One idea, three levels of effort
Write the notation, draw the graph, or say it in English and let a local model write the program — which then has to compile before it is allowed to run.
| Language | What it is | Reach for it when |
|---|---|---|
| Volt | A notation that parses into a graph | The job is “this, for all of those, then combine” |
| Ohm | Lua 5.4 where LaTeX maths is executable syntax | You need real control flow, or the full 49 operators |
| Bolt | English, translated by a local model | You would rather describe it than write it |
Every line names a thing, and the names are the edges. Building the graph is pure, so it is safe to do on every keystroke and on a program you have not read. Because it is a graph rather than a script, the independent work is concurrent, re-running recomputes only what changed, and the cost is countable before anything executes.
# Two companies, two questions. The whole program. sites = [apple, nvidia] topics = [latest earnings, layoffs] jobs = sites × topics found = search(jobs, k: 3) brief = summarize(found, words: 60) table(brief)
Lua 5.4, in which LaTeX mathematics is executable. $ ... $ is an expression;
$$ ... $$ holds statements — assignment, loops, conditionals, operators. Maths and
Lua share one set of variables; neither is a guest in the other.
-- Maths and Lua share one set of variables.
local A = {3, 1, 4, 1, 5, 9, 2, 6}
$$
|A| % length
\sum_{x \in A} x % sum over a list
\max_{x \in A} x
\frac{\sum_{x \in A} x}{|A|} % mean
sq := x \mapsto x^2 % a lambda
\forall x \in A : sq(x) % map over a list
\{1, \dots, 5\} % a range
$$
-- …and an AI operator is something you compose.
$$
answer := \Cite_{\Gamma} \circ \Ground_{\Gamma}(\Summarize_{\Psi}(pages); C)
$$
Each documented once, in the registry, and rendered everywhere from there. They compose with
∘, × and ∥ the way operators in a notation should.
| Family | For | Some of them |
|---|---|---|
\Psi | Reasoning | Ask, CoT, ToT, GoT, Plan, Reflect, Critique, Summarize |
\Gamma | Context | Search, Fetch, Retrieve, Rank, Filter, Compress, Ground, Cite |
\mathcal{M} | Memory | MemRead, MemWrite, MemSearch, MemSummarize, MemForget |
\Sigma | Aggregation | Join, Vote, Debate, Cross, Flatten, Sort, Pick, Zip |
\Lambda | Agents | ReAct, Verify, Fork, Chat |
Describe the job and a local model writes the program. The model is bounded rather than trusted: it produces source, and source can be checked.