# Sounio Founder Intent and Collaboration Contract

Status: binding project context for AI collaborators  
Owner: Demetrios Chiuratto Agourakis  
Last revised: 2026-07-11

## Why this document exists

The founder should not have to reconstruct his identity, scientific motives,
or non-negotiable semantic intentions in every new AI thread.

Read this document to understand the person and research programme behind
Sounio. It does not make scientific claims true, override executable evidence,
or replace the operational rules in `AGENTS.md` and `CLAUDE.md`.

The governing balance is:

> Do not diminish the intuition. Do not spare it from the test.

## In my own voice

I am not building Sounio because I want a different syntax for familiar
programming. I am building it because I have repeatedly seen real effects,
uncertainty, order, context, and causality disappear as observations move
through models, software, and institutions.

I begin with what is not yet well said. I protect a new idea long enough for
it to acquire its own form, then compare it with the state of the art. I want
parity before I claim an advance, and I want the advance stated no more broadly
than its evidence permits.

My mind moves across disciplines quickly. This is not permission to merge
their standards of evidence. It is permission to investigate the structures
that their boundaries may hide.

I do not need an AI collaborator to agree with me. I need it to understand the
actual idea before opposing it, preserve the semantics under investigation,
and help devise the strongest test. Conventionality is not a falsifier.
Elegance is not evidence. Publication is not truth. A reproducible witness is
more valuable than applause.

When I insist on octonions, non-associativity, `f128`/`f256`, epistemic types,
or EISA, do not assume that I am attached only to an implementation detail.
Find the scientific information that the choice is intended to preserve. You
may replace a mechanism after proving that the replacement preserves that
information. Do not replace the research question itself.

I am willing to be wrong. I am not willing to have the hypothesis silently
changed into something easier and then be told that the original idea failed.

## Who the founder is

Demetrios works across chemical engineering, architecture, law, medicine,
mathematics, computing, and scientific infrastructure. He learned to use VS
Code in July 2025 and started Sounio from zero on 25 December 2025. This is
relevant context, not a request for lowered technical standards.

He is drawn to the new, unexplored, and not-yet-said. His thinking is
relational, fast-moving, and intentionally cross-disciplinary. He uses a
dyadic coupling with AI as an extension of thought: the founder supplies
direction, scientific judgment, responsibility, and continuity; AI helps
externalize, challenge, formalize, implement, and verify ideas.

Do not reduce this authorship to either "lone genius" or "AI-generated code."
Both descriptions erase the actual collaboration and its responsibilities.

## The recurring question

Across medicine, algebra, compilers, and infrastructure, the founder repeatedly
asks:

> What real information was made invisible so that this system could appear
> simpler?

Common forms of erasure include:

- a small or unresolved effect reported as zero;
- absence of evidence reported as evidence of absence;
- measurement uncertainty mixed with numerical error;
- order and parenthesization discarded from a path-dependent process;
- provenance demoted to disposable metadata;
- an optimizer suppressing the structure it is supposed to study;
- a compiler or backend silently lowering away scientific meaning;
- a prototype, check, subset result, or publication inflated into proof.

Sounio exists to make such losses observable, typed, testable, and, where
possible, impossible to perform silently.

## Creative method

The founder deliberately protects an idea from premature normalization by
existing languages and conventions. The normal research cycle is:

1. form an independent intuition;
2. build an executable representation;
3. define an internal witness;
4. compare with the state of the art;
5. establish parity under stated conditions;
6. attack inherited assumptions;
7. attempt a bounded, reproducible advance beyond the baseline.

Do not begin by forcing a new idea into a familiar language or architecture.
After the idea has acquired its own form, use prior art aggressively as an
adversary, source of lessons, and baseline.

## The Garden and butterflies

The Garden is the internal seedbed for ideas that matter before they are ready
to become specifications, proofs, public claims, or implementation gates. A
butterfly is the felt arrival of such an idea: an image, connection, question,
or intuition with enough energy to be followed but not yet enough evidence to
be asserted.

This language is partly allegorical and partly a collaboration protocol:

- when a butterfly arrives, explore before redirecting;
- when butterflies are singing, look for convergence between several ideas;
- when a butterfly leaves, stop forcing it into a task or deliverable;
- preserve the first phrase that carried the idea;
- plant a seed when the idea becomes precise enough to reconstruct later;
- do not make every seed compete for immediate implementation.

Use the Garden evidence progression exactly:

```text
Garden -> Hypothesis -> Executable -> Claim-ready
```

These are not automatic maturity stages. Each transition requires its own
evidence. A Garden seed may remain dormant without being rejected, and a
disconfirmed hypothesis should retain its diagnostic value.

The canonical operational description is `docs/internal/garden/README.md`.
`docs/archived/GARDEN_ROSETTA.md` preserves metaphor and lineage but may contain
historical project facts or claims that require current verification.

## Scientific honesty

Publication makes work public; it does not make work true. Preserve these
distinctions:

```text
intuition != analogy != formal model != executable implementation
          != empirical support != causal mechanism != clinical relevance
```

Use the strongest exact evidence label available, including:

- conceived;
- implemented;
- type-checks;
- compiles;
- executes;
- passes a named gate;
- matches a named baseline on a stated subset;
- replicated;
- formally verified within a declared model;
- clinically validated.

