Create a metric dict with point, ci_lower, ci_upper.

Run the full criteria wizard pipeline. Args: input_mode: How the user provides input (text, topic, yaml, examples). w

Check whether this rule is violated. Args: record: The literature record being screened. criteria: The review criteri

Create a ModelOutput with given score, confidence, and decision.

Read literature records from a file, auto-detecting format by extension. Args: path: Path to the input file. Returns: List o

Create a ModelOutput helper.

Normalise a raw record dict from any format into a Record. Args: raw: Raw key-value pairs from a format-specific parser. format_t

Compute screening performance metrics from decisions vs ground truth. INCLUDE is treated as the positive class and EXCLUDE as negative. HUMAN

Load and validate an extraction form from a YAML file. Args: path: Path to the YAML extraction form. Returns: Validated Extr

Load ReviewCriteria from YAML, auto-detecting legacy format. If the file contains the legacy ``PICOCriteria`` format (identified by t

Create a mock adapter that returns framework detection results.

Write Record list to file. Auto-detect format from extension if not given. Args: records: List of Record objects to write. path:

Create a mock adapter with a given extraction response.

Compute all screening metrics with bootstrap CIs. Args: decisions: List of ScreeningDecision objects. gold_labels: Go

Compute overall RoB judgement from per-domain judgements. Applies the Cochrane RoB 2 worst-case rule: - Any HIGH in any domain yields

Route to a final decision and tier. Args: model_outputs: LLM outputs from Layer 1. rule_result: Rule check result fro

Save a wizard session to disk. Args: session: The session to persist. Returns: Path to the saved session fil

Validate extracted fields against the form schema. Checks: 1. Required fields are present. 2. Field values match declared types. 3. N

Create a minimal ExtractionForm.

Aggregate model outputs into a final score and confidence. Applies the CCA formula and computes ensemble confidence from Shannon entr

Render PCC-specific criteria section. Args: criteria: Review criteria with PCC elements. Returns: Formatted

Extract structured data from text using all backends. Args: text: Full text of the paper (from PDF). form: Extraction

Screen a single record through the full HCN pipeline. Args: record: The literature record to screen. criteria: Review

Parse and apply user feedback to a criteria element. Supports simple add/remove commands: - ``+term`` or ``add: term`` → appends to i

Create a mock adapter with a criteria generation response.

Merge multiple model outputs into a single ReviewCriteria. Args: model_outputs: List of parsed JSON dicts from each model.

Helper to create ReviewCriteria with defaults.

Create a metric dict with point, ci_lower, ci_upper.

Build a RoB assessment prompt from tool schema and paper text. Generates prompts dynamically based on the tool schema, so the same function w

Remove markdown code fences from LLM response text. Handles both complete (````...````) and unclosed fences, as well as fences with language

RuleCheckResult with no violations.

Helper to create a ModelOutput.

Create a ModelOutput helper.

Assess risk of bias using a specified tool. Args: text: Full text of the paper. tool_name: Tool identifier ('rob2', '

Fit the Platt calibrator on validation scores. Args: scores: Raw model scores in [0, 1]. labels: Binary labels (0 = i

Return numeric severity rank for ROBINS-I judgements. ROBINS-I uses a 4-level severity scale: LOW (0), MODERATE (1), SERIOUS (2), CRI

Helper to create ScreeningDecision objects.

Split text into overlapping chunks sized for LLM context windows. Splits on paragraph boundaries (double newline). Falls back to character-le

Send a prompt and return the raw text response. Public wrapper around ``_call_api`` for general-purpose LLM calls (e.g., criteria gen

Compute SHA256 hash of a prompt for audit trail. Args: prompt: The prompt string to hash. Returns: 64-character hex string (

Load gold-standard labels from a CSV file. Expects columns: ``record_id`` (or any column containing "id") and one of ``label``, ``included``,

Load model configuration from a YAML file. Args: path: Path to the YAML configuration file. Returns: MetaScreenerConfig with

Merge base layout with caller-supplied overrides.

Create a mock framework detector adapter.

Create a mock generation adapter.

Calibrate a single score. Args: score: Raw model score in [0, 1]. Returns: Calibrated probability in [0, 1].

Compute screening metrics with bootstrap 95% confidence intervals. Converts binary predictions (0/1) to Decision enums, computes screening me

Create a grouped bar chart for the ablation study. Reads exp4_ablation_study.json and plots sensitivity and specificity as grouped bars for e

Create a forest plot of sensitivity across benchmark datasets. Reads exp2_cohen_benchmark.json, extracts per-topic sensitivity point estimate

Detect the appropriate framework for the given input. When ``override_framework`` is provided the LLM call is skipped entirely and th

Detect the primary language of text using Unicode heuristics. Uses character-class ratios to distinguish CJK-family languages from La

Migrate a legacy ``PICOCriteria`` to the new ``ReviewCriteria`` format. Maps PICO-specific fields into the generic ``elements`` dictionary

