Accounting quality, manipulation and fraud detection

Accounting risk, measured. Not guessed. In hours, not weeks.

See how it works

How a study runs

“AI-powered fraud detection” usually means a document in and a summary out, or a clever prompt with a skill bolted on. Lynx is built the other way round: more than 200 models built in house on years of academic and practitioner research run first, as code; generative AI only explains what they found.

It applies the methods forensic accountants, short sellers and accounting researchers use to assess accounting quality and detect manipulation and fraud: it reads your filings, extracts every figure with its page, runs those models to compute 166 indicators, benchmarks them against the company’s peers, and only then writes a cited report you can question.

No card, no passwordPDF (native or scanned), Excel, Word, PowerPoint, images, XBRL and MarkdownEvery number computed, never generated
Atlantis Semiconductor Ltd · Annual report 2025 · sampleAnalysing
Revenue/Sales4,182 p.2
Gross Profit1,081 p.2
Finance Costs84 p.2
Earnings Per Share0.42 p.2
166 indicators · 14 categories
Rating B-
Elevated in the closing yearFY2025 rating · risk index 0.38 · score quality sufficient · coverage 71% · 4 findings rated high
AI-generated analysis · figures computed, not writtenOpen the sample →
Hours, not weeks
A forensic engagement takes weeks of fieldwork; Lynx reads a filing in minutes to a couple of hours. It is the first read: an engagement, if one follows, starts from the questions already found.
200+ in-house models
Run on every study, producing 166 indicators in 14 categories, the same way every time. An engagement is a team on one company at a time.
Pay as you go
Credits you top up: a study fee for each completed study plus the tokens it uses, and Research by its tokens; a medium study is about $106, and you start with $100.00 free. An engagement carries a fee to match its weeks.

Who uses Lynx

Anyone whose decision rests on someone else’s financial statements. The study is the same; what changes is the question you bring to it.

Asset managers

Screen a watchlist or a portfolio each reporting season. Rank holdings by accounting risk, see what moved since last year, and take the page-cited findings into the investment committee.

Family offices

Diligence on a private deal from the CIM and management accounts you were sent, before the advisers are engaged. A rating, the findings and what the documents do not print.

Funds

Test a long thesis or build a short one. Every indicator against the peer group, restatement history laid bare, and Research to interrogate the filing line by line.

Regulators and supervisors

Decide which filings deserve a second look. Consistent, repeatable screening across issuers, with the reasoning and the page for each flag, and a data-quality reading for each study.

Also used by transaction-advisory practices, credit and counterparty-risk teams, forensic accountants and independent analysts.

Three finished reports, before you upload anything

Three finished studies on fictitious companies, published whole: one rated, one where the rating was withheld, one where the filings did not print enough to score. Every tab a real study has, page citations included.

Rated B-Rating B-

Atlantis Semiconductor Ltd FY2021–FY2025

Aggressive but disclosed choices, cash consistent with profit. Where judgement starts, and what a rating rests on.

Open the sample
WithheldWithheld

Avalon Telematics Ltd FY2020–FY2026

Too little of the analysis could be computed from these filings, so the rating is withheld and the report says which figures would restore it.

Open the sample
Insufficient dataNo rating

Lyonesse Logistics FY2024

A teaser that never prints the primary statements. What a study publishes when it cannot score, and what to upload instead.

Open the sample

How a study runs

Every number is computed by code and checked against the page it was read from. Models read, judge and write; they never produce a figure.

Drop the filings
Answer two questions
Lynx reads and benchmarks
Read, then ask

While the player has focus: Space plays or pauses, ← and → change scene. A chapter jumps to its scene.

Not a PDF dropped into a chatbot

Everything below runs before a single sentence of the report is written, and none of it is done by language models.

