Multi-agent regulatory review, made auditable
An AI pipeline reads 86 published Solvency II SFCR reports and grades each one against a 158-point disclosure rulebook — grounded in every report's own text, cross-checked by two models, and rendered as the interactive dashboard below. A controlled accelerator with a full audit trail, not a black box.
Reading the rulebook so the reviewer doesn't have to
An SFCR is a long, structured disclosure with dozens of mandatory items across the Annex XX sections. Checking one by hand is slow; checking a market of them is a project. The rulebook becomes a machine-readable checklist and AI does the first pass — with the discipline an actuary expects: every verdict cites the page it came from, conditional rules are handled, and the quantitative templates are checked by code.
Two models grade the same checklist independently. When they agree, that's a strong signal. When they don't, the conflict isn't hidden behind an average — the dashboard flags it for a human. The point isn't to replace the reviewer; it's to hand them a triaged, cited, defensible starting point.
RAG-grounded rulebook
The 158 checks are derived from the binding texts — Solvency II Directive, Delegated Regulation, the ITS templates, EIOPA Guidelines and the French ACPR overlay. Every verdict is anchored in the report's own words with a page citation: retrieval over the regulation, not the model's memory.
Multi-agent cross-check
Two independent frontier models grade the same rulebook. Where they agree, that's consensus; where they genuinely disagree, it's surfaced as 'to review' rather than averaging the conflict away. Verification, not a single opinion.
Custom interactive dashboard
A bespoke, fully offline, zero-dependency dashboard: a market overview, per-report drill-down, category radars, and a gap list — built to the exact shape of the question, not a generic BI tool.
Actionable remediation
Each shortfall comes with a plain-language fix and the precise regulatory reference, so a finding is the start of a to-do list — not just a red flag.
What 11,465 sourced verdicts revealed
A text-detection benchmark across the public sample—not a legal conclusion about any entity.
The output is a review interface, not a verdict
Every proposed correction retains the chain from report wording to requirement, source, remediation and human decision.
Report text
The exact sentence and page supporting—or failing to support—the finding.
Requirement
The applicable control, article and conditional applicability rule.
Remediation
A grounded correction is proposed in language the reviewer can inspect.
Human decision
The reviewer accepts, edits or rejects; the tool never signs off.
69 verdicts were flagged—not averaged away
In the cross-checked subset, genuine disagreement was marked “to review” and excluded from the score. A plausible page reference was not accepted when it matched a template rather than the report text.
Explore the live demo
Fully interactive and anonymised. Use the market view to compare segments, then open a report to see its category profile and the cited gap list. The interface is in French — the reports are French SFCRs.
The same engine, pointed at your own work
This demo runs on public reports. The same approach runs privately on your material — as a repeatable, audit-trailed review step.
In-house disclosure QC
Run the checklist on your own SFCR / RSR / ORSA drafts before submission — a cited gap list and remediation list, on demand.
RAG over your corpus
A grounded assistant over your regulatory texts, internal policies and model documentation — answers with citations, not guesses.
Model-doc & migration review
Apply the same grounded, cross-checked review to model documentation and Python migrations — completeness and consistency, with an audit trail.
The rulebook is inspectable
The controls are derived from the applicable European disclosure framework and the French supervisory overlay.
Apply the same review discipline to your own draft
Run a defined, cited checklist over an SFCR, RSR, ORSA or another controlled document before submission.
Scope & honesty. This is an automated, text-based completeness review — a decision-support proxy, not a compliance audit or legal advice. Agreement between models indicates consistency, not proven accuracy. The two-model cross-check applies to the subset of reports graded by both models. All entities are anonymised for this public demo.