agentinputlab 0.1.0a1__py3-none-any.whl

agentinputlab/experiments/__init__.py

@@ -0,0 +1,278 @@
+"""Preflight-first experiment manifests, plans, execution, and artifacts."""
+
+# The automatic-measurement import is intentionally delayed below the core
+# schema imports to avoid the selector/experiment package import cycle.
+# ruff: noqa: I001
+
+from .analysis import (
+    analyze_confirmatory_interactions,
+    analyze_selector_policy,
+    load_verified_cell_results,
+    load_verified_terminal_cell_results,
+    summarize_paired_arm_metric,
+    summarize_terminal_execution_outcomes,
+)
+from .artifacts import (
+    ExperimentArtifactIdentity,
+    ExperimentArtifactStore,
+    verify_experiment_run_directory,
+)
+from .auxiliary_matrix import (
+    AUXILIARY_MATRIX_PROTOCOL,
+    AUXILIARY_PROFILE_PROTOCOL,
+    AuxiliaryMethodProfile,
+    auxiliary_pair,
+    auxiliary_static_cells,
+    build_auxiliary_method_cell_matrix,
+    load_auxiliary_method_profiles,
+)
+from .bindings import (
+    FrozenBindingEntry,
+    FrozenBindingReceipt,
+    resolve_frozen_binding_receipt,
+)
+from .blueprints import load_experiment_blueprint
+from .controls import (
+    COST_MATCHED_EXTERNAL_PAYLOAD_PROTOCOL,
+    COST_MATCHED_NO_OP_BUILDER_PROTOCOL,
+    COST_MATCHED_NO_OP_RELEASE_ID,
+    CostMatchedExternalPayloadBuilder,
+    CostMatchedNoOpBuilder,
+    CostMatchedNoOpIntervention,
+    cost_matched_external_payload_builder_factory,
+    cost_matched_no_op_builder_factory,
+)
+from .driver_guard import GuardedExperimentDriver, activate_guarded_driver
+from .executor import (
+    ExperimentCellResult,
+    ExperimentDriver,
+    ExperimentExecutionSummary,
+    build_experiment_result_record,
+    execute_experiment_plan,
+    load_experiment_driver,
+    parse_experiment_cell_result,
+)
+from .loader import (
+    PreflightReport,
+    load_experiment_manifest,
+    preflight_experiment,
+    resolve_target_units,
+)
+from .measurement import (
+    AUTOMATIC_HUMAN_VALIDATION_PROTOCOL,
+    DEFAULT_MEASUREMENT_COUNTS,
+    MEASUREMENT_ANALYSIS_PROTOCOL,
+    MEASUREMENT_ASSIGNMENT_PROTOCOL,
+    MEASUREMENT_KINDS,
+    MeasurementItem,
+    analyze_measurement_annotations,
+    build_measurement_assignment,
+    compare_automatic_human_audits,
+)
+from .paper_analysis import (
+    LEGACY_CONFIRMATORY_PROTOCOL,
+    PAPER_COMMON_SUPPORT_PROTOCOL,
+    PAPER_FAMILY_NATIVE_PROTOCOL,
+    PAPER_RETRIEVER_FACTORIAL_PROTOCOL,
+    PAPER_SELECTOR_PROTOCOL,
+    analyze_confirmatory_protocol,
+    analyze_paper_common_support,
+    analyze_paper_dual_arena_retriever_factorial,
+    analyze_paper_family_confirmation,
+    analyze_paper_retriever_factorial,
+    analyze_paper_selector_claim_hierarchy,
+    build_paper_selector_claim_gates,
+    infer_paper_run_protocol,
+    validate_paper_run_contract,
+    validate_paper_selector_analysis_contract,
+    validate_paper_selector_claim_gates,
+)
+from .parity_identity import (
+    HOST_EXECUTION_RECEIPT_PROTOCOL,
+    ParityImplementationContract,
+    ParityImplementationIdentity,
