A single signaling question in a RoB domain. Attributes: id: Question identifier (e.g., "1.1"). text: Full question text.

Mock LLM adapter that returns predefined responses. Used exclusively for offline testing. Never call this in production. Args: model

Structured review criteria supporting multiple frameworks. Generalises ``PICOCriteria`` to support PICO, PEO, SPIDER, and custom frameworks.

A single criteria element (e.g., Population, Intervention). Represents one dimension of the review criteria with include/exclude terms, ambig

A single literature record to be screened. Attributes: record_id: Unique identifier (auto-generated UUID if not provided). title:

Assessment of a single PICO element from one LLM call. Attributes: match: Whether the record matches this element (None = unable to asses

Orchestrate the 5-step criteria generation pipeline. Steps: preprocess -> detect framework -> generate -> validate -> refine. Args:

Final screening decision for one record (output of Layer 4). Attributes: record_id: Reference to the screened Record. stage: Scre

Orchestrate the full evaluation workflow. Computes screening metrics, AUROC, calibration metrics, and bootstrap CIs. Delegates to Phase 3 ca

Output from a single LLM model for one record. Attributes: model_id: Identifier of the model (e.g., "qwen3", "deepseek-v3"). deci

Save, load, and resume wizard sessions. Persists ``WizardSession`` objects as JSON files on disk, keyed by their ``session_id``. Supports lo

Schema for one RoB assessment domain. Attributes: domain: The RoB domain enum value. name: Human-readable domain name. si

Cochrane Risk of Bias 2 (RoB 2) tool schema for RCTs. Defines 5 domains with 22 signaling questions following the official Cochrane RoB 2 ins

State for an interactive criteria wizard session. Tracks the current step, draft criteria, and Q&A history so that the wizard can be resumed

Title/Abstract screening orchestrator — full HCN pipeline. Runs the complete 4-layer HCN pipeline for each record and produces a ScreeningDec

Definition of a single extraction form field. Attributes: type: The data type of the field. description: Human-readable descripti

Orchestrator for multi-LLM risk of bias assessment. Pipeline: 1. Auto-select tool by StudyType (or manual override) 2. Chunk text (reuse

Routes model outputs to a final decision and confidence tier. Uses configurable thresholds for tier assignment. Implements recall bias at Tie

Structured inclusion/exclusion criteria based on PICO framework. Attributes: criteria_id: Unique identifier for this criteria set.

Calibrated Confidence Aggregator for ensemble LLM outputs. Computes a weighted, calibrated ensemble score and a decision agreement-based ense

Orchestrator for multi-LLM parallel data extraction. Implements the chunk-then-merge strategy: 1. Split text into chunks 2. For each chun

ROBINS-I tool schema for non-randomized studies of interventions. Defines 7 domains with 24 signaling questions following the official ROBINS

QUADAS-2 tool schema for diagnostic accuracy studies. Defines 4 domains with 11 signaling questions following the official QUADAS-2 instrumen

AUROC computation result with ROC curve data points. Attributes: auroc: Area Under the ROC Curve. fpr: False positive rates at ea

Exclude records with non-primary publication types. Triggers on either the ``study_type`` field or title keyword matching. Unknown study type

Orchestrates all Layer 2 rules and produces a RuleCheckResult. Runs hard and soft rules in sequence, collects violations, and computes the to

A single rule violation detected by the semantic rule engine. Attributes: rule_name: Unique name of the violated rule. rule_type:

Load and search built-in criteria templates. Templates are stored as YAML files in the builtin_templates directory.

A single bin in the calibration reliability diagram. Attributes: bin_lower: Lower bound of the bin. bin_upper: Upper bound of the

Platt scaling calibrator using logistic regression. Maps raw scores to calibrated probabilities via a sigmoid. Returns identity (input unchan

Penalize records where majority of models report population mismatch. Checks multiple element keys ("population", "sample") to support PICO,

Quality assessment of generated criteria on five dimensions. All scores are integers in [0, 100]. The ``suggestions`` list contains actionabl

Risk of Bias judgement for a single domain (Module 3). Attributes: domain: The RoB domain being assessed. judgement: Consensus Ro

Penalize records where models disagree on intervention match. Checks multiple element keys ("intervention", "comparison", "exposure", "phenom

Calibration quality metrics with per-bin data. Attributes: ece: Expected Calibration Error. mce: Maximum Calibration Error.

Screening decision for a single paper.

Exclude records whose language is outside the allowed list. If ``criteria.language_restriction`` is set and the record's language is known an

Exclude records whose study design is in the exclusion list. Matches ``record.study_type`` against ``criteria.study_design_exclude`` using ca

Optimize decision router thresholds via grid search. Searches over τ_high × τ_mid × τ_low to maximize automation rate (fraction of records no

Threshold configuration for the DecisionRouter. Attributes: tau_high: Confidence threshold for Tier 1 (unanimous). tau_mid: Confi

Learn and manage per-model weights for CCA aggregation. Weights are learned via SLSQP minimization of binary cross-entropy between weighted m

Generate review criteria via LLM inference with optional multi-model consensus. When initialised with a single backend the generator operates in

Detect the most appropriate SR framework using an LLM. Args: backend: LLM backend for inference.