Every figure is read from a page and checked.
Each value carries the page it came from and the quoted line. Figures that fail the statement identities, or disagree on a second reading, are held back rather than used.
Every indicator is computed by code.
More than 200 models built in house compute 166 indicators in 14 categories, from accrual quality to digit conformity, each with its sample size and benchmark population recorded. Language models never produce, move or retype a number in the report.
Every claim in the prose is verified against the figures.
A separate check reads the finished report and strikes any sentence the numbers do not support, before you see it. A judge argues the opposite conclusion before the verdict stands.
Not clever prompting, not an agent’s “skill”.
The detection is done by more than 200 quantitative models built in house on decades of published research and practitioner method, not by asking language models what they think. If the filings do not print enough to support a rating, Lynx says so and withholds it. It never fills a gap with a guess.
See the full methodology
Report and Research · generative AI explainstwelve sections, every numeric claim verified · cited answers on demand, a figure calculated in code citing its inputs · the only layer language models write; they never produce a figure in the report
Interpretation · models read the numbers23 readings of the computed indicators · red flags read from the text with quotes · a synthesis that must answer a judge arguing the opposite
Quantitative analysis · code89 line items read with page citations and re-read · statement identities · 200+ in-house models · 166 indicators in 14 categories · peer benchmarks · coverage and confidence · a rating per fiscal year, withheld when the filings cannot support one

The methods, not a black box

More than 200 models built in house, from the published research on accounting quality, manipulation and fraud, the tests forensic accountants run, and the patterns short sellers write up. Their 166 indicators are grouped into 14 categories; each one you can open in your study.

Accruals and earnings quality

Discretionary and abnormal accruals, accrual quality and persistence, earnings smoothing, accruals against cash from operations.

Real-activities manipulation

Production, discretionary-spending and sales patterns that move reported margins without moving the business.

Distress and manipulation composites

The bankruptcy-risk, manipulation-detection and financial-strength composites, in the variants that fit listed and private companies.

Digit conformity

Whether the filed figures follow the leading-digit distribution that unmanipulated data follows, year by year, and where they depart from it.

Asset, cash and working-capital quality

Soft assets, goodwill movements, capitalised costs, non-cash items in operating cash flow, receivable and inventory days against peers.

Gearing, credit and governance

Leverage and refinancing pressure, interest cover, audit fees and opinions, options and insider signals.

Growth, margins and valuation

Where growth outruns cash, where margins move against the peer group, what the market price implies, and how capital spending follows.

8 red-flag typologies

Read from the text: revenue recognition, expense capitalisation, reserves, related parties, off-balance-sheet items, cash-flow classification, non-GAAP measures, going concern. Quoted with the page.

Every indicator is benchmarked against the company’s peer group, or with all companies where that group is too small, in the nearest fiscal year with a complete population, and every departure from a published definition is listed in the methodology note of each report.Methodology in full

What you get for every study

Every figure, with its page

Both the reported and the restated series, kept apart. Hover a cell for the quote; click for the page.

A rating per fiscal year

A+ to F on a benchmarked risk index, and withheld when the filings do not print enough to support one.

Findings you can verify

Severity, year, the quoted passage and the page. 8 red-flag typologies read from the text.

Research that cites

Ask across one study or several. Out of scope is refused visibly. Answers come with the page.

Pricing

Pay as you go. Add credits: a completed study draws a fixed study fee plus the tokens it uses, a Research question its tokens alone, shown to the cent before and after. New accounts start with $100.00 of free credits: enough for a study of a single filing and a dozen Research questions.

Pay as you go
$25.00per completed study
plus $4.00 per million tokens
  • $100.00 of free credits to start, no card
  • Add credits from $25.00; valid for 6 months
  • An estimate before every run, the exact cost after it; Research metered per question at the same token rate, no fee
  • Everything included; unlimited members in a workspace, one set of credits
  • A run that fails, stops before scoring or withholds its rating is not charged
Enterprise
Contact us
  • Invoice billing and committed volume rates
  • Single sign-on, dedicated data region, custom retention
  • API access and screening across many issuers
  • A named contact and a security review on your paper
Talk to us

Typical studies: a 22-page information memorandum about $44 · two annual reports, 150 pages, about $106 · six annual reports, 437 pages, about $245. A worked example