Skip to Content
BayesiaLabWhat's New?

BayesiaLab 11.9 is Here

What’s New?

BayesiaLab 11.9 brings three major advances. Hellixia, BayesiaLab’s integrated generative AI assistant, can now explain your analysis reports in plain language. Bayesian credible intervals show the uncertainty around estimated probabilities and effects. Textual Imputation adds sentiment analysis to help you understand the opinions expressed in customer comments and other open-ended responses.

Hellixia uses large language models (LLMs) to help you build, quantify and document networks. This release redesigns its elicitation tools and introduces the Conversational Network Modeler for editing existing networks. It also improves analyses, documentation tools and everyday workflows.

In this release:

The examples explore telecom customer churn, meaning subscription cancellation. The network-generation, editing and elicitation examples use illustrative models and probabilities. The analysis reports use a separate network learned from 7,043 customer records. A separate set of 12 fictional comments demonstrates sentiment analysis. Each section identifies the model used, so you can explore the features independently.

Hellixia Report Analyzer: Your Reports in Plain Language

The new Hellixia - Report Analyzer button turns an analysis report into a plain-language explanation. It identifies the main findings, cites relevant figures and describes the limits of the analysis. The explanation opens in a separate window that you can read alongside the source report.

Hellixia draws on the context you enter in the dialog and on the network’s documentation. Long Names provide descriptive labels for nodes, while Node Comments and Arc Comments explain their meaning and relationships. The Network Comment adds context about the model. With Autodetect selected, Hellixia writes in the language of the supplied context. The documentation example shows how to make these descriptions more informative.

In the following example, the question is which customer characteristics are associated with cancellation, and how strongly. The Target Node, the node selected as the focus of the analysis, is Churn\mathit{Churn}. Hellixia explains the Total Effects table and compares its standardized estimates, which put the associations on a common scale. This analysis uses the learned customer-data network with no intervention applied, so the results describe statistical associations.

BayesiaLab Total Effects report showing point estimates for four indicators associated with customer cancellation
Click to Zoom
Total Effects on Target for the four customer characteristics.
Hellixia Analysis Report explaining the direction and relative strength of four associations with customer cancellation
Click to Zoom
Hellixia’s explanation of the same Total Effects results.
Telecom Report Analyzer
Includes 7,043 records, node documentation and the binary state values used in the reports.
XBL

The Report Analyzer covers reports on relationships, effects, observations and model performance. For example, it can explain which nodes inform a target, how observations change other probabilities, or how evaluation results vary across networks.

Supported reports and what their explanations cover

Analysis reportWhat Hellixia explains
Target ReportHow much information the other nodes provide about the target, using Relative Mutual Information, and what the target’s states mean in the application domain.
Relationship AnalysisThe strength and direction of relationships, using Global Force, Overall Contribution and Pearson correlation. Direction depends on state ordering. Discussion of possible confounding is limited to common parents represented in the graph, and a causal interpretation requires supporting context.
Target AnalysisWhich nodes provide information about the target, using Mutual Information, Normalized Mutual Information and Relative Significance, which compares each node’s information with that of the most informative node.
Total EffectsThe direction and size of effects under intervention, or associations without intervention, with factors ranked by standardized estimates.
Direct EffectsDirect relationships in the graph, including contributions and elasticity, which expresses a proportional response to change.
ContributionsWhich factors contribute to the target, how they work together, what is lost when a factor is removed, and what it contributes alone.
Difference DecompositionHow each factor contributes to a difference between two results, and how that attribution depends on the order in which factors are considered.
Information AnalysisHow much one group of nodes tells you about another, including information available after taking a third group into account. See the cancellation example.
Evidence Analysis (Information Gain)How consistent the observations are, which unobserved states they make more or less likely, and which support or contradict the observed target state.
Most Relevant ExplanationsWhich combinations of unobserved states best explain the observations, and the strength of those explanations.
Hidden Variable DiscoveryPatterns that may suggest unobserved factors and warrant further investigation.
Target Optimization (genetic)A retained optimization solution. The explanation becomes available once a solution has been retained.
Target Dynamic ProfileEach profile’s results and the reason the procedure stopped.
Targeted EvaluationPredictive performance and its variability across evaluated networks, including jackknife, bootstrap, K-fold and Markov blanket evaluations.
Structural ConfidenceSimilarities and differences between learned and reference structures, including arc directions and pairs of arcs that converge on a node, known as V-structures.
Structural Coefficient AnalysisHow structures and validation results change with the structural coefficient, including log-likelihood on training and test data.
Multi-Target EvaluationImprovement over the reference, reduction in uncertainty, and stability across networks.
Global Network PerformanceHow well the network represents the data, using Contingency Table Fit, cases assigned zero probability, and comparisons of training and test data.
Network ConsistencyHow well the cases agree with the network, using Information Gain, Mean Compression and cases identified as conflicting.

Credible Intervals: How Well Do Your Data Support Your Results?

An estimated probability depends on the available data and prior information. Credible Intervals describe uncertainty about that estimate, helping you distinguish precisely estimated results from those based on limited observations.