Layer 1 orchestrator: builds prompt + runs all backends in parallel. Uses ``PromptRouter`` to select the framework-specific prompt template,

Isotonic regression calibrator. Maps raw scores to calibrated probabilities using isotonic regression, which produces a monotonically non-dec

Routes criteria framework to the appropriate prompt template. Maps ``CriteriaFramework`` enum values to their corresponding ``ScreeningPrompt

Screening performance metrics. Attributes: sensitivity: True positive rate (recall). specificity: True negative rate. pre

Unsupported file format.

A warning generated during extraction validation. Attributes: field_name: The field that triggered the warning. message: Human-re

Bootstrap confidence interval result. Attributes: point: Point estimate of the metric. ci_lower: Lower bound of the 95% CI.

Invalid or incomplete PICO criteria.

User-defined extraction form loaded from YAML. Attributes: form_name: Human-readable name of the form. form_version: Version stri

AI-assisted extraction form generator. Uses a single LLM backend to generate a draft extraction form from a research topic description.

Result of framework auto-detection. Attributes: framework: The detected or overridden framework. confidence: Confidence score (0.

Screening prompt for CUSTOM or unsupported frameworks. Renders all elements in ``criteria.elements`` using their ``name`` field as the label,

Penalize records where majority of models report outcome mismatch. Checks multiple element keys ("outcome", "evaluation") to support PICO, PE

Runs multiple LLM backends in parallel and collects their outputs. Args: backends: Sequence of LLMBackend instances to run in parallel.

Complete Risk of Bias assessment for one paper (Module 3). Attributes: record_id: Reference to the assessed Record. tool: Name of

Result of the semantic rule engine for one record (Layer 2 output). Attributes: hard_violations: List of hard rule violations (each trigg

A single validation issue found during rule-based checks. Attributes: severity: Issue severity ('error', 'warning', 'info'). elem

Systematic review criteria framework type.

Pre-built criteria template for common review types. Templates provide a starting point for criteria generation and can be customised by the

Complete evaluation report with all metrics and CIs. Attributes: metrics: Screening performance metrics. auroc: AUROC with ROC cu

Numeric validation constraints for an extraction field. Attributes: min: Minimum allowed value (inclusive). max: Maximum allowed

Audit trail for AI-assisted criteria generation. Records how criteria were generated, which models participated, and the raw per-model output

LLM adapter using the OpenRouter API. Args: model_id: Internal model identifier (e.g., 'qwen3'). openrouter_model_name: OpenRoute

Screening prompt for PEO (Population, Exposure, Outcome).

Screening prompt for PICO (Population, Intervention, Comparison, Outcome).

Hierarchical decision routing tier (Layer 4). Lower value = higher priority.

Base for all LLM-related failures.

Failed to parse LLM JSON response.

LLM API rate limit exceeded.

LLM API call timed out.

Root configuration for MetaScreener. Attributes: models: Registry of available LLM models. thresholds: Decision router threshold

Screening prompt for PCC (Population, Concept, Context).

Failed to extract text from PDF.

Risk of Bias judgement for a single domain.

Screening prompt for SPIDER framework. SPIDER: Sample, Phenomenon of Interest, Design, Evaluation, Research type.

Screening stage identifier.

Complete TRIPOD-LLM audit trail entry for one screening decision. Captures all information needed to fully reproduce a screening decision, in

Calibration failed due to insufficient data.

Minimal concrete implementation for testing.

Data extraction result for one paper (Module 2). Attributes: record_id: Reference to the source Record. form_version: Version of

LLM inference configuration. Attributes: temperature: Sampling temperature (0.0 for deterministic). timeout_s: Timeout per LLM ca

Base exception for all MetaScreener errors.

A single LLM model configuration entry. Attributes: name: Full model name (e.g., "Qwen/Qwen3-235B-A22B-Instruct"). version: Versi

Abstract base class for Risk of Bias tool schemas. Each concrete subclass defines the domains, signaling questions, and overall judgement log

Abstract base class for screening prompt templates. Subclasses implement ``build_criteria_section`` to render framework-specific criteria (PI

Decision router threshold configuration. Attributes: tau_high: Confidence threshold for Tier 1 (unanimous). tau_mid: Confidence t

Mock adapter that returns invalid JSON to trigger parse errors.

Qualitative confidence level.

Merge criteria from multiple LLM outputs using exact-string consensus. Current implementation uses exact string matching for merging. Semanti

How criteria input is provided.

Read/write ReviewCriteria as YAML with legacy format migration.

Validate review criteria via rule-based checks and LLM quality assessment. Provides two validation layers: - ``validate_rules`` (Layer 1): Fa

Data type for extraction form fields.

Base for file I/O failures.

Static utility class for preprocessing raw text input. Handles PDF noise removal, text truncation with word-boundary awareness, and simple Un

Abstract base class for all LLM adapters. Subclasses must implement `_call_api()` and `model_version`. All calls use temperature=0.0 for repr

Risk of Bias assessment domains (shared prefix convention).