01
Classify
Determines query type, complexity, risk level, and which FCA sourcebooks apply before anything else runs.
Ref FG-001 · 2026
Built for FCA-regulated firms
The system that argues
with itself.
Ask a compliance question. One model generates the answer. Ten agents try to tear it apart. You only see what survives.
0
Pipeline Stages
0
Verification Agents
0+
Adversarial Tests
Scroll
9-Stage PipelineSentence-Level VerificationKnowledge Graph Retrieval10 Independent AgentsDebate-Based VerificationSM&CR Audit TrailsCitation Existence ChecksFCA Handbook CoverageTrust Score CalibrationEvidence Bundle Export9-Stage PipelineSentence-Level VerificationKnowledge Graph Retrieval10 Independent AgentsDebate-Based VerificationSM&CR Audit TrailsCitation Existence ChecksFCA Handbook CoverageTrust Score CalibrationEvidence Bundle Export
How it works
Nine stages between your question and the answer.
Four phases. Intake, generation, verification, output. Every answer passes through all nine stages before you see it.
Phase 01 · Intake2 stages
Understand the question and gather the evidence before any LLM speaks.
02
Retrieve
Expands your question into multiple phrasings. Searches vectors and the regulatory knowledge graph. Follows explicit cross-references between provisions and links to relevant enforcement notices.
Phase 02 · Generation1 stage
A first answer — treated as a hypothesis, not a verdict.
03
Generate
Produces the initial answer using retrieved evidence. The next six stages check whether that answer is actually correct.
Phase 03 · Verification3 stages
The hypothesis is broken into individual claims and challenged.
04
Verify
Ten agents check the answer independently with different evidence subsets. Sentence-by-sentence. If a cited section does not exist, the score gets hard-capped.
05
Analyse
Aggregates agent verdicts with weighted voting. Tier 1 agents carry 1.5× weight. Detects disagreement patterns and flags low-consensus claims.
06
Cite
Links every claim to its source. Confirms each cited FCA section exists, is current, and matches the rule type the answer implies.
Phase 04 · Output3 stages
The answer becomes a memo with a trust score that means something.
07
Calibrate
Adjusts raw confidence so the percentages mean something. If the score says 80%, the system should be right about 80% of the time. Calibration corrects for overconfidence.
08
Explain
Generates counterfactual explanations. Shows what would need to change to flip the verdict. Makes the reasoning auditable.
09
Score
Produces the final trust score from five weighted components: verification, dimensional analysis, citations, calibration, and retrieval quality.
Live sample
Here's what comes out the other end of the pipeline.
Four real questions. Pick one — see the answer, citations, judge verdict, and 10-agent breakdown. Same output you'll see signed in.
Question
What are the FCA SM&CR requirements for senior managers?
Judge · VerifiedSuper-agents · 9/10Trust · 86%Citation gate · strong3.14s
Answer
Senior Managers under the SM&CR are subject to several core requirements. They must be pre-approved by the FCA to hold a Senior Management Function (SMF) before performing the role (SUP 10C.10). Each Senior Manager must have a Statement of Responsibilities that clearly sets out their personal accountabilities, and the firm must maintain a Management Responsibilities Map showing how all responsibilities are allocated (SYSC 25). Senior Managers are subject to the Duty of Responsibility — they can be held personally accountable for misconduct in their area if they fail to take reasonable steps to prevent it. They must comply with the Senior Manager Conduct Rules (COCON 2.2), which include taking reasonable steps to ensure the firm complies with regulatory requirements, the firm has adequate systems and controls, and any delegation is to an appropriate person. Annual fitness and propriety certification is required (FIT 1.3), with a declaration to the FCA via Form J.
Citations (4)
FCA Handbook
93%SUP 10C.10
A firm must take reasonable care to ensure that no person performs an FCA-designated senior management function without prior approval.
FCA Handbook
91%SYSC 25.2.1R
A firm must ensure that, at all times, one or more Senior Managers have overall responsibility for each of the activities…
FCA Handbook
89%COCON 2.2
You must take reasonable steps to ensure that the business of the firm for which you are responsible complies with the relevant requirements…
FCA Handbook
87%FIT 1.3
A firm must annually certify that each Senior Manager is and remains fit and proper to perform their role.
Snapshot of real pipeline output. The full demo page shows the per-agent breakdown and judge reasoning.