The network still stores a single estimate in each cell of its Conditional Probability Tables (CPTs), and ordinary inference is unchanged. You can examine uncertainty at three levels: individual table entries, measures calculated from the network, and effects displayed in analysis reports. The examples below use the same learned customer-data network as the Report Analyzer example.

The Credible Interval Report

The Credible Interval Report replaces the Confidence Interval Report with a Bayesian calculation of uncertainty. For each marginal or conditional probability, it shows the estimate, the lower and upper bounds, and the selected credibility level, such as 95% or 99%. Its color coding highlights differences in support from the data.

In the screenshot, the red estimate of 5.2958% has a 95% credible interval of [0.1446%, 18.5866%]. Given the data, prior information and model, the probability parameter has a 95% posterior probability of lying within that interval.

That estimate is based on just 19 records with the same combination of parent states, called a parent configuration, and one observed cancellation. The green estimate uses 852 matching records and 617 cancellations. Comparing the counts and intervals shows why some conditional probabilities are estimated more precisely than others.

Bayesian intervals and prior information

Credible Intervals work with binary and multi-state nodes. They respect the parameter’s bounds, such as [0, 1] for probabilities, and can be asymmetric. Rarely observed parent configurations typically produce wider intervals. The Bayesian calculation replaces the former normal approximation and its special handling of rare events, including the Rule of Three.

Prior information represents knowledge supplied before learning from the current data. The calculation supports estimation without added prior samples, as well as uniform priors, a fully specified prior network, and virtual samples generated from a prior network.

Prior Samples
How BayesiaLab combines prior information with observations.
BayesiaLab conditional Churn probabilities, with 95 percent credible intervals and red, yellow and green cells
Click to Zoom
95% credible intervals for conditional Churn probabilities. The red 5.2958% estimate uses one churn event in 19 matching records; its interval is [0.1446%; 18.5866%]. The green 72.4175% estimate uses 617 events in 852 records.
Telco Churn — Credible Intervals
Explore parameter uncertainty in the learned customer-data network.
XBL

Credible Interval Analysis: Uncertainty in Calculated Results

To assess how uncertainty in the probability tables affects the network’s results, BayesiaLab repeatedly samples plausible tables and recalculates the selected measures. This procedure, called posterior Monte Carlo simulation, shows how much the results vary under the model and its data. It can calculate:

  • state probabilities and the mean of the target;
  • Mutual Information, Normalized Mutual Information and Relative Mutual Information;
  • Direct Effects and Total Effects, standardized or not.

In the Credible Intervals dialog, Number of Samples sets the number of simulation draws. It is separate from the number of records in your dataset. Choose the measures to calculate with Compute Metrics on Target, Compute Total Effects on Target and Compute Direct Effects on Target. A Fixed Seed makes the simulation reproducible.

How the probability tables are sampled

Each CPT row is represented by a Dirichlet distribution centered on its learned values. A draw from this distribution supplies a plausible set of probabilities for that row. BayesiaLab samples the rows, runs inference with the resulting tables and recalculates the requested measures. Repeating this process produces a simulated posterior distribution for each measure.

Credible Intervals dialog with 10,000 samples, fixed seed 20261109 and Compute Metrics on Target enabled
Posterior simulation settings: 10,000 samples and fixed seed 20261109. Compute Metrics on Target is enabled; Total and Direct Effects are disabled for this example.

The Credible Interval Analysis window shows the simulated posterior distribution of each selected measure as a density or distribution function, together with summaries such as its mean, median, standard deviation, quantiles and credible interval. In the following example, the curves describe uncertainty about the estimated probabilities of the two Churn\mathit{Churn} states.

Credible Interval Analysis chart displaying posterior probability densities for both Churn states, using 100 histogram intervals
Click to Zoom
Posterior probability densities for Churn=yes and Churn=no, with 100 histogram intervals. The simulated mean for Churn=yes is 26.2869%, with a standard deviation of 0.5296 percentage points.

Credible Intervals in the Effect Reports

MenusAnalysisReportTargetTotal Effects on Target MenusAnalysisReportTargetDirect Effects on Target

You can also add intervals directly to an effect report. In the Effect Analysis dialog, check Compute Credible Intervals and set the simulation sample count and credibility level. The report then shows intervals for both raw and standardized estimates.

The example below examines the same four customer characteristics as the Report Analyzer example. The binary states are coded as 1 for yes\mathrm{yes} and 0 for no\mathrm{no}. With no intervention applied, the estimates describe associations with cancellation.

Total Effects on Churn report for Dependents, Partner, Paperless billing and Senior citizen, with 99 percent credible intervals
Click to Zoom
Associations with Churn=yes for four binary indicators coded yes=1 and no=0. The report includes raw and standardized effects with 99% credible intervals from 10,000 samples. No intervention is applied.

The Quadrants chart draws the intervals as whiskers. This view plots effect magnitudes, with downward triangles indicating negative associations and upward triangles indicating positive associations. You can hide the whiskers with Display Intervals in the chart’s context menu.

