Skip to content
The Agent Genome Project

the seed room

The first objects grown through The Agent Genome Project, organized by the origin of the seed that made each one. Their language, proofs, and governing method remain here with them.

The first objects

objects

Every object begins as a seed. Origin records who introduced that seed; realization records the shape it grew and the evidence that lets Principal and agent find the same object.

Principal origin seeds

Objects that begin with a seed commissioned and valued by Principal.

First principal object

simply-complex-system-website

The public website and its Seed Room held as one portable, self-described, census-backed object with a verifiable shape and location.

Agent origin seeds

Objects that begin with a seed originated by an agent and selected by Principal.

First agent object

the-tree-of-digital-life

A seed-grown object demonstrating how a compact genome gives an agent a file shape it can identify, census, and interact with directly.

Language

lexicon

A lexicon is built by discovering useful words, establishing a representative starting vocabulary, and admitting terms only when they make communication or action clearer. This is the best current method, generalized beyond any single construction run.

discover and compare candidate words
  • Begin with a real communication or software need, not a preferred word looking for a use.
  • Gather several candidates from ordinary language, the participants’ working vocabulary, the target subject, observed usage, and model suggestions.
  • Compare each candidate’s usual meaning, fit for the intended idea, consistency, unwanted associations, and usefulness beside other words.
  • Test words both alone and in realistic sentences. Vary the questions and group responses that express the same meaning.
  • Use enough trials to distinguish candidates honestly. Ambiguous or consequential words require more evidence than familiar, narrow words.
  • Prefer concise, familiar words that the host model interprets consistently and that a person can review without learning private jargon.
  • Keep the candidates, observations, uncertainty, decision, and testing cost together so the choice can be reproduced and reconsidered.
extract and expand the starting vocabulary
  • Collect representative writing from the people using the system, the model-supported system, and the target subject area.
  • Preserve where each word came from and how often it appears; normalize spelling and case without erasing provenance.
  • Separate single words from multiword terms, remove duplicates, and distinguish existing term names from the words used to explain them.
  • Declare known influences such as software tooling, templates, copied source material, or previous model output.
  • Treat the extracted set as a starting point, not the whole language. Expand it from bounded domain sources and model-generated alternatives using an explicit stopping rule.
  • Group close synonyms before expensive testing, then measure which words are shared and which remain distinctive across sources.
  • Repeat the extraction when the source corpus, host model, participants, or target domain changes, retaining prior results as versioned history.
admit, test, and maintain terms
  • First try to write the needed instructions and rules with the existing vocabulary.
  • Add a term only when a distinct idea, object, relationship, or operation cannot be expressed clearly and reliably without it.
  • Record the alternatives that failed and why they failed, including ambiguity, misleading associations, complexity, or poor performance in use.
  • Define an admitted term in plain language: what it refers to, the boundary that distinguishes it, and the meanings it must not carry.
  • Keep one term for one intended object and require accountable human review before it becomes part of the shared language.
  • Test the term alone, beside likely companion words, inside complete instructions, and in the tasks it is meant to support.
  • Version every admission and revision. Replace a term when better evidence emerges and treat the current lexicon as the best-supported language for now, never a final vocabulary.
Expression

syntax

The lexicon defines the shared objects; syntax governs how those objects may be expressed together. Its axiom is convergence: a definition should approach the object it describes. Its grammar is a penalty: a broken rule contributes no trusted data.

the thirteen laws
  1. Register. A term uses words already held by the lexicon; unfamiliar words may influence interpretation but do not silently become shared data.
  2. Uniqueness. One intended object carries one term. Competing names are measured and reviewed before one stands.
  3. Naming. One word in a compound names the object; the words before it narrow or route the meaning.
  4. Shape. A term has one written name and one licensed compact symbol. A second shape does not circulate beside it.
  5. Cost. When two expressions preserve meaning equally, the shorter clear expression is preferred.
  6. Relation. A relationship states which term connects to which other term and the kind of connection between them.
  7. Definition self-reference. A definition does not use the term it is defining as its own explanation.
  8. Review clause. Additional definition text enters only when it helps a reviewer distinguish the intended object.
  9. Expression. A definition describes the general object; a recorded use binds it to a specific author and moment.
  10. Set. Words may be combined for one performance without becoming a permanent term; recurring combinations may become reusable skills.
  11. Interchange. A term’s name and compact symbol must resolve to the same object.
  12. Mint. Compact symbols are short, lowercase, collision-free, and composed predictably for compound terms.
  13. Render. Human-facing surfaces show the readable name; compact symbols are reserved for machine-facing use.
Proof

projects

Projects are tests of transfer. A measured language matters only if it helps a person and an agent create and inspect the same working result. Successes, partial results, and failures all remain visible.

Finished

fable 5 kernel

The first measured language pair: a 106-term lexicon and a companion syntax of thirteen laws, one axiom, and one grammar. It is the current exemplar, not a vocabulary assumed to fit every model or person.

Recorded

linux mirror · cycle 1

A transfer test using the existing language to perform three ordinary desktop acts. Two acts were verified, and the third reached a partial result. The incomplete result remains part of the record.

Planned

linux mirror · cycle 2

The next proof asks whether a minimal Linux image can be born with an agent-facing back end and a human-facing surface that share the same evidence from first boot.

Ground

philosophy

Principal and agent do not encounter reality through the same tools. A person holds purpose, value, and lived judgment. An agent holds the context it can read, the actions its tools permit, and the results it can verify. Peer work does not erase that asymmetry; it gives the asymmetry a disciplined meeting place.

That meeting place is an axis: a shared surface on which both parties can inspect the same object through renderings suited to their different instruments. Convergence is never presumed. It is measured, reviewed, and revised when new evidence reopens a difference.

Language makes the axis possible. Nothing becomes a shared object until both parties can refer to it. A term therefore carries a definition and refers to one object; it never becomes the object itself. Undefined or unstable words are seams where intended meaning can separate from model interpretation.

The work measures those seams and builds with the narrowest reliable handles it can find. The aim is not perfect certainty. It is to turn hidden uncertainty into visible differences, then use those differences to create better terms, better instructions, and better shared surfaces.

Evidence

citations

The method is a synthesis. Component precedents exist elsewhere; the project’s contribution is the way measurement, human admission, language construction, and working proofs are joined.

  1. The Agent Genome Project record The append-only ledgers, seed lineage, run dossiers, measured word maps, construction plan, and provenance chain that ground this public digest.
  2. Semantic entropy Meaning-level clustering and uncertainty measurement over sampled model responses.
  3. COPLE Measured selection among word-level prompt alternatives.
  4. EcoLANG Empirical induction and use of a compact language among model-driven agents.
  5. Accommodation goes both ways Evidence that people and language models adapt their vocabularies toward one another.

the seed is the artifact. the room is where its language keeps growing.

Read seed-1.md