Every capability, in one place.

A fairness triage reading in one sitting, not an audit: declare what you bring, get a plain-language opinion and a report built from the engine's result, without writing code.

Guided, end-to-end fairness assessment in four pathways (predictive, LLM, agent and multi-agent), with gates and an explicit decision at each stage.

Self-directed path: run any of the 29 fairness modules à la carte, in whatever order your project needs, instead of a fixed wizard.

Scan datasets for representation gaps, historical patterns and statistical disparities.

Part of module 2.3.1 Bias Detection & Prioritization

Surface features that quietly stand in for protected attributes.

Part of module 2.3.2 Feature Analysis & Proxy Detection

Inspect how features relate to protected attributes, and compare fairness-aware transforms.

Residualize features to break proxy chains.

Build a DAG and classify each path from a protected attribute to the outcome as direct, indirect or proxy discrimination.

Part of module 3.1.1 Causal Modeling Template

Flip demographics, see what the model actually changes.

Part of module 3.2.1 Counterfactual Fairness Tool

Auto-detects protected attributes and scans for fairness violations across all built-in metrics.

Split total effects into natural direct (NDE) and indirect (NIE) effects through chosen mediators.

Part of module 3.1.1 Causal Modeling Template

20+ fairness criteria (demographic parity, equalized odds, calibration) and a guided way to choose between them.

Part of module 2.1.1 Fairness Definition & Trade-Off Decision

A prioritised fairness risk register with three-dimension scoring, an EU AI Act Annex III check and governance sign-off.

Part of module 1.3.2 AI Fairness Risk Framework

Link business objectives to value drivers and KPIs, so the fairness work traces back to why the system exists.

Part of module 1.2.1 Business Objectives & Guidelines Alignment

Turn the obligations that apply to your system into a document plan and a delivery roadmap.

Part of module 4.2.1 Compliance Pathway Navigator

The regulatory catalogue behind the Navigator and the Compliance Pathway: obligations across jurisdictions, risk tiers, domains and data categories, each with its legal source.

Part of module 4.2.1 Compliance Pathway Navigator

Rebalance, rebin and transform data before training.

Part of module 3.4.1 Pre-Processing Fairness Techniques

Demographic Parity, Equalized Odds, Equal Opportunity, Adversarial, Counterfactual.

Part of module 3.4.2 In-Processing Fairness Techniques

Exponentiated Gradient and Grid Search over fairness constraints.

Part of module 3.4.2 In-Processing Fairness Techniques

Drop-in, sklearn-compatible fair model wrappers.

L2 parity, conditional independence and HSIC penalties.

Part of module 3.4.2 In-Processing Fairness Techniques

Group-specific temperature and Platt scaling during training.

Platt, Isotonic, Beta and Temperature scaling per subgroup.

Part of module 3.4.3 Post-Processing Fairness Techniques

Per-group thresholds for any fairness constraint.

Part of module 3.4.3 Post-Processing Fairness Techniques

Map the accuracy ↔ fairness frontier and pick a defensible point.

Multiplicative or additive adjustments at inference time.

Part of module 3.4.3 Post-Processing Fairness Techniques

Boost confidence for marginal cases near the decision boundary.

Part of module 3.4.3 Post-Processing Fairness Techniques

One API for classification, regression and ranking metrics.

DP, EO, Equal Opportunity, Exposure, MAE/RMSE parity, ECE/MCE.

Bootstrap and Bayesian intervals, safe on small samples.

Plain-language explanations for every metric.

Detect fairness gerrymandering: strategic subgrouping that hides bias.

Part of module 2.4.2 Robustness Testing

Automated intersectional analysis across multiple protected attributes simultaneously.

Part of module 2.2.1 Affected Populations & Intersection Definition

Show which features drove any single decision, behind one interface, with explanations tiered to the audience. Built for the GDPR Art. 22 right to an explanation.

Part of module 2.3.3 Explainability (XAI)

Decompose a group-fairness disparity into the per-feature contributions that drive it.

Part of module 2.3.3 Explainability (XAI)

Actionable what-if explanations: the smallest change that would flip an individual outcome.

Part of module 2.3.3 Explainability (XAI)

Check that an explanation is faithful and stable, not merely persuasive.

Part of module 2.3.3 Explainability (XAI)

Measure how strongly word and sentence embeddings tie social groups to loaded concepts, before that bias reaches a model.

Part of module 2.3.3 Explainability (XAI)

Check whether a toxicity, sentiment or moderation classifier scores text differently just because it names a particular group.

Part of module 2.3.3 Explainability (XAI)

CI/CD gate that blocks unfair deploys: overall, per-attribute, intersectional.

Part of module 4.2.3 CI/CD Integration

A second reviewer independently signs off each assessment; the server rejects self-validation, so no author can approve their own result.

Pre-training dataset check, no surprises downstream.

Validate fairness configs and model cards before code lands.

Fairness assertions for your existing test suite.

Log fairness next to accuracy in every run.

Multi-level CI/CD gate that enforces overall, per-attribute and per-intersection thresholds together.

See EU AI Act requirement coverage derived from the evidence in your Navigator assessments.

One page per assessment with its readiness verdict, sign-offs and a release approval bound to the approver.

A release gate computed from measured values; exceptions need a second named approver.

Part of module 4.2.2 Fairness DoD Gate

Sliding-window bias tracking in production.

Wavelet + KS, MMD, SPRT: multi-scale drift signals.

Trends, cycles and seasonality of fairness over time.

Self-learning thresholds with regulatory-aware prioritization.

Rank upstream variables by their share of explained drift in outcomes.

A portfolio view of deployed systems: drift, alerts, re-assessment triggers and Art. 73 incident deadlines.