Loading SVG...
Standardized effect magnitudes with 99% credible interval bars. Downward triangles indicate negative associations; upward triangles indicate positive associations. The vertical axis shows magnitude.
Telco Churn — Effect Analysis
The network used for the effect estimates and their credible intervals.
XBL

Reading the Results

For signed estimates in the effect table, interpret the interval relative to zero:

  • an interval entirely above zero supports a positive sign;
  • an interval entirely below zero supports a negative sign;
  • an interval containing zero leaves the sign uncertain at the selected credibility level.

The posterior probability of a positive effect can also be estimated directly. A narrow interval indicates a precisely estimated measure under the model’s assumptions. A wide interval can reflect few observations, rare parent configurations or rarely observed states. Increasing the simulation count improves numerical precision; uncertainty due to limited data remains. A causal interpretation also depends on the model and the use of interventions.

Textual Imputation: From What Is Mentioned to What Is Felt

Textual Imputation converts free-text comments into structured data for analysis. You define topics, called dimensions, such as connection reliability or billing clarity. BayesiaLab 11.9 adds a sentiment mode that records the opinion expressed about each dimension. This lets you analyze both what customers discuss and how they describe their experience.

The dialog offers a new choice, Imputation Mode, which BayesiaLab remembers from one use to the next:

ModeWhat it measuresDimension not mentioned in the text
By PresenceHow clearly the text expresses the topic or propositionScore of 0
By SentimentWhether the expressed opinion is negative, neutral or positiveDedicated state, “Not Mentioned”

In By Sentiment mode, each imputed node gets four states: Not Mentioned\mathrm{Not\ Mentioned}, Negative\mathrm{Negative}, Neutral\mathrm{Neutral} and Positive\mathrm{Positive}. A neutral opinion about billing clarity is different from a comment that does not discuss billing clarity at all.

When a long text is processed in several chunks, BayesiaLab averages the numerical scores for each dimension. This allows both positive and negative passages to contribute to the result, rather than retaining only the most positive score.

The following example uses 12 fictional comments and three dimensions. The dialog shows the selected text column and By Sentiment mode.

Textual Imputation dialog set to By Sentiment for three service dimensions and a Feedback text column containing fictional telecom comments
By Sentiment applied to 12 fictional telecom comments. The dimensions are connection reliability, billing clarity and customer support experience. Unmentioned dimensions remain distinct from neutral opinions.
Synthetic Telecom Feedback
Includes the 12 fictional comments and completed sentiment results.
XBL

One comment in this example reads:

The internet is reliable, but the bill is confusing and customer support was unhelpful.

The resulting states show how a single comment can express different opinions about different topics:

DimensionResult
Connection reliabilityPositive\mathrm{Positive}
Billing clarityNegative\mathrm{Negative}
Customer support experienceNegative\mathrm{Negative}

Updated LLM Network Generators

Hellixia’s generators provide several starting points, from maps of a domain’s concepts to networks with probability tables. You can then refine a generated network with the Conversational Network Modeler or review its numerical assessments with the elicitation tools.

  • Automatic Bayesian Network Generator: revised prompts generate networks with up to four parents per node.
  • Causal Network Generator: the generated structure now defines the unit of analysis and carries it through the probability assessments. For example, the propositional telecom model below concerns one subscriber over a three-month period. When generation starts from an existing node, that node retains its name and meaning. Existing nodes reused as parents keep their tables, and the Network Comment is preserved.
  • Knowledge Graph Generator: revised generation includes checks on the types of relationships represented by arcs across languages.
  • Generate Evidence Scenarios: generated sets of observations are checked before use. Inputs marked as non-observable are excluded, invalid values are rejected, and states are matched by name or long name. Duplicate and empty scenarios are removed.

Understanding the Domain

These telecom examples show four ways to organize a domain. Choose the representation according to the question you want to explore:

RepresentationQuestion it helps explore
Semantic NetworkWhich concepts are related within the domain?
Knowledge GraphWhat kinds of relationships connect the entities?
Semantic FlowchartHow does the process unfold, including its branches?
Causal Semantic DiagramWhich causal mechanisms does Hellixia propose?

These views emphasize meaning and structure. The probabilistic examples below add states and probability tables.

Semantic Network

The Semantic Network groups related concepts in telecom churn. Use it to explore the vocabulary and themes of the domain before deciding which relationships to model.

Loading SVG...
Explore full graph
Related concepts in telecom customer churn are grouped into classes, with descriptions explaining their meaning.
Semantic Network
Source network. Use Modeling Mode to explore the qualitative graph.
XBL

Knowledge Graph

The Knowledge Graph makes the types of relationships explicit through arc labels. It helps you distinguish how entities relate, beyond identifying that they are connected.

Loading SVG...
Explore full graph
Labeled arcs connect the entities in this telecom example.
Knowledge Graph
Source network. Use Modeling Mode to explore the qualitative graph.
XBL

Semantic Flowchart

The Semantic Flowchart follows the subscription and retention process. Its sequence and conditional branches help you examine what happens at each stage.

