Quantitative analysis first. Intelligent reporting second.
Lynx assesses accounting quality and detects signs of manipulation and fraud using the methods forensic accountants, short sellers and accounting researchers use. It is not clever prompting or a skill bolted over a PDF: the detection is done by more than 200 models built in house on years of academic and practitioner research, and this page sets out what is computed, what is read, what the AI does, and what the output is not.
How a figure becomes a finding
Read
Every page of every filing is converted, scanned pages through OCR, and each table’s currency and scale is recorded as printed. Each document is identified: type, company, fiscal periods, where the primary statements sit.
Extract, with the page
89 line items are read for every fiscal year the filings cover. Each value keeps the page, the table and the quoted line it came from. Sign and scale are applied by code, never by the reader.
Check
Every published figure must pass the statement identities it takes part in. Values that were contested, or that a second independent reading disagreed with, are held back and reported as missing rather than used.
Compute
More than 200 models built in house, coded from the published definitions, compute 166 indicators in 14 categories from the checked figures. Each carries its sample size, and where a regression is involved, its residual degrees of freedom. Too little history yields no number.
Benchmark
Each indicator is compared with the company’s peer group, or with all companies where that group is too small, in the nearest fiscal year with a complete population. The comparison population and its size are recorded with every reading, and stated in words in the report.
Rate, or withhold
A risk index combines the benchmarked indicators into a rating from A+ to F per fiscal year. A full reading needs 60% of the weighted indicators computed; between 40% and 60% the rating is published with a reduced-reliability caution, to be read as indicative; below 40% it is withheld and the report says why and what would unblock it.
Interpret
Only now do models read: 23 readings of the computed indicators, red flags read from the filing’s own text with quotes, and a synthesis that must answer a judge arguing the opposite conclusion.
Verify, then publish
A separate check reads every numeric claim in the finished report against the figures and strikes what the numbers do not support. Coverage, confidence and a data-quality page are published beside the report.
What the AI does, and does not do
Models read, judge and write
They identify documents, locate figures on a page, read the text for red flags and quote it, interpret computed indicators, argue both sides of a conclusion, and write the report’s prose and the Research answers.
Models never produce a number in the report
Nothing written by AI calculates, moves, rounds or retypes a figure in the report. Every value in a table was read from a page and checked by code; every indicator, benchmark, coverage figure and rating was computed by code. A number the prose states is verified against those tables before publication, and struck if it does not match. Research can calculate what a question needs: every figure it calculates is computed in code, never written, and cites its inputs.
What is tested
More than 200 in-house models produce 166 indicators in 14 categories, grouped into 23 analyses. The definitions follow the published research; where the implementation departs from a paper, the departure is listed in the methodology note of every report.
| Category | What it looks for | Examples of indicators |
|---|---|---|
| Accruals management | Earnings that are not backed by cash | abnormal accruals, accrual volatility, accruals relative to net income |
| Smoothing and income | Results steadier or better than the business | smoothing measures, tax-driven results, non-recurring items |
| Real activities | Operations run to hit a number | abnormal production, discretionary spending, cost management |
| Asset quality | Value parked on the balance sheet | intangibles and goodwill movements, capitalised costs, soft assets |
| Cash quality | Operating cash that is not what it seems | non-cash items, working-capital releases, interest and tax payments |
| Working capital | Sales and stock out of step with cash | receivable and inventory days, provision rates |
| Gearing and credit | Pressure to present well | leverage, short-term financing, interest cover, distress composites |
| Governance | Incentives and oversight | audit fees and opinions, options, insider holdings, digit conformity |
| Growth, margins, investing, valuation | Where the story and the numbers diverge | growth against cash, margin moves against peers, capex patterns, market-implied measures |
8 red-flag typologies are also read from the text of the filings: revenue recognition, expense capitalisation, reserve manipulation, related parties, off-balance-sheet items, cash-flow classification, non-GAAP measures and going concern. Each finding quotes the passage and its page.
Benchmarks, ratings and their limits
Peer groups
The sector you choose selects a peer group of listed companies. Each indicator is compared with that group for the same fiscal year, and the size of the population is recorded with the reading.
Context, declared up front
Where the company is in its life, how it makes money, the accounting standard and declared events change how a reading is interpreted, never the reading itself, and only on the risk-raising side. Declared events are looked at harder, not explained away.
Withheld means withheld
A study with too little coverage publishes its figures, findings and a data-quality page, but no rating. A study that cannot find the core statements publishes only what would unblock it.