Demos

Twelve of the .su programs in examples/ are exercised end-to-end by the smoke test (the directory holds more — see Other examples below):

git clone https://github.com/EmmaLeonhart/Sutra
cd Sutra
python examples/_smoke_test.py

The smoke test compiles each .su source through the reference codegen path, executes the emitted Python, and compares the output to a hardcoded expected table. Seed and dimensions are fixed, so results are deterministic.

The smoke-tested programs

#	File	What it demonstrates
0	hello_world.su	Embed plus retrieve. Three candidate phrases, argmax_cosine, name lookup at the edge.
1	fuzzy_branching.su	Weighted-superposition conditional. Four program variants × four inputs. All branches contribute to a weighted sum; argmax_cosine commits at the end.
2	role_filler_record.su	Structured record as a flat vector. bundle(bind(role, filler), …); decode a field by unbind(role, record) followed by argmax_cosine.
3	classifier.su	Bundled-prototype classifier. Three classes, three examples each, bundle averages them into a per-class prototype.
4	analogy.su	Associative pair memory. Five (capital, country) pairs bundled into one vector; query a capital, recover the country.
5	knowledge_graph.su	Bundled triples with compositional query. bind(object, bind(subject, predicate)); lookup is unbind(predicate, unbind(subject, graph)).
6	predicate_lookup.su	Multi-object superposition. When a (subject, predicate) key has multiple objects, all members score above all non-members.
7	fuzzy_dispatch.su	N-way dispatch returning structured records. Each branch returns an (action, target) record; the winner is decoded with two unbind calls.
7b	content_addressed_read.su	NTM-style content-addressed read head. A colour key recalls the bound value by associative lookup.
8	nearest_phrase.su	20-phrase codebook, clean and noisy retrieval. Target plus 0.2·distractor still returns target.
9	sequence.su	Position-bound sequence encoder. A 5-token sequence is one vector; decode any position with unbind(pos_i, record). Two sequences compared by cosine.
10	semantic_faq.su	Semantic FAQ matcher. A paraphrased question matches the right canned answer by meaning — embed + argmax_cosine over a question codebook.
11	strings_and_formatting.su	String concat, interpolation, and int_to_string — text assembled entirely on the substrate.
12	fizzbuzz.su	The real FizzBuzz, 1..15. No if — a softmax-weighted select superposition picks each word; a loop threads the growing String accumulator and returns it as a value.
13	loop_forms.su	Every loop call form on small checkable loops: by-reference slot mutation, the expression form (int x = loop f(...)), multi-state tuple-destructure ((a, b) = loop g(...)), and a String accumulator by reference.

Loops use first-class declared functions. Start with do_while_adder.su for the minimal shape, then loop_forms.su for all three call forms and fizzbuzz.su for a loop doing real work. See the Loops page for the surface.

The big Unix tools, on a completely neural computer

The substrate can run the classic Unix utilities — not as a metaphor, but with the actual byte-processing done on the substrate and checked char-for-char against the real coreutils binaries. The machine is NTM-style (external addressable RAM and a persistent disk, not a plain RNN); the host does only I/O, the substrate does the transform. Fifteen tools are implemented and verified:

echo · cat · wc · head/tail · tr · rev/tac · cut · uniq · sort · grep (fixed string and regex) · sed · awk (field/pattern subset) · cat FILE/ls/cp/mv/rm · find

One primitive does most of the work: an exact codepoint indicator — 1 at the target character, a hard 0 at every other — which composes into counters (wc), gates (head/cut), codebook maps (tr), and comparators (uniq/sort). Regex (grep -E/sed/awk) runs a Thompson NFA as a genuine vector-valued substrate state, stepped by matrix multiplications. Source + per-tool self-tests: experiments/ntm_ram/ (see its README).

Other examples in the directory

examples/ contains additional programs that aren’t part of the smoke-test asserted-output table but exist as reference material:

• do_while_adder.su — minimal do_while declared-function loop.
• imperative_reversible.su — slot-based reversible state demo.
• classes_demo.su — empty-body class declarations (the MVP form, see Ontology).
• analogy_minilm.su, analogy_mxbai.su — substrate-sweep variants of analogy.su.
• protein_record.su — the same role-filler shape applied to ESM-2 protein-language-model embeddings.
• rotation_hashmap.su, rotation_book_catalog.su, rotation_record.su — rotation-binding library patterns.
• tutorial.su — companion source for the tutorials.
• wait_keyword_demo.su — the wait reserved-keyword shape.

These don’t have asserted outputs in the smoke test but parse and (where the codegen supports them) execute under the standard pipeline.

The primitives the smoke-tested demos exercise

Operation	What it computes	Demo files
embed(name)	embed a string through the substrate	all
bundle(a, b, …)	sum and L2-normalize	1, 2, 3, 4, 5, 6, 7, 9
bind(role, filler)	rotation binding: Q_role @ filler	1, 2, 5, 7, 9
unbind(role, record)	inverse rotation: Q_role^T @ record	2, 5, 6, 7, 9
similarity(a, b)	cosine similarity	1, 6, 9
select([scores], [options])	softmax-weighted superposition	1, 7
Scalar-vector multiply, vector add	weighted superposition	1, 7
map<vector, string> lookup	the single edge bridge from vector to host string	all

There is no if, while, for, or switch in any of these programs. Every conditional is a weighted sum across all options; the commitment to a discrete answer happens at the final cleanup step or map lookup. Loop primitives show up in do_while_adder.su and the dedicated test suite, not in the smoke-tested set.

Reading the source

The .su files are deliberately short (30–100 lines each including comments) and meant to be read front-to-back. Start with hello_world.su for the minimal shape; role_filler_record.su and knowledge_graph.su are the richest for understanding bind/unbind composition; do_while_adder.su is the smallest example of the loop surface and loop_forms.su walks every loop call form.

To inspect the generated Python for any demo:

sutrac --emit examples/knowledge_graph.su

The emitted module is self-contained — it instantiates a small _VSA runtime class and calls into it for every Sutra primitive.