Intersectional A/B testing with proper statistical machinery.

Part of module 4.1.2 Experimentation & A/B Testing

Sample sizes that actually detect what you care about.

Part of module 4.1.2 Experimentation & A/B Testing

Pareto frontier, causal decomposition and recommendations.

Part of module 4.1.2 Experimentation & A/B Testing

Placebo, subset and robustness checks for causal claims.

Part of module 3.1.1 Causal Modeling Template

Privacy-aware metrics warehouse with health scoring.

Publication-grade SVG report templates with zero runtime dependencies.

One-click regulator-grade artifacts from a completed assessment: EU AI Act serious-incident reports, a fundamental-rights skeleton, NYC LL144 ratios and a machine-readable evidence file.

Plotly + Dash, with what-if threshold sliders.

DPIA, Risk Register and Model Card, auto-drafted.

External-audit-ready fairness scorecard.

SHA-256 hash chaining, HMAC signing, zero-knowledge encryption.

A structured model card in the Navigator, designed against the EU AI Act Annex IV, the GDPR, NIST AI RMF, ISO/IEC 42001 and Model Cards for Model Reporting.

Built-in Plotly and Matplotlib renderer for every metric, with multi-theme support.

What an LLM assessment established at each of three evidence layers, on one page: behavioural, model and process, system and decision.

Part of module 2.4.3 LLM Evidence Record

Swap demographics in prompts, measure output drift.

BBQ, BOLD and HolisticBias out of the box.

Truthfulness, stereotypes, toxicity and robustness in one harness.

Sentiment, toxicity and refusal-rate profiling.

Noise floor and consistency baselines per model.

Does the chain-of-thought actually match the answer?

One interface for OpenAI-compatible, Anthropic and custom LLM endpoints.

Paired-artifact bias on CVs, loan applications, tickets.

Which tools an agent calls for whom, and how often.

Retrieval bias caught before generation.

Bias propagation across multi-step agent chains.

Delegation and routing decisions, audited.

Component-fair, system-fair? Compare side by side.

Bias that only appears when agents interact.

Echo chambers and consensus collapse, flagged.

Inter-agent bias amplification under adversarial conditions.

Orchestrator-level demographic routing, surfaced.

Turn-by-turn fairness drift across negotiations.

Trace capture for autogen, crewai and langgraph.

A structured digital-trust assessment built on the iceberg.digital model: ten constructs, 128 measurable cues.

A guided, step-by-step path through an iceberg.digital trust assessment, from stakeholder listening to ship-ready governance.

A board-ready snapshot of trust capability across the iceberg.digital constructs and four layers.

The four-layer trust model (agency, engineering, governance, institutional) with ten constructs and 128 cues.

Browse, analyse and report on trust incidents and positive trust evidence across sectors; in owner preview while the Organisation assessment is built.

The iceberg.digital trust model as one connected graph of constructs, cues and the trust incidents they help explain; in owner preview.

Trends, cohorts and signals across iceberg.digital audits and trust incidents over time.

A consulting method, not software: the three-phase approach behind every Validant engagement.

Facilitator-ready sessions that operationalize the iceberg.digital model: stakeholder mapping, risk perception, trust drift.

Governance artifacts aligned to the iceberg.digital model: DPIAs, risk registers, charters and comms templates.

A curated best-practice gallery of first-person trust patterns, organised by a 13-moment journey: the live Person surface today, ahead of self-sovereign identity.

Credentials a person can hold, such as the Decision Receipt: built and demonstrated on the Swiss e-ID test environment, not yet offered.

The shared knowledge graph behind the Navigator and Dante: fairness concepts, metrics and regulations, the jurisdictions and domains they apply to, and their sources.

Hybrid semantic and keyword retrieval over validant.ai's curated knowledge corpus, feeding Dante.

Every RAG answer ships with source provenance you can audit and challenge.

The shared layer behind Dante: two processing modes, grounded retrieval, written module contracts and guardrails.

Dante, the in-app assistant in every module: it answers questions, helps read results and drafts text, grounded in curated knowledge.

Automated outreach, nurture, evidence requests and incident notifications, fully templated.

Wire Validant events into Slack, Teams, Jira, GitHub or any downstream system.

Unified inbox for drift alerts, incidents, gate failures and stakeholder asks.

Plan and run assessments as a team: boards, backlog, sprints and capacity, with every card linked to the assessment it belongs to and all text encrypted.

A help centre with a searchable FAQ, tickets from the app or by email, a ticket number at once and a progress tracker.

Message colleagues inside the platform, encrypted with your organisation's key.

Compose reports from your assessment's evidence, pick a template and a design, and get a PDF on every plan; on Pro it is sealed and comes with an editable Word copy.

Optional AI-generated narrative sections, customizable in tone and prompt, generated on local hardware so evidence never leaves your environment.

A per-assessment record of who changed what and when: key edits, overrides, sign-offs, iteration closes and module completions.

Okta, Entra, Google Workspace. SCIM provisioning, group sync, MFA enforcement.

Six organisation roles, enforced on the server: owner, admin, project lead, analyst, reviewer and viewer.

Invite teammates by email into the seats your plan includes; members share the organisation's plan and credits.

The data every analysis runs on is encrypted in your browser; the cloud only ever holds ciphertext.

Seals a completed assessment into a tamper-proof credential anyone can verify, with no login and without trusting us. Built; issuance is switched off in production.

An assurance grade on three axes, carried inside the Trust Seal, with the headline set to the weakest of the three, never an average.

On Pro, every PDF report is sealed, so anyone can check at validant.ai/verify that it is the exact file validant.ai issued.

The open-source engine behind every analysis: pip install vfairness, with the verification state of every public function published.