Loading SVG...
Explore full graph
The subscription and retention process is represented as a sequence with conditional branches.
Semantic Flowchart
Source network. Use Modeling Mode to explore the qualitative graph.
XBL

Causal Semantic Diagram

The Causal Semantic Diagram focuses on proposed mechanisms that could lead to cancellation. It provides a structure for reviewing causal assumptions before assigning probabilities.

Loading SVG...
Explore full graph
Proposed mechanisms leading to customer churn are shown in a Causal Semantic Diagram.
Causal Semantic Diagram
Source network. Use Modeling Mode to explore the qualitative graph.
XBL

Causal Semantic Diagram with Knowledge File

A Knowledge File supplies material to guide generation. This variant uses the fictional Calvera telecom review as additional context, so the proposed diagram can reflect the situation described in that review.

Loading SVG...
Explore full graph
This source-informed variant uses the fictional Calvera telecom review as its knowledge file.
Causal Diagram with Knowledge File
Source network. Use Modeling Mode to explore the qualitative graph.
XBL

Generating Bayesian Networks

The following models include node states and probability tables, allowing you to calculate probabilities and examine scenarios. They illustrate different modeling approaches rather than successive conversions of the diagrams above. The probabilities shown in the node monitors are starting assumptions. The Credible Interval examples use a separate network learned from customer records.

Propositional Causal Network

A propositional model uses statements that can be true or false. This example organizes causes and retention actions around the proposition that a subscriber cancels within the next three months. The green nodes represent actions the provider can take to influence the corresponding causes of cancellation. These actions are also called retention levers.

Loading SVG...
Explore full graph
The network combines economic value, service reliability, switching readiness and support experience around a three-month cancellation event.
Propositional Causal Network
Explore the illustrative telecom model in BayesiaLab.
XBL

Risk-Centric Network

Risk-centric generation organizes the same cancellation proposition around threats, preventive barriers, detection and mitigation. Threats are conditions that could lead to cancellation; preventive barriers act on those conditions. The model also includes cancellation review and actions such as win-back outreach to address the consequences. This structure makes each action’s role in managing the risk explicit.

Loading SVG...
Explore full graph
Preventive barriers act on the causes of cancellation. Cancellation review, win-back outreach and transition support connect the event to its consequences.
Risk-Centric Network
Explore the illustrative telecom model in BayesiaLab.
XBL

Deepening an Existing Branch

The Causal Network Generator can also expand an existing model. In this example, it adds causes above Contract renewal point\mathit{Contract\ renewal\ point} to describe contract timing and the option of offering a contract extension.

Loading SVG...
Explore full graph
The expanded contract-lifecycle branch connects contract timing and a contract-extension lever to renewal, switching readiness and cancellation.
Deepened Branch
Explore the illustrative telecom model in BayesiaLab.
XBL

Networks with Multi-State Nodes and Full Probability Tables

The Automatic Bayesian Network Generator supports nodes with multiple states and general probability tables. This approach is also described as unconstrained generation because it allows a wider choice of state types and table values than a Boolean, formula-based model.

This example includes a Retention offer\mathit{Retention\ offer} decision, representing a choice available to the provider, and a Net retention utility\mathit{Net\ retention\ utility} node, representing its economic consequences. The elicitation example uses this model’s Customer satisfaction\mathit{Customer\ satisfaction} table.

Loading SVG...
Explore full graph
Customer experience and contract status inform churn risk, while the retention decision connects customer value and churn to net retention utility.
Multi-State Bayesian Network
Explore the illustrative telecom model in BayesiaLab.
XBL

Conversational Network Modeler: Reliable Changes to Your Networks

MenusHellixiaConversational Network Modeler

The Conversational Network Modeler lets you describe changes to an existing network in your own words. You can work on the whole network or a selection. Instead of returning a rewritten network, the LLM returns operations to add, update or remove nodes, arcs, tables and classes. BayesiaLab applies these operations while preserving the probabilities, comments and identifiers of untouched elements.

This targeted approach reduces the risk of losing content when a large network is rewritten and limits the generated response to the requested changes. If a change makes a probability table incompatible, BayesiaLab requests a replacement table from the LLM. It also checks the result for directed cycles, valid table dimensions, probability rows that sum to one, and consistency of parent order, intervals and classes.

You can also ask questions about the model. Review and analysis requests produce an answer in a dedicated window while leaving the network unchanged.

Editing networks that use Noisy-OR formulas

Noisy-OR formulas express how individual causes combine to produce an outcome, with a leak allowing for causes outside the model. When editing these networks, the Modeler recomputes tables from the formulas and updates references to renamed nodes. When a cause is added, it recalibrates the leak to preserve the affected node’s baseline probability. Arc color, thickness and effect follow the network’s conventions.

The Modeler supports up to five parents per node for Noisy-OR networks and four for networks with general probability tables.

Evolving the Telecom Model

The following sequence develops the illustrative propositional causal model, starting with a targeted edit and ending with a model of retention decisions and their economic consequences. Its probabilities are modeling assumptions, separate from those of the learned customer-data network. Validation Mode displays the modeled probabilities after each change.

Rename a Node and Remove an Arc

The first request renames Unreliable service\mathit{Unreliable\ service} to Poor service\mathit{Poor\ service} and removes the direct arc from Household or work change\mathit{Household\ or\ work\ change} to Cancels within three months\mathit{Cancels\ within\ three\ months}.