+    build_parity_implementation_contract,
+    validate_host_process_cell_result,
+)
+from .plan import (
+    ExperimentCell,
+    ExperimentPlan,
+    build_experiment_plan,
+    validate_experiment_plan,
+)
+from .readiness import (
+    ACCEPTANCE_PROTOCOL,
+    ADAPTER_READINESS_PROTOCOL,
+    AdapterGate,
+    MethodAdapterReadiness,
+    ParityAcceptanceSpec,
+    ReadinessCatalog,
+    load_adapter_readiness_matrix,
+    load_parity_acceptance_catalog,
+)
+from .schema import (
+    ExperimentArmKind,
+    ExperimentArmSpec,
+    ExperimentBenchmarkTarget,
+    ExperimentManifest,
+    ExperimentPrerequisite,
+    ExperimentStage,
+    PrerequisiteKind,
+)
+from .stateful_analysis import (
+    STATEFUL_ARENA_REPLICATION_ARTIFACT,
+    STATEFUL_ARENA_REPLICATION_PROTOCOL,
+    analyze_stateful_arena_replication,
+    validate_external_arena_record_receipts,
+)
+from .transfer_analysis import (
+    HELDOUT_MODEL_TRANSFER_PROTOCOL,
+    analyze_heldout_model_transfer,
+    validate_heldout_model_transfer_run,
+)
+from .unseen_analysis import (
+    PAPER_UNSEEN_TAU3_PROTOCOL,
+    UNSEEN_TAU3_ELIGIBILITY_PROTOCOL,
+    UNSEEN_TAU3_SELECTOR_DECISIONS_ARTIFACT,
+    analyze_unseen_tau3_transfer,
+    validate_unseen_tau3_analysis_plan,
+    validate_unseen_tau3_selector_decisions,
+)
+
+# Imported after the core experiment schema because the automatic rule path
+# reuses the public selector feature extractor, whose package also exposes
+# experiment-aware discovery helpers.
+from .automatic_measurement import (
+    AUTOMATIC_INPUT_LABEL_PROTOCOL,
+    AUTOMATIC_MEASUREMENT_RULE_REVISION,
+    AUTOMATIC_REPAIR_QUALITY_PROTOCOL,
+    AUTOMATIC_TRAJECTORY_PROTOCOL,
+    audit_summary,
+    automatic_input_label,
+    automatic_proximal_repair_quality,
+    automatic_trajectory_validity,
+    deterministic_stratified_sample,
+    frozen_state_from_visible,
+    infer_visible_contract,
+    visible_contract_retention,
+    visible_defect_scores,
+)
+
+__all__ = [
+    "ACCEPTANCE_PROTOCOL",
+    "ADAPTER_READINESS_PROTOCOL",
+    "AUTOMATIC_HUMAN_VALIDATION_PROTOCOL",
+    "AUTOMATIC_INPUT_LABEL_PROTOCOL",
+    "AUTOMATIC_MEASUREMENT_RULE_REVISION",
+    "AUTOMATIC_REPAIR_QUALITY_PROTOCOL",
+    "AUTOMATIC_TRAJECTORY_PROTOCOL",
+    "AUXILIARY_MATRIX_PROTOCOL",
+    "AUXILIARY_PROFILE_PROTOCOL",
+    "COST_MATCHED_EXTERNAL_PAYLOAD_PROTOCOL",
+    "COST_MATCHED_NO_OP_BUILDER_PROTOCOL",
+    "COST_MATCHED_NO_OP_RELEASE_ID",
+    "DEFAULT_MEASUREMENT_COUNTS",
+    "HELDOUT_MODEL_TRANSFER_PROTOCOL",
+    "HOST_EXECUTION_RECEIPT_PROTOCOL",
+    "LEGACY_CONFIRMATORY_PROTOCOL",
+    "MEASUREMENT_ANALYSIS_PROTOCOL",
+    "MEASUREMENT_ASSIGNMENT_PROTOCOL",
+    "MEASUREMENT_KINDS",
+    "PAPER_COMMON_SUPPORT_PROTOCOL",
+    "PAPER_FAMILY_NATIVE_PROTOCOL",
+    "PAPER_RETRIEVER_FACTORIAL_PROTOCOL",
+    "PAPER_SELECTOR_PROTOCOL",
+    "PAPER_UNSEEN_TAU3_PROTOCOL",
+    "STATEFUL_ARENA_REPLICATION_ARTIFACT",
+    "STATEFUL_ARENA_REPLICATION_PROTOCOL",
+    "UNSEEN_TAU3_ELIGIBILITY_PROTOCOL",
+    "UNSEEN_TAU3_SELECTOR_DECISIONS_ARTIFACT",
+    "AdapterGate",
