Expand description
Clean-room primitives for autonomous research loops.
The crate models the reusable parts of autoresearch systems: experiment lineage, objective metrics, gates, outcomes, and frontier selection. It intentionally avoids host-specific agent execution and any scanner or exploit implementation so it can stay reusable and publishable.
Re-exports§
pub use acceptance::AcceptanceCriterion;pub use acceptance::AcceptanceVerdict;pub use adapter::Candidate;pub use adapter::EvaluationBatch;pub use adapter::GateRunner;pub use adapter::ObjectiveScores;pub use adapter::ReflectiveDataset;pub use adapter::ReflectiveEntry;pub use adapter::ResearchAdapter;pub use cache::EvaluationCache;pub use cache::cached_evaluate;pub use cache::candidate_hash;pub use embedding::Embedder;pub use frontier::FrontierStrategy;pub use frontier::RankedCandidate;pub use frontier::enforce_acceptance;pub use frontier::rank_frontier;pub use graph::ExperimentGraph;pub use graph::ExperimentId;pub use graph::ExperimentNode;pub use graph::ExperimentStatus;pub use graph::GraphError;pub use graph::Hypothesis;pub use graph::ResearchMode;pub use lm_loop::AbsorbReport;pub use lm_loop::ContextPack;pub use lm_loop::EvidenceConfidence;pub use lm_loop::ExperimentPlan;pub use lm_loop::ExperimentState;pub use lm_loop::HypothesisCandidate;pub use lm_loop::HypothesisState;pub use lm_loop::LedgerEvent;pub use lm_loop::Observation;pub use lm_loop::ResearchDossier;pub use lm_loop::ResearchLedger;pub use lm_loop::ResearchResult;pub use lm_loop::ResultVerdict;pub use lm_loop::WorkBudget;pub use lm_loop::render_dossier;pub use loop_opt::OptimizationState;pub use loop_opt::RejectionRecord;pub use loop_opt::optimize;pub use opportunity::DispatchClass;pub use opportunity::DispatchHint;pub use opportunity::ResearchBias;pub use opportunity::ResearchOpportunity;pub use opportunity::SurfaceKind;pub use opportunity::dispatch_class_for;pub use opportunity::rank_opportunities;pub use outcome::Metric;pub use outcome::MetricDirection;pub use outcome::Objectives;pub use outcome::OutcomeStatus;pub use outcome::TaskScore;pub use outcome::TrialOutcome;pub use pareto::FrontierKey;pub use pareto::FrontierType;pub use pareto::ParetoFront;pub use pareto::non_dominated_set;pub use pareto::pareto_front;pub use policy::Gate;pub use policy::GateOutcome;pub use policy::GatePhase;pub use policy::ResearchPolicy;pub use proposer::DEFAULT_REFLECTION_TEMPLATE;pub use proposer::LanguageModel;pub use proposer::Proposer;pub use proposer::ReflectiveMutation;pub use stop::LoopSnapshot;pub use stop::StopCondition;
Modules§
- acceptance
- Acceptance criteria for proposed candidates.
- adapter
- Execution-agnostic adapter boundary.
- cache
- Evaluation cache: skip redundant rollouts for (candidate, example) pairs.
- embedding
- Host-implemented text embedding.
- frontier
- graph
- ids
- lm_loop
- loop_
opt - The optimization loop body.
- opportunity
- outcome
- pareto
- Pareto frontier over the evaluated pool.
- policy
- proposer
- Proposers generate new candidates from a parent and a reflective dataset.
- stop
- Enforceable stop conditions for the optimization loop.