Open full demoModules
15 modules, grouped by how you'll actually use them.
Every module has a production backend, API endpoints, and test coverage. No vapourware — if it's on this page, it ships in the build you sign up for.
Core workflow
5 modules
Used on every visit. The shortest path from "regulator question" to "memo I can paste into a reply."
Compliance Q&A
Ask anything about FCA rules in plain English. The answer passes through nine stages before you see it; every citation is checked against the actual handbook.
Decision Vault
Encrypted decision records for SM&CR compliance. SHA-256 signed approvals, escalation history, five-year retention. Generates certificates auditors can verify.
Evidence Bundle
Generates a ZIP ready for FCA inspection. Cover sheets, per-agent verification certificates, audit logs, decisions, firm compliance profile.
Pulse Engine
Monitors FCA, PRA and Bank of England feeds 24/7. Triages each publication by relevance to your firm. Storm alerts when something material lands.
Regulatory Genome
22,084 FCA provisions indexed and linked. Click any provision to see its cross-references, in-force dates, and enforcement history.
Verification
4 modules
What makes FINGAURD different — every answer goes through this layer before you see it.
Compliance Gateway
Checks AI-generated content before it reaches clients. Multi-stage analysis, data-boundary checks, PII scanning. Sits between your AI systems and your output.
Audit Simulator
Simulates a Section 166 review. Stress questions, vulnerability scan against enforcement actions, scoring with category breakdowns. Run the audit before it runs you.
Persona Auditor
Maps SM&CR roles to AI-agent behaviour. Detects when an agent drifts from its risk profile. Flags gaps before the regulator asks.
Enforcement Intelligence
Analyses the graph of FCA enforcement actions. Repeat offenders, co-violation patterns, fine trajectory. Tells you which rules the FCA is enforcing now.
Specialised
6 modules
Use when the scenario calls for it. Each has a working backend; not every customer needs every module.
Regulatory Forecast
Reads consultation papers and Dear CEO letters. Scores each against your firm type and tells you which changes will affect you.
Policy Generator
Drafts internal policies mapped to specific FCA provisions. Every statement links to the regulation it satisfies.
Code Scanner
Analyses source code against FCA compliance rules. Flags data handling, consent flows, and record-keeping gaps that standard code review misses.
Sandbox Simulator
Tests fintech products against sandbox acceptance criteria across multiple jurisdictions. Run the scenario before it is real.
Regulatory Arbitrage
Compares rules across jurisdictions. Where do UK and EU overlap, where do they conflict? Parallel queries across regulatory frameworks; friction scored.
Inclusion Scanner
Consumer Duty vulnerability assessment. Checks against twelve FCA vulnerable-customer categories for access, comprehension and pricing barriers.
Verification agents
Ten agents. None of them trust each other.
They run in parallel. The important ones carry more weight. If the two most critical agents (Citation Verifier · Temporal Validity) both fail, the score gets capped at 40% regardless of what the other eight say.
Weight tiers
Critical · 1.5× · 2 agents
High · 1.2× · 4 agents
Standard · 1.0× · 1 agent
Soft · 0.6× · 3 agents
01
Citation Verifier
1.5×Character-by-character match against the FCA Handbook.
If you cite SYSC 10.1.8R, this agent confirms that exact section exists — and says what you claim it says.
02
Temporal Validity
1.5×In-force, superseded, or pending — knows which.
Catches answers that cite revoked provisions as if they still applied.
03
Logical Consistency
1.2×Finds contradictions in the answer itself.
If the answer says "firms must" but cites Guidance rather than a Rule, that's a logical error. Guidance is not binding.
04
Cross-Jurisdiction
1.2×UK-only? EU? Both? Validates the regime.
Catches when UK-specific advice accidentally references EU law that no longer applies post-Brexit.
05
Completeness
1.2×Uses the Regulatory Genome to find missed obligations.
A correct but incomplete answer is still a problem in compliance.
06
Adversarial
1.2×Tries to break the answer. Six attack types.
Edge cases · outdated rules · bad assumptions · missing info · misinterpretation risks · scope problems.
07
Regulatory Intent
1.0×Does it match why the rule exists, not just what it says?
A technically correct answer that misses the point still gets flagged.
08
Plain Language
0.6×Readable by someone who is not a lawyer.
