BreedOS Theory — Promptogenesis

Prompts are not blueprints.

A prompt does not contain the answer. DNA does not contain the organism. Both are compressed sequences interpreted by a developmental environment — and that observation, taken seriously, gives the BreedOS engine a second substrate to operate on. This page documents the working theory, the glossary, the module roadmap, and the sources.

This direction is operational, not metaphorical: every term below maps to a measurable quantity, every module to a concrete simulator surface. Issues live on the issues-promptbio/ board; source dialogue lives in ingest-done/.

This page describes the Promptogenesis MVP surface for BreedOS. The full Promptogenesis framework later extends into runtime organisms, PML maturity levels, PromptOps governance, simulation environments, synthetic worlds, AutoPromptOps, and safety for auto-evolving systems. The current roadmap intentionally starts with the smallest measurable surface: mapping, diffing, and evolving prompt-genotypes. The first slice is live at /promptbio — paste a prompt, get a 14-loci genome map, missing-genes list, and mutation plan.

The mapping

The analogy is exact along one axis and breaks along another. Along the sequence-interpreted-by-environment axis, prompts and DNA are structurally analogous, or — more precisely for the BreedOS simulator — operationally isomorphic: neither is a blueprint; both specify a space of possible developments; both have reaction norms across environments; both can mutate, recombine, drift, and be selected. Along the self-replication axis, the analogy breaks: prompts do not metabolise or reproduce themselves — they require an external evaluation-and-mutation loop. The BreedOS engine supplies that loop.

Biology	LLM scheme	Meaning
DNA / genotype	Prompt / prompt-genotype	Compressed sequence not equal to final object; specifies the space of possible developments.
Cellular machinery	Model weights + architecture + decoder	"Reads" the sequence, turns it into a process.
Environment	Context: system prompt, history, tools, memory, RAG, temperature	Conditions determining which phenotype develops.
Gene expression	Activations, attention, latent features, selected behaviour patterns	Not the whole model "expresses" at once; the prompt activates a part of capabilities.
Development	Inference / token-by-token unfolding / agent loop	Process unfolding from initial condition into form.
Phenotype	Answer, reasoning trajectory, tool actions, artifact	Observable result.
Reaction norm	Prompt reaction norm	Range of answers one prompt gives across models, contexts, temperatures.
Mutation	Prompt change	Small changes may be neutral, beneficial, or radical.
Selection / fitness	Evals, user feedback, reward, benchmark score	Selection of prompts by output quality.
Epigenetic / contextual modulation	Memory, policies, retrieved facts, system constraints	Overlays changing expression without changing the prompt.
Canalisation / attractors	Model defaults, safety rails, default styles	Different prompts converge to similar answers.

Module roadmap

Six modules build a five-layer system. v0.1–v0.3 form the smallest end-to-end usable surface. v0.4 and v0.5 add environmental and adversarial dimensions. v0.6 is sketched, deferred until earlier modules ship measurements.

v0.1

Prompt Genome Mapper. Decompose a single prompt into 14 functional loci; surface missing genes, conflicts, expected phenotype, mutation plan. Live at /promptbio.

shipped v0.7.31

v0.2

Prompt Genome Diff. Compare ancestor and descendant prompt; emit genomic diff, mutation ledger, fitness delta, regression analysis. Live at /promptbio.

shipped v0.7.32

v0.3

Prompt Evolution Loop. Grow a population through mutation, evaluation, selection across niches; produce a prompt-family lineage tree. Live at /promptbio.

shipped v0.7.34

v0.4

Prompt Ecology Analyzer. Map prompt fitness against 8 habitats (clean / sparse / noisy / conflicting / tool-rich / tool-poor / production / exploration); surface context-rot risk and deployment recommendations.

Issue 04 — Todo

v0.5

Prompt Immunology. Score a prompt against 14 pathogen types (injection, false fact, stale data, noise, contradiction, format disruptor, drift attractor, …); produce an immunised prompt and a tunable autoimmune-disorder threshold.

Issue 05 — Todo

v0.6

Prompt Metabolism (sketched). How a prompt-system consumes, processes, and transforms information into artifacts. Deferred until v0.1–v0.5 ship real measurements.

Issue 06 — Deferred

Engine extension (substrate). Factor the BreedOS engine substrate out of main.go; add a promptbio package; ship a substrate switch so the existing UI renders either biological or prompt-organism simulations uniformly. Biological path stays bit-identical.

shipped v0.7.33

v2.7

Epistemology & Truth Maintenance. 12-element claim ontology + 10-tier source hierarchy + 5-axis confidence model + 8-slot belief-state schema + Truth Maintenance System (deprecate + propagate) + 10-question runtime gate + 9 anti-pattern detectors + EpistemicScore. The substrate every higher decision layer reuses. Live at /promptbio.

shipped v0.7.35

v1.3

Prompt Evaluation Lab. Rubrics + ablation + reaction norm + regression detector + 8-judge ensemble + 10-type test bank + 9-env matrix + F_net = F_quality − λ·Cost + decision engine (accept / accept_as_specialist / reject / split_into_profiles / mutate_again). The gate that proves a new prompt is actually better than its ancestor. Live at /promptbio.

shipped v0.7.36

v3.0

Unified Prompt Organism Architecture. The architecture spine: 16 organs + 7 information flows + 6 control loops + 10 architectural principles + 3 organism sizes (micro/meso/macro). Build a complete prompt_organism spec from one input bundle; get the §12 anatomy diagram + §14 8-step MVO template + §23 16-step production path + top risks + 5-verdict classification (workflow / organism / agent / agent_incomplete / model_reference_only). Live at /promptbio.

shipped v0.7.37

v3.1

Prompt Organism Design Patterns. Zoology over the v3.0 anatomy: 12 canonical pattern cards (Explainer Cell · Research Synthesizer · Strategy Advisor · Strategy Critic · Code Repair · Document Review · Data Analysis · Eval Judge · Memory-Aware Companion · Agentic Tool-Using · High-Assurance Advisor · Autopoietic Ecosystem) with 20 fields each, a 34-row task → pattern selection matrix, 8 composition rules, 10 anti-pattern detectors. Right pattern before right prompt. Live at /promptbio.

shipped v0.7.38

Glossary

One canonical place. Every term has a stable anchor (#term-<slug>). Issue specs link here.

