Fair-Lending Bias-Simulation Appliance
FL-BSA
FL-BSA (Fair-Lending Bias-Simulation Appliance) is a self-hosted fair-outcomes evidence appliance for regulated credit decisions. It runs inside the customer-controlled environment and helps teams review credit-decision, affordability, pricing, underwriting, and policy-review workflows.
Each run packages a PDF report, metrics, manifests, and certificates for governance, model-risk, audit, and regulatory review.
FL-BSA highlights
Customer-hosted
Run inside your VPC, VM, or other customer-controlled environment.
Outcome simulation
Stress-test lending decision patterns on synthetic borrower cohorts instead of live production decision traffic.
Evidence packs
Generate reports, manifests, and certificates that support audit, model-risk, and regulatory review.
How it works
Four steps from source snapshot to evidence prepared for governance review.
Step 1: Snapshot the baseline
Point FL-BSA at a bounded, read-only snapshot of features and decisions (prepared CSV/Parquet or a warehouse pull) to establish the baseline for the run.
Step 2: Simulate two branches
Train two generative branches. Amplification asks: are we amplifying the unfairness already in our history? Intrinsic asks: would the model still discriminate if history were fairer? The gap between them is the audit object.
Step 3: Generate hash-linked evidence
Every run emits a tamper-evident evidence manifest (SHA-256) plus certificates that link source data, model settings, and outputs.
Step 4: Review findings and act
Use the review-oriented report findings to tune models or document Less Discriminatory Alternatives (LDAs).
In build: measurement on your own decisions. Today FL-BSA evidences fairness on simulated decision patterns learned from your historical data; it never executes your models. We are building a production-targeted measurement mode for unsigned aggregate fairness measurements over customer-authored decision or score columns. It is in active development and is not part of the current release.
Deployment
Choose the customer-hosted path that fits your procurement and infrastructure model.
Guided AWS AMI
The current AWS path is controlled guided-pilot access to the AMI-first customer path or an agreed private handoff. Public AWS Marketplace access is not yet available.
Container
Deploy the container stack via docker-compose in your VPC, VM, or on-prem environment.
See Procurement & Deployment for deployment steps and prerequisites.
Working with us
Three engagement shapes, smallest first. You control the infrastructure; we licence the appliance through controlled qualification.
Evidence-readiness assessment
About 5 business days. For teams that want a fast stop/go answer. You get a short findings memo on your workflow, schema, data boundary, and evidence gaps, with a recommendation. The recommended first step.
Evidence readiness sprint
About 2 weeks. For teams preparing a controlled evidence run. You get an evidence-gap memo, a draft evidence-pack acceptance checklist, and a scoped plan for the run.
Fair outcomes evidence sprint
6 to 10 weeks. One controlled, customer-hosted evidence scenario delivered under controlled-pilot posture, with evidence-pack deliverables and a governance readout.
Start with the assessment. Pricing for each engagement is provided on request. Requesting the pack gets you a written reply with the buyer and procurement pack (evidence-bundle samples, control mapping, deployment and security materials, and a commercial overview); it commits you to nothing further.
Controlled pilot access
- Access request: Evidence-readiness assessment, pricing, or guided pilot qualification.
- Pilot scope: One regulated-credit workflow, evidence question, data boundary, and deployment path.
- Commercial terms: Finalized during qualification and may be delivered through a Private Offer or agreed private handoff.
Annual and volume licence structures are post-qualification options, not public cold-sell offers.
Governance evidence
Materials and controls designed to support security review, vendor due diligence, and regulatory dialogue. The EU AI Act applies to high-risk credit scoring from 2 August 2026; these mappings support that preparation. This is not legal advice.
FL-BSA is a simulation and evidence-packaging tool. It does not certify regulatory compliance or replace legal, compliance, or model-risk judgment.
Product scope
- Self-hosted appliance: Guided AWS deployment / Docker deployment.
- Data boundary: No borrower, model, or protected-attribute data sent to vendor SaaS in self-hosted deployments.