+    "AuxiliaryMethodProfile",
+    "CostMatchedExternalPayloadBuilder",
+    "CostMatchedNoOpBuilder",
+    "CostMatchedNoOpIntervention",
+    "ExperimentArmKind",
+    "ExperimentArmSpec",
+    "ExperimentArtifactIdentity",
+    "ExperimentArtifactStore",
+    "ExperimentBenchmarkTarget",
+    "ExperimentCell",
+    "ExperimentCellResult",
+    "ExperimentDriver",
+    "ExperimentExecutionSummary",
+    "ExperimentManifest",
+    "ExperimentPlan",
+    "ExperimentPrerequisite",
+    "ExperimentStage",
+    "FrozenBindingEntry",
+    "FrozenBindingReceipt",
+    "GuardedExperimentDriver",
+    "MeasurementItem",
+    "MethodAdapterReadiness",
+    "ParityAcceptanceSpec",
+    "ParityImplementationContract",
+    "ParityImplementationIdentity",
+    "PreflightReport",
+    "PrerequisiteKind",
+    "ReadinessCatalog",
+    "activate_guarded_driver",
+    "analyze_confirmatory_interactions",
+    "analyze_confirmatory_protocol",
+    "analyze_heldout_model_transfer",
+    "analyze_measurement_annotations",
+    "analyze_paper_common_support",
+    "analyze_paper_dual_arena_retriever_factorial",
+    "analyze_paper_family_confirmation",
+    "analyze_paper_retriever_factorial",
+    "analyze_paper_selector_claim_hierarchy",
+    "analyze_selector_policy",
+    "analyze_stateful_arena_replication",
+    "analyze_unseen_tau3_transfer",
+    "audit_summary",
+    "automatic_input_label",
+    "automatic_proximal_repair_quality",
+    "automatic_trajectory_validity",
+    "auxiliary_pair",
+    "auxiliary_static_cells",
+    "build_auxiliary_method_cell_matrix",
+    "build_experiment_plan",
+    "build_experiment_result_record",
+    "build_measurement_assignment",
+    "build_paper_selector_claim_gates",
+    "build_parity_implementation_contract",
+    "compare_automatic_human_audits",
+    "cost_matched_external_payload_builder_factory",
+    "cost_matched_no_op_builder_factory",
+    "deterministic_stratified_sample",
+    "execute_experiment_plan",
+    "frozen_state_from_visible",
+    "infer_paper_run_protocol",
+    "infer_visible_contract",
+    "load_adapter_readiness_matrix",
+    "load_auxiliary_method_profiles",
+    "load_experiment_blueprint",
+    "load_experiment_driver",
+    "load_experiment_manifest",
+    "load_parity_acceptance_catalog",
+    "load_verified_cell_results",
+    "load_verified_terminal_cell_results",
+    "parse_experiment_cell_result",
+    "preflight_experiment",
+    "resolve_frozen_binding_receipt",
+    "resolve_target_units",
+    "summarize_paired_arm_metric",
+    "summarize_terminal_execution_outcomes",
+    "validate_experiment_plan",
+    "validate_external_arena_record_receipts",
+    "validate_heldout_model_transfer_run",
+    "validate_host_process_cell_result",
+    "validate_paper_run_contract",
+    "validate_paper_selector_analysis_contract",
+    "validate_paper_selector_claim_gates",
+    "validate_unseen_tau3_analysis_plan",
+    "validate_unseen_tau3_selector_decisions",
+    "verify_experiment_run_directory",
+    "visible_contract_retention",
+    "visible_defect_scores",
+]