Flags jargon and ambiguous phrasing the compliance officer can't paste verbatim.
09
Precedent
0.6×Checks against how the FCA has actually enforced.
What the Handbook says and what the FCA does about it are sometimes different things.
10
Audit Trail
0.6×Records every step. Reproducible by design.
Who checked what, when, what they found. The verification process itself needs to be auditable.
Jurisdictions
UK FCA is the primary jurisdiction.
Structural coverage across the FCA Handbook sourcebooks. The multi-regime engine is built and tested; EU and UAE data ingestion is in progress, and FCA full-text coverage continues to expand under the Handbook licence work.
UK FCA
PrimaryPrimary jurisdiction. Structural coverage of the FCA Handbook across the listed sourcebooks; full-text ingestion expanding progressively as the Handbook licence agreement completes.
SYSCCOBSSUPMCOBICOBSBCOBSDISPPRINCONCCASSGENPRUBIPRUFEESGENMARSMCRWDPG
EU MiCA
Backend BuiltMarkets in Crypto-Assets Regulation. Cross-jurisdiction agent built and tested. Data ingestion in progress.
EU AI Act
Backend BuiltEU Artificial Intelligence Act compliance mapping. Multi-regime engine supports cross-referencing with FCA rules.
EU DORA
Backend BuiltDigital Operational Resilience Act. Backend agents ready. Regulatory genome supports DORA obligation extraction.
UAE VARA
Backend BuiltVirtual Assets Regulatory Authority. Sandbox simulator evaluates products against VARA acceptance criteria.
Contact us for EU/UAE access timeline.
Transparency
We show you how it works. And where it falls short.
I
How we verify
Four mechanisms run on every answer. None of them are optional. None of them are reversible by the LLM that wrote the answer.
- 01The answer is broken into individual claims. Each claim is classified and verified against source evidence independently.
- 02Each verification agent gets a different subset of evidence on purpose. The adversarial agent deliberately gets the weakest evidence — prevents groupthink.
- 03Deterministic checks confirm cited sections exist, are current, and have the right rule type. A fabricated citation caps the trust score regardless of what the agents say.
- 04Agents challenge each other through four rounds of structured debate. They can concede, rebut, or maintain. Fully deterministic and reproducible.
II
What we cannot do
Honest disclosure. The product would be vapourware without these four limits explicit.
- 01AI makes mistakes. Verification catches more than a single model, but it will not catch everything. The trust score tells you how much the agents agreed, not that the answer is definitely right.
- 02This is not legal advice. Always talk to a qualified advisor before making compliance decisions. Your team makes the final call.
- 03FINGAURD Ltd is not authorised or regulated by the FCA. We build technology tools, not regulated financial services.
- 04Confidence calibration improves with data. Early scores may be over- or under-confident until enough labelled queries have flowed through the system.
Security & data handling
ICO Registered
UK data-protection compliant
Encrypted in transit + at rest
TLS 1.2+ · AES-256
SCCs in place
International transfers covered
API data not used for training
OpenAI policy + our contract
Measured, not claimed
FCABench runs every release. Numbers, not promises.
Most compliance AI gives you marketing claims. We give you scores. Each release goes through the adversarial test suite below before any customer sees the update.
01
Adversarial test cases
31+
Pulled from real FCA enforcement notices, Dear CEO letters and Section 166 reviews — including deliberate hallucination traps.
02
Categories tested
18
Governance · conduct rules · Consumer Duty · market abuse · operational resilience · financial crime · CASS · MCOB · ICOBS · CONC · prudential · enforcement · temporal · cross-reference · hallucination · prin_2a · smcr_fit · SUP.
03
Metrics scored
5
Overall accuracy · citation recall · citation precision · hallucination detection rate · verdict accuracy. Every release. CI-verified.
Methodology
backend/app/core/fcabench.py· verified by CI on every commit to main.Why this exists
I watched compliance teams trust AI nobody was checking.
Single-model LLMs will confidently cite FCA rules that do not exist. They will tell you a revoked provision is still current. They will give you a technically correct answer that misses the three other rules that also apply.
Compliance officers started making decisions based on these outputs. No verification. No audit trail. No way to know what was fabricated. That is the problem I set out to solve.
