FEAGI Trainer ADR Set
Status: Proposed
Date: 2026-06-07
Owners: FEAGI Trainer Architecture Working Group
Supersedes: None
Related: docs/FEAGI_TRAINER_ARCHITECTURE_AND_DESIGN.md, docs/EXPERIENCE_CAPTURE_ARCHITECTURE_AND_DESIGN.md (upstream dataset producer)
ADR-001: Module Placement and Runtime Boundary
Status
Proposed
Context
The Trainer currently spans frontend-oriented patterns (legacy feagi-react-core-pro training UI/client flow) while the target architecture requires deterministic orchestration, provenance, compatibility validation, and FEAGI binding selection. These are service responsibilities, not UI responsibilities.
Without a clear boundary, protocol logic leaks into UI and creates inconsistent execution paths.
Decision
Implement FEAGI Trainer as a feagi-desktop plugin surface backed by a dedicated Trainer service boundary:
- UI lives in desktop plugin windows and handles researcher interactions.
- Orchestration, validation, adapter execution, FEAGI bindings, and run artifact lifecycle live in backend service/runtime.
- UI must not own FEAGI protocol semantics for production benchmark runs.
Responsibility boundary (Trainer vs Desktop) — clarification
FEAGI Trainer is one piece of the puzzle, not the home of the whole vision. The boundary is:
- feagi-trainer (this module) owns the train/evaluate/benchmark engine: adapters, samplers, encoder/decoder binding selectors, metric packs, deterministic run execution, and the generation of
PredictionRecords and localScorecards for a given dataset + connectome + protocol. It is a reusable Rust component with a Control API, usable headless. - feagi-desktop (and Composer) owns the product-level surfaces that the Trainer plugs into: the experiment infrastructure (
experiment_id/session_id, lifecycle, genome auto-save/versioning), Brain Hub integration and genome publishing, scorecard publishing, competitions/leaderboards/scoreboards, hosted dataset assets, Composer sync, and the cross-experiment UI. - Experience Capture (
feagi-experience-capture) owns upstream dataset production: live acquisition from sources/devices, labeling, validation, and packaging into Experience Dataset Packages. The Trainer is a downstream consumer that imports these packages via an adapter; it does not capture or label data. Producer-side decisions (capture format, label schemas, source profiles) live inEXPERIENCE_CAPTURE_ARCHITECTURE_AND_DESIGN.mdand its own decision log, not in this ADR set.
In short: Experience Capture produces datasets; the Trainer produces verifiable results from them and plugs into experiment infra, Brain Hub, and more; the desktop decides how those results are published, compared, ranked, and hosted. Nothing competition/leaderboard/publishing-specific is implemented inside feagi-trainer.
This boundary is also a licensing boundary: the engine ships as the open-source feagi-trainer crate (Apache-2.0, in feagi-core), while the closed-source "FEAGI Trainer" app (in feagi-desktop) and Composer consume it one-way. See ADR-006 for the two-artifact split and the open/closed dependency invariants.