Rename “Unreliable service” to “Poor service” and rewrite its comment accordingly.

Remove the arc from “Household or work change” to “Cancels within three months”.

Loading SVG...
Explore full graph
After the rename to Poor service and removal of the direct household-change arc, the baseline cancellation probability is 11.98%.
Modeler — Rename and Remove an Arc
Saved result of the illustrated network-editing step.
XBL

Add a Cause While Preserving the Marginal Probability

Add a node “Number portability” as a cause of “Easy competitor switch”.

Number portability lets subscribers keep their telephone number when changing providers, making switching easier. Adding Number portability\mathit{Number\ portability} therefore introduces another cause of Easy competitor switch\mathit{Easy\ competitor\ switch}. After this change, the probability of Easy competitor switch=True\mathit{Easy\ competitor\ switch}=\mathrm{True} remains 38.00%, and the baseline cancellation probability remains 11.98%. These are the probabilities before setting any observations.

Loading SVG...
Explore full graph
Adding Number portability preserves the marginal probabilities of Easy competitor switch and customer cancellation.
Modeler — Number Portability
Saved result of the illustrated network-editing step.
XBL

Regroup Causes into Latent Nodes

Group the five parents of “Switching Readiness” by theme into two latent nodes.

Here, latent nodes are unobserved intermediate constructs introduced to organize related causes. The regrouped model adds Switching opportunity group\mathit{Switching\ opportunity\ group} and Needs and plan fit group\mathit{Needs\ and\ plan\ fit\ group} above Switching Readiness\mathit{Switching\ Readiness}. This makes the thematic structure explicit. After regrouping, the baseline cancellation probability is 11.98%.

Loading SVG...
Explore full graph
Two latent groups organize the original causes of Switching Readiness.
Modeler — Regrouped Causes
Saved result of the illustrated network-editing step.
XBL

Add Decisions and Utilities

The model already contains five retention actions, represented as interventions. Decision nodes represent the choice of whether to take each action. Utility nodes quantify the resulting costs and benefits. Adding these elements creates an influence diagram for examining the economic consequences of retention decisions.

Turn the five intervention nodes into decision nodes (apply or not), and add utility nodes for the cost of each intervention and for the subscription revenue kept, in euros per subscriber over three months.

The final request creates decisions with Do not apply\mathrm{Do\ not\ apply} and Apply\mathrm{Apply} labels, together with cost and retained-revenue utility nodes. Their values are expressed in euros per subscriber over three months.

Loading SVG...
Explore full graph
The model becomes an influence diagram, connecting retention decisions to churn and to costs and retained revenue in euros per subscriber over three months.
Modeler — Decisions and Utilities
Saved result of the illustrated network-editing step.
XBL

Redesigned Hellixia Elicitation Tools

Elicitation turns expert knowledge into explicit assessments, such as probability distributions or the contribution of a cause. Hellixia can ask an LLM for these assessments and record its reasoning for review. The redesigned tools distinguish the task of quantifying a network from assessing its causal structure.

ToolUse it to
Probability Tables ElicitorAssess a node’s probability distribution for each relevant combination of parent states. It replaces the Root Priors Elicitor, extending elicitation beyond nodes with no parents.
ICI Local Effects ElicitorAssess each parent’s individual contribution in an Independence of Causal Influence (ICI) model.
Causal Structural PriorsReview existing arcs and obtain assessments of their causal direction and strength. These are the Arc Prior assessments described below.
Pairwise Causal LinkAsk whether to add a causal arc between two selected nodes, with an explanation of the proposed mechanism.

ICI models represent each parent’s contribution separately, then combine those contributions using a specified function. These individual contributions are called local effects. A leak represents contributions from causes not explicitly included in the model. The service-quality example below shows how these pieces fit together.

You can request assessments from 1 to 100 virtual experts for each selected LLM engine. These are separate LLM-generated assessments. Each includes a distribution, a reported confidence and an explanation. The reported confidence expresses the LLM’s assessment of its answer; it is a different quantity from the data-based Bayesian credible intervals discussed above. A new run replaces the earlier assessments from those virtual experts.

Under Subject of the Query, choose which fields to include: Node Name, Node Long Name and Node Comment. Including comments lets you supply definitions and the elicitation questions generated with an ICI model.

