation; genius is re-applying principles there. Next: the nerve graph finds a principle of the gathered worlds that matches a new situation by its structure, and applies it.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/docs/nerves/