chimera.training

chimera.training is the synthesis layer: drive an agent through multiple passes against a spec until tests pass.

Top-level exports

from chimera.training import (
    Trainer,
    Spec,
    Architecture,
    Constraint,
)
from chimera.training.strategies import (
    TestConvergenceStrategy,
    TreeSearchStrategy,
    CurriculumStrategy,
    EnsembleStrategy,
    MajorityVotingStrategy,
    AIMOEnsembleStrategy,
    PassthroughStrategy,
    CEGISStrategy,
    IncrementalStrategy,
)

Symbol	Module	Purpose
Trainer	chimera.training.trainer	Synthesis orchestrator. Trainer(agent_factory, strategy).run(spec).
Spec	chimera.training.spec	Task specification: description + tests + constraints.
Architecture	chimera.training.architecture	Multi-layer build composition (frontend / backend / tests / docs).
Constraint	chimera.training.constraints	Synthesis constraint base class (formal, example-based, type-based).

Strategies

Strategy	Module	When to use
TestConvergenceStrategy	chimera.training.strategies.test_convergence	Iterate until all tests pass. Default.
TreeSearchStrategy	chimera.training.strategies.tree_search	Branch-and-prune over candidate solutions.
CurriculumStrategy	chimera.training.strategies.curriculum	Easy-first ordering of subtasks.
EnsembleStrategy	chimera.training.strategies.ensemble	Run N agents in parallel, pick best.
MajorityVotingStrategy	chimera.training.strategies.majority_voting	Pick the answer N agents agree on.
AIMOEnsembleStrategy	chimera.training.strategies.aimo_ensemble	AIMO-tuned ensemble.
PassthroughStrategy	chimera.training.strategies.passthrough	One pass, no retries.
CEGISStrategy	chimera.training.strategies.cegis	Counter-example-guided synthesis.
IncrementalStrategy	chimera.training.strategies.incremental	Build up sub-functions incrementally.

See also

• chimera.eval for benchmark-driven evaluation.
• chimera.composition for runtime composition (Pipeline / Ensemble / Supervisor) — strategies are about synthesis-time iteration.