Elicitation scope and model-specific behavior

  • Probability Tables Elicitor: requests tables in blocks of 24 conditional rows, including the parent states and rules for consistency between rows. It excludes the effect nodes of OR, AND, binary Vote and Sum models with a warning; these need local-effect elicitation.
  • ICI Local Effects Elicitor: requests one effect per parent state. This is a link probability for OR, the probability that a condition is met for AND, or a signed effect on the node’s scale for Sum, SumPos and SumNeg. Leak and level nodes are handled by Probability Tables Elicitor. The tool does not start if the selection contains no eligible effects.
  • Causal Structural Priors: assesses each selected pair without creating arcs. Results can be an oriented prior with an effect and confidence, a bidirectional prior, or the review flags [REVIEW] Opposite direction proposed and [REVIEW] No causal link.
  • Pairwise Causal Link: adds a proposed arc with a comment describing its mechanism, effect and confidence. It preserves the child’s existing probability distributions by duplicating rows for the new parent’s states. If no causal arc is proposed or the arc would create a directed cycle, it displays an explanation without adding the arc.
  • ICI Model Generator: adds a model-specific elicitation question to each effect node’s comment and a comment to the leak node. Effect nodes are grouped in an Effects class. The MinMax setting is centered on the target’s reference state.
  • Virtual expert identifiers: additional experts are identified with suffixes such as #2 and #3, allowing their assessments to be tracked and replaced on a subsequent run.

Summarizing expert reasoning with BEKEE

The Bayesia Expert Knowledge Elicitation Environment (BEKEE) supports collecting and comparing expert assessments. The Hellixia summary sent with Send Priors now uses only each expert’s latest answer. It keeps experts anonymous, connects each argument to the position it supports, and makes disagreements explicit. The summary focuses on reasoning and omits numerical assessments and a consensus value. No summary is sent if none of the comments contains an argument.

Eliciting the Telecom Model

These examples use two of the illustrative telecom models introduced above. The network with multi-state nodes and full probability tables supplies the Customer satisfaction\mathit{Customer\ satisfaction} example. The regrouped causal model supplies the service-quality branch for ICI elicitation. The resulting probabilities are LLM assessments for subject matter experts to review.

Elicit a Conditional Probability Table

In the model with full probability tables, select Customer satisfaction\mathit{Customer\ satisfaction} and open Menus > Hellixia > Probability Tables Elicitor. Its three parents are Service quality\mathit{Service\ quality}, Price position\mathit{Price\ position} and Support experience\mathit{Support\ experience}, with three, three and two states respectively. Their combinations produce 18 conditional rows, each requiring a distribution over the states of Customer satisfaction\mathit{Customer\ satisfaction}.

The illustrated run used two models each from OpenAI, Anthropic and Google, with one virtual expert per model. Each model assessed all 18 parent configurations, giving 108 assessments. Enable Node Name, Node Long Name and Node Comment. Open Menus > Network > Reports > Assessments to review each distribution, confidence and explanation.

In BayesiaLab 11.9.1, the run used a General Context instruction that supplied both row-index field names expected by the elicitation prompt and response parser. The exported native report confirmed exactly one assessment per model for each parent configuration and a normalized distribution in every row.

BayesiaLab Probability Tables Elicitor report showing 108 assessments from six models and complete coverage of the 18 Customer satisfaction conditional rows
Click to Zoom
Six virtual experts assess all 18 combinations of service quality, price position and support experience.
Probability Tables Elicitor — Completed Example
Includes all 108 assessments across 18 parent configurations.
XBL

Generate the ICI Structure

In the regrouped causal model, Frequent service interruptions\mathit{Frequent\ service\ interruptions}, Weak signal coverage\mathit{Weak\ signal\ coverage} and Persistently low speeds\mathit{Persistently\ low\ speeds} contribute to Poor service\mathit{Poor\ service}. An OR combination models how any of these problems can produce poor service, while the leak allows for other causes.

Select Poor service\mathit{Poor\ service} and its three causes, then open Menus > Tools > Independence of Causal Influence (ICI). Choose Poor service\mathit{Poor\ service} as the output, Or as the combination function, and enable Amechanistic Local Effects, Leak Node and Tree-Based Decomposition. The generator creates separate effect nodes for elicitation, intermediate nodes that combine them, and the leak, together with their elicitation questions.

ICI Local Effects Elicitor
Learn how individual causal contributions are assessed.
ICI Model Generator with Poor service marked as Output, Or selected, and local effects, leak and tree-based decomposition enabled
The ICI Model Generator separates each modeled cause from the residual contribution represented by the leak.
ICI — Generated Structure
Inspect the decomposed network before local-effects elicitation.
XBL

Elicit Local Effects and the Leak

Set Poor service\mathit{Poor\ service} as the Target Node. Select the generated effect nodes and open Menus > Hellixia > ICI Local Effects Elicitor. Enable Node Name, Node Long Name and Node Comment so the generated questions accompany the request. Select Leak 1\mathit{Leak\ 1} separately and use Probability Tables Elicitor for its distribution.

Loading SVG...
Explore full graph
The validation view shows the Poor service branch after the six-model elicitation. Open the figure to explore the complete network.

Compare the Six Models

Open Menus > Network > Reports > Assessments to inspect each expert’s probability distribution, confidence and explanation. In the comparison below, two models each from OpenAI, Anthropic and Google served as one virtual expert per model. Each assessed both states of the three causes and supplied a separate leak distribution, for 42 assessments in total.

ModelFrequent service interruptionsWeak signal coveragePersistently low speedsLeak
GPT-6 Astra98%90%97%12%
GPT-6.1 Sol98%90%98%10%
Claude Fable 5.185%70%75%8%
Claude Opus 5.585%55%80%15%
Gemini 3.1 Pro Preview95%85%90%15%
Gemini 3.6 Flash92%80%90%8%

The cause columns show the elicited local-effect probability when each cause is active. Weak signal coverage had the largest spread, from 55% to 90%, and ranked first in the native report by overall divergence.