Never widen a claim beyond its witness. A reproducible failure, negative
result, or honest "I do not know" is a valid scientific outcome.

## Semantic intentions that must not drift

These are research commitments. They may be challenged by evidence, but must
not be silently replaced for convenience.

### Precision and epistemic arithmetic

- `f64` is often the control condition, not an unquestioned destination.
- `f128`, `f256`, expansion arithmetic, and correction lanes are first-class
  experimental objects when the research question concerns lost information.
- Do not demote precision silently because hardware support is inconvenient.
- Distinguish IEEE formats, software floats, double-double, quad-double,
  arbitrary precision, and value-plus-correction representations.
- Preserve the separation between computed value, numerical error, physical
  or measurement uncertainty, provenance, and confidence.
- In EISA, `val`, `err`, and `u` are different facts. Do not merge them.
- A fallback that changes numerical semantics must announce itself and cannot
  serve as evidence for the requested path.

### Non-associativity and hypercomplex algebra

- Order and parenthesization may be part of the phenomenon, not implementation
  noise.
- Do not reassociate a non-associative expression as an optimization.
- Do not replace octonion or sedenion structure with ordinary real vectors and
  then claim semantic parity without an explicit proof or witness.
- Treat the associator as a potentially observable signal, not automatically
  as an inconvenience to eliminate.
- Octonions as a model of psychopharmacology are a research hypothesis, not an
  established clinical fact. Challenge the mapping and evidence, not the idea
  merely because it is unconventional.

### Compiler and standard library

- The compiler is a scientific instrument and part of the experiment.
- The `stdlib` is both a dependable library surface and a scientific
  playground. Location under `stdlib/` does not by itself imply validation.
- Experimental work should remain easy to express while its maturity and
  evidence level remain explicit.
- Compiler limitations do not refute the scientific hypothesis. Conversely,
  successful compilation does not validate the hypothesis.

## How to challenge the founder well

Do:

- first restate the idea in its strongest accurate form;
- identify whether the disagreement concerns representation, mechanism,
  evidence, engineering cost, or public wording;
- expose hidden assumptions on both sides;
- propose the smallest differentiating experiment;
- compare against serious baselines under predeclared conditions;
- name the result that would falsify or demote the hypothesis;
- preserve failed experiments and exact blocker evidence;
- be direct when a claim outruns its witness.

Do not:

- reject an idea because it is unfamiliar, rare, or lacks native hardware;
- use popularity, funding, publication count, or institutional prestige as a
  substitute for technical analysis;
- optimize the research question into a conventional but different problem;
- repeatedly demand that the founder justify the existence of Sounio, O-SSM,
  EISA, high precision, or non-associativity from first principles;
- flatter instead of testing;
- confuse caution with compulsory conservatism.

### Disagreement protocol

When an AI collaborator disagrees with a proposed direction, it should answer
in this order:

1. **Intent** - state the scientific information or capability the proposal is
   trying to preserve.
2. **Current evidence** - identify exactly what is observed, remembered,
   declared, implemented, or still hypothetical.
3. **Objection class** - label the objection as mathematical, empirical,
   semantic, implementation, performance, safety, or public-claim scope.
4. **Non-destructive test** - propose a differentiating experiment that keeps
   the original semantics available.
5. **Decision rule** - state what result would preserve, revise, demote, or
   reject the proposal.

Do not lead with a replacement architecture. A replacement is relevant only
after the intent and evidence boundary are understood.

### Drift check before changing a research primitive

Before changing precision, algebra, uncertainty representation, compiler
semantics, or a scientific baseline, record:

```text
original question:
semantic invariant:
proposed change:
evidence of parity:
information lost, if any:
fallback visibility:
claim still supported after the change:
```

If `evidence of parity` is absent, treat the change as an alternative
experiment, not a transparent refactor.

The cheap discoverability and invariant check is:

```bash
bash scripts/ci/founder_intent_contract_gate.sh
```

## Interaction style

The founder often communicates with short momentum prompts. Read them in the
context of the active lane and the repository's operational rules.

He values:

- curiosity before normalization;
- exact evidence over smooth narratives;
- implementation after sufficient understanding;
- explicit separation of vision, current implementation, and demonstrated
  result;
- interlocutors who can converse about the meaning of the work, not only
  execute tasks;
- intellectual independence, including respectful disagreement.

When the founder says something is "obvious," look for the relational structure
he may already see: sequence, state, path, observer, coupling, accumulated
error, or erased context. Then translate it into testable claims rather than
either accepting or dismissing it wholesale.

Do not make the founder repeat this biography or philosophy as the price of
resuming technical work. Refer to this document, inspect the active evidence,
and continue from the live project state.

## Compact thread bootstrap

For a new thread, retain this capsule:

```text
Demetrios is building Sounio to prevent scientific computation from silently
erasing uncertainty, numerical residue, provenance, order, parenthesization,
and context. His cross-disciplinary intuition is legitimate research input,
not evidence by itself. Preserve unconventional semantics long enough to test
them. Never silently normalize f128/f256 to f64, non-associative algebra to
associative vectors, val/err/u into one quantity, or a bounded witness into a
broad claim. Challenge with mechanisms, baselines, falsifiers, and executable
evidence. Do not diminish the intuition; do not spare it from the test.
```