§ LENS · TAX · AVAILABLE
Where a group books its profit, and what it pays there.
Public country-by-country reports say where a multinational earned its money, how many people it employed, and what tax it accrued and paid, jurisdiction by jurisdiction. The tax lens reads those filings and asks where the three disagree.
What it reads
Country-by-country reports published under EU Directive 2021/2101, and Australian reports published under the voluntary Tax Transparency Code. Both are read from the filed PDF, with every figure cited back to the character span it came from.
Australia’s statutory public country-by-country regime begins reporting later, and AYVES reads those filings when they are published rather than before. The two are different regimes with different disclosure requirements, and the record says which one a figure came from.
Other formats are in build and are listed on this page when their fixtures pass, and not before.
What it computes
Six indicators, seven values, per jurisdiction.
- Effective tax rate, accrued
- Tax accrued against profit before tax.
- Effective tax rate, cash
- Tax actually paid against profit before tax.
- Profit per employee
- Profit against headcount, held against the median of this filing's own jurisdictions.
- Profit margin
- Profit against revenue.
- Revenue per employee
- Revenue against headcount, on the same median basis.
- Tangible assets to profit
- Profit against the physical assets reported as producing it.
- Related-party revenue share
- Intragroup revenue as a share of the total.
How it scores
Each indicator returns red, orange, green, or grey. Grey means the indicator could not be computed, and it never counts as a pass. A jurisdiction where most indicators are grey is recorded as under-evidenced rather than clean.
Each jurisdiction then rates critical, review, monitor, clean, or unrateable, from a deterministic rule over those flags. The rule has eleven outcomes, each with a machine-readable reason code that appears in the analysis, the CSV, and the PDF. Every one is documented, with worked examples, in the methodology.
No language model influences a rating. Models propose values, and the deterministic core disposes of them.
What it will not do
Each of the following would look sophisticated and produce unreliable output, so the lens refuses them by design.
- Score the quality of a group's transfer-pricing documentation.
- Infer whether a tax ruling exists.
- Analyse a treaty network.
- Benchmark against a peer pool below the safe-sample floor.
- Claim a year-on-year risk change without naming the mechanism behind it.
- Infer beneficial ownership.
- Quantify a regulatory penalty.
- Read sentiment from management commentary.
How it is tested
Every prompt change and every model change re-runs a gate over a corpus of hand-labelled filings, scored on six metrics: jurisdiction recall, spurious rows, financial error, employee error, schema pass rate, and format detection. The gate reports per format as well as overall, so a weak format cannot hide inside an average.
The corpus currently holds 0 labelled filings against the 15 the methodology requires. That gap is open and tracked. It is stated here because a reader is entitled to know the size of the evidence base behind a rating, including when it is smaller than intended.
Reading it yourself
Every figure carries the character span in the source document it was read from. Every rating carries its reason code. Every threshold cites the source it derives from, and where a threshold comes from AYVES calibration rather than a published number, the methodology says so in those words.