The idea was simple: what if the system argued with itself before answering? That question turned into ten verification agents, a nine-stage pipeline, claim-level hallucination detection, and the first FCA-specific benchmark suite.
Every module on this site has a working backend. 3,669 tests pass. I do not put features on a roadmap and call them products.
MD
Madhu Dasari
Founder, FINGAURD Ltd
FINGAURD Ltd · Company No. 16310538 · ICO Registered
Pricing
One pipeline. Four price points. Cancel anytime.
Every plan runs the same 9-stage verification pipeline and the same 10 super-agents. Higher tiers unlock more modules, scheduled monitoring, and the MCP server. Prices exclude VAT.
Early Adopter
£250/mo
100 queries/month · £2.50/query
For solo compliance officers at small FCA firms.
- Full 9-stage verification pipeline
- 10 super-agent verdicts on every query
- 5 core modules — Q&A, Archive, Firm Lookup, Pulse, Insights
- FCA Register API live lookup
- Email support
Most popular
Starter
£1,000/mo
500 queries/month · £2.00/query
For small compliance teams (2–10 seats).
- Everything in Early Adopter
- All 10 modules including Decision Vault audit trails
- Citation hallucination gate with claim-level verification
- Inspector Pack export (JSON) for SM&CR record-keeping
- Watchers — re-run a query if its cited regulation changes
- Priority support (4-hour first-response SLA)
Professional
£2,500/mo
5,000 queries/month · £0.50/query
For mid-size firms with a dedicated compliance function.
- Everything in Starter
- All 14 modules — Risk Analytics, Audit Simulator, Genome, Graph
- Evidence Bundle generation for FCA inspection
- Bulk decisions + memo export
- Cross-firm anonymised benchmarking (coming soon)
- Dedicated Customer Success Manager
Enterprise
£5,000+/mo
No query cap
For asset managers, banks, and regulated platforms.
- Everything in Professional
- MCP server (10 compliance tools for your AI agents)
- Scheduled compliance monitoring + Storm alerts
- UK-hosted, fully isolated infrastructure option
- SSO (SAML / OIDC) · custom API · per-tenant SLA
- Quarterly business review with the founding team
Why per-query pricing? Each FINGAURD query runs through ~10 LLM calls (one per super-agent). The cost is real. Per-query pricing keeps the model aligned: we make money when you get answers, not when you log in.
Frequently asked
Questions compliance teams ask us.
01How is this different from ChatGPT with a compliance prompt?+
ChatGPT gives you one answer from one model with no verification. We run the answer through nine more stages after generation. One agent breaks it into individual sentences and verifies each one. Another confirms every cited FCA section actually exists. A third tries to break the answer on purpose. If cited sections do not exist, the trust score gets capped.
02Is this regulated by the FCA?+
No. FINGAURD is a technology tool. We do not provide financial advice or regulated services. Your compliance team makes the decisions. The system includes a guardrail that blocks any answer that crosses into giving legal advice.
03What happens when the AI gets it wrong?+
You see it. Each claim is classified as supported, unsupported, contradicted, unverifiable, or fabricated. If more than 30% of claims fail verification, the answer gets blocked entirely. The explainability engine shows what would need to change to flip the verdict.
04Can we use this for audit evidence?+
Yes. The Decision Vault stores encrypted, signed decision records with five-year retention. The Evidence Bundle generates a ZIP file with cover sheets, per-agent verification certificates, audit logs, and your compliance profile. Built for SM&CR record-keeping.
05Where is our data stored?+
Queries go through OpenAI API, which does not train on API data per their current policy. Frontend on Vercel, backend on dedicated UK infrastructure. The shared FCA Handbook knowledge base sits in Pinecone (1536-d embeddings); per-tenant data — audit logs, decisions, KYC cases — is partitioned by tenant_id in Postgres. Encrypted in transit (TLS 1.3) and at rest (AES-256). Enterprise plans offer fully isolated UK-hosted instances.
06What is the MCP server?+
Model Context Protocol. It exposes ten compliance tools that any AI agent can call directly: query regulations, verify answers, check firm status on the FCA Register, get enforcement history. Claude, GPT, LangChain, CrewAI, and AutoGen can all connect. Available on Enterprise plans.
Contact
Want to see it?
Try the platform directly, or request a 20-minute walkthrough of the live pipeline.
Or request a demo below and we'll reach out.