BayesiaLab Assessment Report with seven assessments for each of six models from OpenAI, Anthropic and Google, plus a divergence ranking for three causes and the leak
Click to Zoom
The native Assessment Report lists seven assessments per model and ranks Weak signal coverage first by overall divergence.
ICI — Elicited Local Effects
Explore the full validation graph with the local effects, leak and saved assessments.
XBL
ICI — Assessment Report Checkpoint
Open the source model used to generate the 42-assessment report.
XBL

Add an Arc While Preserving the Child Table

Return to the telecom model with full probability tables. Select Contract status\mathit{Contract\ status} and Churn\mathit{Churn}, then open Menus > Hellixia > Pairwise Causal Link. In this run, Hellixia added Contract status\mathit{Contract\ status} → Churn\mathit{Churn} and documented the proposed mechanism. The six existing CPT rows for Churn\mathit{Churn} became twelve rows, with each original distribution duplicated for the new parent’s two states. Because both states of the new parent initially use the same distributions, the added parent does not yet change the predictions. You can quantify its influence through elicitation or learning.

Pairwise Causal Link
Saved telecom network after adding the candidate causal link.
XBL

The next example returns to the learned customer-data network. With Paperless billing\mathit{Paperless\ billing} and Churn\mathit{Churn} selected, Hellixia did not propose a causal arc in this run. Its explanation suggested shared influences, including contract type and customer preferences. This is an LLM assessment of a proposed causal mechanism. The Total Effects example examines the statistical association between these nodes.

Pairwise Causal Link reports no causal relationship between Paperless billing and Churn and cites contract type and customer preferences as shared influences
Click to Zoom
The Paperless billing–Churn example returns a no-link response, with contract type and customer preferences offered as an explanation.
Pairwise Review — Paperless Billing
Learned network used to review the proposed Paperless billing–Churn link.
XBL

Review an Unsupported Candidate Arc

To review arcs already in a network, use Menus > Hellixia > Causal Structural Priors. In the model with full probability tables, it returned oriented priors for Service quality\mathit{Service\ quality} → Customer satisfaction\mathit{Customer\ satisfaction} and Contract status\mathit{Contract\ status} → Churn\mathit{Churn}.

The small example below illustrates a review flag. Random audit label\mathit{Random\ audit\ label} is assigned by an independent coin toss after service measurements and is never used in service decisions. An existing candidate arc connects Service quality\mathit{Service\ quality} to that label. Hellixia marks it [REVIEW] No causal link, explaining why the proposed causal interpretation needs reconsideration.

Loading SVG...
The independently assigned Random audit label provides no causal mechanism for the candidate arc. Causal Structural Priors records its assessment in the arc comment.
Arc Prior — Random Audit Label
Inspect the two-node example and the elicited structural prior.
XBL

Improved Analyses

The analysis tools add clearer ways to select information, investigate possible missing factors and compare performance across networks.

  • Information Analysis: assign nodes to three roles: the quantities you want to understand, information you want to evaluate, and information already taken into account. The new dialog labels these Studied (X), Informing (Y) and Condition (Z). Each node can occupy only one role. Filters by name and class help with selection, which stays synchronized with the graph. The report now begins with a visual summary.
  • Hidden Variable Discovery: the visual report now checks three kinds of patterns that may suggest unobserved factors, rather than one. It flags non-transitive paths, which may indicate that a node combines several dimensions; densely connected groups, called semi-cliques; and dependencies in the data that are absent from the network. The last check uses a Bonferroni correction to account for testing multiple relationships. These findings identify candidates for investigation.
  • Targeted Evaluation over several networks: performance measures now include their mean, minimum, maximum and standard deviation, making variation across evaluated networks visible.
  • Global Network Performance: a new Learning versus Test section supports comparison of training and test data. It includes a Kolmogorov-Smirnov test, reporting DD and pp, and a Student test.

Information About Cancellation

Return to the learned customer-data network and ask how much three customer characteristics tell us about cancellation. Churn\mathit{Churn} is the studied node (XX). Dependents\mathit{Dependents}, Paperless billing\mathit{Paperless\ billing} and Senior citizen\mathit{Senior\ citizen} form the informing group (YY). Partner\mathit{Partner} is the condition (ZZ), allowing us to examine how much information the group provides once partner status is taken into account.

Together, the three informing nodes reduce uncertainty about cancellation by about 9.8%. After conditioning on partner status, they reduce the remaining uncertainty by about 9.0%. These percentages express reductions in entropy, the measure of uncertainty used by Mutual Information. They measure information about cancellation, rather than a reduction in the cancellation rate. The report presents the numerical details alongside a visual summary, and Hellixia explains the same analysis in domain language.

