A model ecology separates doing from measuring
Runtime specialists produce useful work. Independent evaluators produce trusted evidence. The lineage DAG records where every descendant came from. The resource ledger keeps the ecology frugal. The evolution controller proposes, compares, promotes, archives, or no-ops. This separation makes model improvement easier to understand and easier to adopt.
Control-plane components
| Component | Positive role |
|---|---|
| Runtime specialists | Focused models, adapters, cascades, or tool pipelines that do the work they are fit for. |
| Router | Chooses the right specialist or stack for the request contract and budget. |
| Independent evaluators | Produce trusted measurement outside the candidate population. |
| Fitness vector | Keeps utility, cost, novelty, local privacy, and human benefit visible. |
| Lineage DAG | Makes learning reusable by recording parents, operators, and evidence. |
| Registry | Stores packages, manifests, model cards, and release packets. |
| Resource ledger | Makes frugal AI measurable through latency, memory, energy, and maintenance cost. |
| Evolution controller | Proposes, compares, promotes, archives, or no-ops based on evidence. |
| Release evidence | Converts experiments into confidence-building adoption packets. |
| Local-first package format | Lets artifacts be copied, hashed, audited, and run on controlled hardware. |
| Human review and choice | Keeps expertise in the loop and turns human judgment into durable capability. |
Data objects
STRUCT Genome
id
parent_ids
base_model
adapters
merge_recipe
quantization
routing_policy
mutation_budget
provenance
created_at_utc
END STRUCT
STRUCT FitnessVector
utility
calibration
robustness
latency
memory
energy
privacy
novelty
maintainability
human_benefit
evidence_uri
evaluated_at_utc
END STRUCT
STRUCT ReleasePacket
candidate_id
champion_id
intended_use
comparison_summary
fitness_delta
resource_delta
lineage_uri
evaluation_set_uri
release_stage
rollback_target
reviewer_notes
created_at_utc
END STRUCTGuide map
- Reference architecture
- Local model ecology stack
- Hybrid local/cloud routing
- Local model innovation stack
- Lineage DAGs make capability reusable
- Fitness vectors for useful descendants
- Browser-local breeding workbench
- Zero-dependency Rust browser LLM
- TinyRustLM runtime bridge
- Adapter and merge registry
- Router and coalitions
Local and sovereign stack
The local AI adoption reports make the architecture section more practical: the model ecology must run well on controlled hardware, compact model formats, local runtimes, adapters, private memory, registries, and evaluators.
Local AI architecture expansion
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.
Implementation directivesModelBreeder v2.7.0 Positive Exploration DirectiveCurrent operating directive · 35.8 KB
Architecture and project resourcesModelBreeder Architecture and ProjectsSource report · 53.0 KB
Architecture and project resourcesModelBreeder Architecture and ResourcesSource report · 50.9 KB
Architecture and project resourcesModelBreeder Architecture and Resources Part 2Source report · 33.5 KB
Evolution lab and dashboardsExploring Ecology Dashboard ArchitectureSource report · 64.7 KB
Architecture and model mergingModel Breeder Architecture Deep DiveSource report · 76.7 KB
Architecture and runtime explorationModelBreeder Architecture ExplorationSource report · 48.9 KB
Browser and edge runtimeArchitectural Advancements for Zero-Dependency In-Browser Large Language ModelsEmerging implementation architecture · 54.0 KB