Get a tool schema instance by name. Args: tool_name: One of 'rob2', 'robins_i', 'quadas2'. Returns: Instantiated tool schema

Load a session by ID. Args: session_id: The session identifier. Returns: The session if found, None otherwis

Save experiment results as JSON with metadata. The output file includes a ``_metadata`` key with timestamp, seed, version, and experiment nam

Build prompt to detect which SR framework fits the user's input. Args: user_input: User-provided text (topic description or criteria text

Build prompt to parse free-text criteria into structured elements. Args: criteria_text: User-provided criteria description. frame

Call LLM with a pre-built prompt and parse the response. Handles both ``element_assessment`` and ``pico_assessment`` keys in the LLM

Remove noise from raw text input. Processing steps: 1. Remove form feed characters. 2. Remove control characters (kee

Format a metric with CI in Lancet Digital Health style. Uses middle dot (U+00B7) instead of decimal period and en dash (U+2013) for CI ranges

Plot a Receiver Operating Characteristic curve. Args: auroc_result: AUROC computation result with FPR/TPR data points. Returns:

Plot overlapping histograms of inclusion scores by true label. Args: scores: Predicted inclusion scores for each record. labels:

Run the full ablation study across all 6 configurations. Loads the dataset, creates backends, runs each configuration (single-model, ensemble

Run cost and time analysis on a dataset. Screens each record individually, measuring wall-clock time and estimating token usage and API cost

Save ReviewCriteria to a YAML file. Args: criteria: The criteria to serialize. path: Target file path.

Create 4 mock LLM backends with distinct model IDs. Each backend returns INCLUDE with confidence 0.85 and score 0.80. Model IDs: mock-qwen3,

Run rule-based validation checks (Layer 1, no LLM cost). Checks for: - Date range contradictions (date_from > date_to). - Inc

Convert value to str or None. Args: value: Any value. Returns: String or None if empty.

Assess risk of bias with auto-selected tool based on study type. Args: text: Full text of the paper. study_type: Stud

Compute bootstrap confidence interval for a metric function. Args: metric_fn: Function that takes a tuple of parallel arrays and

Build prompt to generate review criteria from a topic description. Args: topic: Research topic or question. framework: Framework

Build a framework-specific screening prompt. Args: record: The literature record to screen. criteria: Review criteria

Build the standardized screening prompt for a record. Args: record: The literature record to screen. criteria: The PICO inclusion

Build prompt to assess quality of review criteria. Args: criteria_json: JSON string of the ReviewCriteria to assess. Returns:

Run all rules and produce a RuleCheckResult. Automatically converts ``PICOCriteria`` to ``ReviewCriteria`` for backward compatibility

Generate an extraction form from a research topic. Args: topic: Research topic description. seed: Reproducibility see

Generate Table 2: Cohen 2006 Benchmark Results. Args: results_dir: Path to the directory containing experiment results. Returns:

Auto-select the appropriate RoB tool based on study type. Args: study_type: The study design classification. Returns: Instan

Run all backends with a framework-specific prompt. Args: record: The literature record to screen. criteria: Review cr

Load the most recently modified session. Returns: The latest session if any exist, None otherwise.

Optimize thresholds via grid search. Each validation_data entry is a tuple of (model_outputs, final_score, ensemble_confidence) for o

Parse and validate JSON response from an LLM. Args: raw_response: Raw string response from the LLM API. model_id: Model identifie

Plot a calibration reliability diagram. Args: cal_metrics: Calibration metrics containing per-bin data and ECE. Returns: A P

Render a single criteria element as prompt lines. Args: label: The display label for the element (e.g., "POPULATION"). element: T

Run the HCN screening pipeline on a single Cohen topic. Loads the topic CSV, screens all records via TAScreener, converts decisions to binary

Screen a batch of records sequentially. Args: records: Records to screen. criteria: Review criteria (shared across al

Load experiment results JSON from the results directory. Args: results_dir: Path to the directory containing result JSON files. e

Build prompt to infer review criteria from labeled example papers. Args: examples: List of dicts with 'title', 'abstract', and 'label' ke

Build prompt to refine a criteria element based on user feedback. Args: element_name: Name of the element being refined. current_

Calibrate a single score. Args: score: Raw model score in [0, 1]. Returns: Calibrated probability in [0, 1].

Remove sessions older than max_age_days. Args: max_age_days: Maximum age in days before cleanup. Returns: Nu

Close the underlying HTTP client.

Compute Area Under the ROC Curve with curve data points. Args: scores: Predicted probability scores (higher = more likely positive).

Compute calibration quality metrics with per-bin data. Args: scores: Predicted probability scores in [0, 1]. labels: Binary groun

Extract full text from a PDF file using PyMuPDF. Args: path: Path to the PDF file. Returns: Extracted text content. Empty st

Generate criteria from a research topic (Mode B). Args: topic: Research topic description. framework: SR framework to

Return weights for given model IDs. If fitted, returns learned weights. If unfitted, returns equal weights that sum to 1. Ar