- Fairness metrics: Adverse Impact Ratio (AIR), selection-rate gap, statistical parity, and equal opportunity/odds, per protected group and per branch.
- Evidence bundle: PDF report, evidence manifest (SHA-256), certificate appendix.
Controls & mapping
- Control mapping: FCA, EU AI Act, and ECOA touchpoints documented.
- Security posture: Hardened images, optional Prometheus metrics.
- Change management: Versioned releases.
Licensing & usage
- Usage terms: Finalized during pilot qualification and can be provided through a Private Offer or agreed private handoff path.
- Marketplace access: Via Private Offer or agreed private handoff once qualified.
- SLA and services: Custom SLA terms and professional services statements of work are executed separately.
Open-source components are catalogued at /legal/.
Documentation
Public materials for evaluation, plus licensed technical docs for implementation teams.
Whitepaper
Public demo whitepaper covering FL-BSA's regulated-credit evidence model. The intake ZIP is an unsigned, non-evidence-grade diagnostics bundle with aggregate synthetic-data and fairness inputs used for the public whitepaper build. See the whitepaper page for provenance and verification steps.
Example report
Pinned public demo report generated from the balanced scenario in the public GOLD run fixture. The ZIP is the supporting synthetic/demo GOLD run pack: balanced scenario artifacts, companion validation scenarios, row-level synthetic demo datasets, certificates, manifests, and evidence guides.
Evidence-readiness note
Five practical checks to make before testing one credit workflow for fair outcomes: decision scope, the evidence question, data boundaries, named owners, and action criteria.
Public demo release: v5.0.0-rc9-public-fix-2724455 (non-commercial prerelease). The downloads above are selected synthetic/demo assets from that public release; they are not customer output, legal advice, or proof of a public Marketplace launch.
Technical docs
Licensed customers receive technical documentation, API references, and integration guides through our secure portal.
FAQ
Does any customer data leave our environment?
No. FL-BSA is self-hosted; all compute happens in your VPC or VM. It does not exfiltrate model parameters, raw data, or SCPD to Equilens-controlled services. Optional AWS Marketplace usage metering is enabled only by configuration.
How long does a typical run take?
Reference targets on a modern CPU: ~10k rows (full pipeline) typically in the ~20-25 minute range; ~100k rows (training) ≤ ~45 minutes. 1M+ rows are long-running capacity-planning scenarios. Actual performance depends on hardware, data complexity, and configuration.
Which regulatory frameworks are covered?
Evidence maps to ECOA/Reg B (US), EU AI Act, and FCA Consumer Duty. Certificate families include input validation, data profiling, hyperparameter tuning, model, generation process, synthetic quality/validation, and regulatory mapping.
How can FL-BSA be deployed?
Public AWS Marketplace access is not yet available. Contact Equilens for controlled guided-pilot access to the AMI-first AWS customer path, or agree a controlled container/VM deployment where appropriate. CPU-only and GPU-preferred profiles are supported.
What is included in the evidence bundle?
A PDF report, an evidence manifest (dataset hash, RNG seed, software version), and certificates. Artifacts are hash-linked for auditability and can be retrieved by Task ID.
Does FL-BSA test our actual production model?
Not in the current release. FL-BSA simulates decision patterns learned from your historical data; it does not load or execute your scorecard. We are building a production-targeted measurement mode for unsigned aggregate fairness measurements over customer-authored decision or score columns. It is not yet shipped.
Can we verify the evidence ourselves?
Yes. Evidence bundles are hash-linked (SHA-256) with a documented hashing spec shipped in the bundle. Where signatures are enabled, signed release manifests and vendor-authored evidence attestations can be verified against the published Equilens vendor trust root at /fl-bsa/trust-root.json. Direct-AMI/customer-local evidence may prove bundle consistency without proving vendor authorship. Verification is offline; no Equilens service is required.
Ready to scope readiness?
Request the buyer/procurement pack or a readiness conversation. It commits you to nothing beyond a written reply.