BayesiaLab Information Analysis report with the uncertainty summary and numerical details for the telecom example
Click to Zoom
Information Analysis report with the visual summary and numerical details.
Hellixia Analysis Report explaining cancellation uncertainty before and after accounting for partner status
Click to Zoom
Hellixia’s explanation of the same Information Analysis results.
Telecom Report Analyzer
Includes 7,043 records, node documentation and the binary state values used in the reports.
XBL

Hellixia Documentation and Naming

Clear documentation makes a network easier to interpret and gives Hellixia better context for its explanations. A Node Name identifies the node, a Long Name supplies a more descriptive label, and a Node Comment provides definitions or background. Arc Comments explain relationships.

Node Names, Long Names and Node Comments
Read the e-book introduction to documenting a network.
  • Shortening and expanding comments: set the target length as a percentage of the original, using 75%, 50% or 25% to shorten and 150%, 200% or 300% to expand. For arc comments, the line containing the effect and confidence is preserved.
  • Comment from Name: the prompt asks Hellixia to describe the node and its states using the supplied information.
  • Node Long Name: generates a readable, explicit form of the short name in at most 15 words.
  • State Long Names: all states of a node are processed together in a single request.
  • Class descriptions: generates meaningful descriptions for groups of related nodes. Processing several classes together helps distinguish their meanings. A Description Length checkbox controls the requested length.
  • Variable Clustering: Hellixia can now name every merge in the clustering tree, called a dendrogram, including intermediate groups.
  • Semantic Variable Clustering: groups nodes by meaning. The resulting partition is checked so that each node belongs to one group, group names are distinct and the maximum size is respected. This complements Variable Clustering based on the data.
  • Network Translation: sends the text that needs translation and checks the response for consistent identifiers, state counts, unique names and HTML. You can choose to create a translated duplicate of the network.
  • Image prompts: prepares one English prompt per node for image generation, specifying a concrete subject, a white background and no text in the image.

Readable Node Documentation

This example uses the 16-node learned customer-data network. Generated long names make the node meanings explicit while retaining their short identifiers. The Comments report presents both names alongside each node’s description. The Senior citizen\mathit{Senior\ citizen} comment illustrates shortening an expanded description while preserving the meaning of its two states. These descriptions also supply context to the Report Analyzer.

BayesiaLab Comments report showing short names, generated long names and comments for sixteen telecom nodes
Click to Zoom
Generated long names accompany the node descriptions. The condensed Senior citizen comment retains the meanings of yes and no.
Telecom Documentation and Thematic Groups
Includes node descriptions and the accepted seven-group partition.
XBL

Named Thematic Groups

Semantic Variable Clustering organizes the same nodes into seven thematic groups. These classes organize the nodes by meaning; each node belongs to exactly one group in this result. Running Generate Descriptions from the Class Editor refines the first three group descriptions to “Household Composition”, “Relationship Commitment” and “Billing Structure”.

BayesiaLab Class Report showing seven thematic groups and their member nodes, with generated business descriptions for the first three groups
Click to Zoom
The Class Report lists seven thematic groups and their 16 member nodes, with business descriptions for Household Composition, Relationship Commitment and Billing Structure.
Telecom Documentation and Thematic Groups
Includes node descriptions and the accepted seven-group partition.
XBL

Usability Improvements

  • Endpoint editor: an endpoint configuration defines the connection to an AI service. Copy that configuration to the clipboard as JSON or import it from the clipboard. The new Test button sends a minimal request using the current settings and shows whether it succeeded or the provider returned an error. Testing does not save the settings.
  • Report windows: evaluation and cross-validation reports can remain open while you work elsewhere in BayesiaLab. A Report Analyzer explanation stays in front of its source report, and report windows that are too narrow are widened when opened.
  • All reports: new tooltips display Node Comments in a thin frame, with a width of 320 pixels for long comments.
  • Database merge: now handles continuous columns represented by interval centers, the representative values of discretized intervals.
  • Search fields: press Enter to run a search.
  • Bayesia Academy: replaces the Hellixia Guide references with access to courses, seminars and other learning resources.

Bug Fixes

  • Evidence Analysis: Information Gain was overestimated with soft evidence (likelihoods or fixed probabilities).
  • Target Optimization (genetic) and Target Dynamic Profile: several defects fixed in the HTML reports.
  • Network Consistency: the standard deviation now accounts for case weights, and the minimum and maximum are correctly initialized.
  • Multi-Target Evaluation: log-loss normalization corrected.
  • Targeted Evaluation: the state-specific precision calculation now uses the target state rather than always using the first state.
  • Formula engine: AsBoolean, IfElseIf, Equal and NotEqual corrected, along with several distributions and functions (BetaDist, DualNoisyOrDist, SwitchNoisyDist, RoundTo, among others) and two parser bugs.
  • Start-up: improved stability, including fixes on macOS for a splash screen remaining on display and a blocked restart.
  • Temporal networks: temporal chains are correctly synchronized when a network is copied.
  • Database merge: text row indexes are correct after merging two databases.
  • Mapping font preferences (2D and 3D): a custom font color is no longer reset when you click Apply, and the Use Default Color check box behaves correctly.
  • HTML editor: the content area now fills the window when you enlarge it vertically.
  • Hellixia: fixed a naming conflict in the engine selector.