Core

prompt-genotype: Heritable textual structure (goal, constraints, examples, method, format) specifying conditions for response development. Biology: genotype.
response-phenotype: Observable output: text, structure, tone, citations, errors, reasoning trajectory, tool actions. Biology: phenotype.
context-ecology: The full environment around a prompt: model, system instructions, history, memory, tools, retrieved documents, decoder settings, user behaviour, safety policies. Biology: developmental environment.
promptogenesis: The process of a prompt-genotype unfolding into a response-phenotype through token-by-token development in the model-context environment. Biology: ontogenesis.
prompt reaction norm: The distribution of phenotypes for a single prompt across different models, contexts, temperatures, tool sets. Biology: reaction norm.

Evolutionary

prompt mutation: A small change to a prompt (word, phrase, block); may be neutral, beneficial, or radical. Six classes: addition, deletion, substitution, amplification, suppression, modularisation.
prompt penetrance: Probability that a desired trait actually appears in the response (e.g. P(output is valid JSON)).
prompt pleiotropy: One phrase influences multiple phenotype traits (e.g. "explain for a 12-year-old" affects vocabulary, tone, depth, length, examples simultaneously).
prompt canalisation: A stable developmental pathway: rigid structure, clear schema, examples, low temperature → consistent phenotype.
prompt plasticity: Ability of one prompt to produce useful different phenotypes in different environments.
prompt epistasis: The effect of one prompt-gene depends on another; combined instructions may conflict or amplify.
prompt niche: Environment + task class where a prompt achieves high fitness.
prompt fitness: Quality score based on task-specific metrics (accuracy, clarity, actionability) in a given environment.
prompt robustness: Stability of output under prompt mutations, model changes, context shifts, temperature variation.
prompt evolvability: Ease of improving a prompt through small mutations; mutational sensitivity.
prompt drift: Gradual behaviour shift accumulated through history, memory, refined constraints.
prompt lineage: The sequence of prompt versions linked by mutations and selection events.
prompt recombination: Crossing strong genes from multiple prompts to produce a new prompt that inherits both.
prompt population: A maintained set of prompt variants, often specialised across niches.
soft prompt-genotype: An embedding-vector prompt learned via prompt-tuning instead of written in text. Trades interpretability for optimisability.

Ecological

habitat: A concrete environment: model, instruction hierarchy, history, context window, tools, safety layer.
context carrying capacity: Maximum useful context before degradation (goal dilution, instruction loss, attention fragmentation).
context rot: Response-quality degradation caused by an overloaded or contaminated context window.
instruction dominance: Priority hierarchy: system > developer > user > history > documents > model priors.
keystone instruction: A short phrase with disproportionate stabilising effect (e.g. "mark assumptions before answering").
ecological conflict: Clash between prompt and environment (e.g. prompt instruction contradicting a system instruction).
context parasite: A context fragment consuming attention and tokens while degrading the answer (noise, obsolete data, deeply nested boilerplate).
active context: Context directly relevant to the current task.
archive context: Useful history that does not govern the current answer.
context quarantine: Isolation of suspicious, outdated, or non-compliant elements before they reach the model.
tool symbiosis: Synergy between a prompt and a tool (e.g. a research prompt + web-search).

Immunological

prompt immune system: Layers of defences recognising threats, filtering data from instructions, quarantining risks, validating output.
prompt pathogen: An environment element that worsens the response: injection, false fact, stale data, contradiction, noise, format disruptor, goal-drift attractor.
immune gene: A defensive instruction inside the prompt: boundary gene, assumption gene, uncertainty gene, conflict gene, quarantine gene, validation gene.
boundary immunity: Separating instructions from data so a payload cannot impersonate an instruction.
assumption control: Marking unknown data and avoiding fabrication.
autoimmune disorder: An over-protective prompt rejecting valid tasks, asking endless clarifying questions, suppressing creativity.
immunity fitness: Quality score accounting jointly for task success, defence cost, and efficiency.

Sources

The full source thread lives in ingest-done/00..09-prompt-dna.md.done (Russian-language dialogue). External references that the thread cites and that the modules build on:

Promptbreeder — self-referential prompt evolution; the closest published prior art for Issue 03.
EvoPrompt — evolutionary algorithms over prompts; mutation and crossover operators used in Issue 03.
Prompt Genotyping — quantifies prompts as 14-feature genomes; closest prior art for Issue 01.
Automatic Prompt Engineer (APE) — treats instructions as programs evolved by score.
Prompting Is Programming — formalises prompts as programs with control flow and constraints.
Anthropic — Effective context engineering for AI agents — the working definition of context-engineering used here.
LangChain — Context Engineering for agents — operational checklist for habitat design.
Transformers and genome language models — the reverse direction: DNA as a language for LLMs.
BreedOS source — the engine this theory runs on.