Ledger Proof: role-aware citation checking for AI-written financial documents

The general idea of checking every AI-written number against its source is taken. Kepler, Energent, Vouch, Finrep, V7 and the open-source FinGPT pipeline all do it for filings-based work, so the Research Note Linter would be entering a competitive field. Checking against US filings is the easiest part to copy, because XBRL facts already carry period and context. Kepler builds verification into its own drafting tool, and Finrep sells to issuers rather than analysts. The clearer opening is the UK wealth side, the Suitability Letter Verifier. The UK tools I found (Aveni, Fintel, Curvestone, AdvisoryAI and others) market completeness checks: COBS 9.4 sections, needs vs. suitability, vulnerability, and whether the fact-find was filled in. None of them say they check each figure's role: that a number comes from the right fact-find field, platform valuation, statement date and metric (contributions vs. growth, gross vs. net, last year's value vs. this year's). I found nothing on how PlannerPal itself checks figures. A new entrant could win as a vendor-neutral checking layer that sits on top of several drafting tools. It would read the source data that matters in the UK, such as platform and wrap statements and back-office systems like Intelliflo and Curo, which the US filings tools don't cover. It would also produce a figure-by-figure audit trail for Consumer Duty reviews. Risks: Drafting vendors (PlannerPal, Aveni, AdvisoryAI) could add figure checks themselves. Fintel and Curvestone already own the compliance-lead buyer and the file-review workflow. UK advice networks may see this as a feature rather than a separate product, so it may sell best as an API or partnership with those incumbents.

• Aveni (Aveni Assist / Detect)

  UK advice AI vendor that pitches compliance checks on AI-drafted suitability reports: checking the COBS 9.4 required sections, flagging gaps, and keeping review/approval audit trails tied to transcripts, CRM data and fact-finds. Its material focuses on whether required sections are present, not on checking that each figure comes from the right source.

• Fintel (threesixty / Simplybiz) AI file checking

  Large UK compliance-support network rolling out AI file checks to advice firms. It flags files for a human to look at. It already sells to the same compliance-lead buyer.

• Curvestone AI

  AI compliance file review covering suitability against needs, affordability, vulnerability and fact-find completeness, claiming 100% file coverage. Customers include Pivotal Group and White Rose Finance. It is mortgage-focused for now and its suitability letter product is marked as not live yet. It could move into wealth.

• AdvisoryAI / Recordsure / Automwrite

  UK tools for drafting suitability reports and checking them case by case. Drafting tools like these, and PlannerPal, could add their own figure-to-source checks, which is the main risk to a standalone verifier.

• Kepler

  Founded 2025 by ex-Palantir and Citadel people. A verifiable AI research platform that links every number to filing, page and line item. It uses a financial ontology to map terms like EBITDA and free cash flow to line items, and removes figures it cannot verify. Kepler claims it picks the right line item 94% of the time, against 38–46% for frontier models (its own figure). Customers are PE, hedge funds and banks. This is the closest match to the Research Note Linter, though it is built into its own drafting tool rather than checking third-party drafts.

• Energent AI Citation Checker / Energent Audit

  Free tool that checks AI-written claims against source files, traces each value to a file, row and field, recomputes derived figures, and gives pass/partial/fail verdicts. Energent Audit runs a separate agent as an independent auditor.

• Vouch API

  API that keeps checking AI-output claims against SEC EDGAR filings, consensus estimates and internal data, with certification and revocation of outputs.

• Finrep

  Checks AI-drafted financial commentary before publication. It matches each narrative number to the statements, recalculates percentage changes, traces non-GAAP figures to the reconciliation table, and checks period and restated-comparative references. That covers the period and GAAP-vs-adjusted checks, but it sells to issuers and CFOs, not research shops.

• FinGPT XBRL Validation Pipeline (open source)

  Open-source pipeline that checks numerical claims in an agent's answer against SEC XBRL facts. Each claim gets Verified, Mismatch, Skipped or Not Applicable, plus the filing used. Because XBRL facts carry period and context, wrong-period checks are cheap to copy for US filers.

• V7 Labs Financial Accuracy Agent

  Agent that reports discrepancies in financial documents, each backed by a visual citation.

• Caseware Validate / OnCorps

  Review tools for financial reports. They check consistency between tables, notes and narrative and flag prior-period variances. OnCorps cross-checks asset-manager reports against other data sources.

• FinGround / FISCAL (academic)

  2025–26 research on checking financial claims one at a time. Finds that general detectors miss 43% of calculation errors. FISCAL's data and scripts are open source, which makes it easier for others to enter.