The record
A toll road, wind farm or hospital built with borrowed money is sized from one large spreadsheet, the financial model, built by the project's backers or their adviser. Before the banks sign, an independent firm called a model auditor checks it. The auditor reads the formulas or rebuilds the results by another method, compares the model with the loan and project contracts and with tax and accounting rules, and sends back a numbered list of issues. The modeller fixes them and sends a new version. This repeats until the list is closed and the auditor signs a letter the lenders can rely on.
One case is therefore a chain: first version, first issue list, second version, second list, and so on to a final model and an opinion.
A real example, from a report a British health regulator published in its committee papers. In 2013 the firm Mazars reviewed two 1 revenue-forecast models for the regulator. It built its own shadow model and ran software tests. After the first review there were three 1 further rounds. In all the firm raised 68 1 comments, each placed in one of three 1 grades, from a possible material error down to information only, and each marked open or closed. None fell in the most serious grade. The opinion covered only the fourth 1 version. An audit for a project loan follows the same steps at larger scale.
Why it might matter for training AI
Vendors' pages show labs asking for work in spreadsheets and for fixes to real failures. Finding an unnamed error in a spreadsheet is something current systems do badly. A public test released in 2026 has a debugging section where the system must find and repair errors nobody has pointed out. The best system tested got 12.00% 2 of those tasks right, against 34.89% 2 across the whole test.
An audit file would supply the missing answer key: errors a professional confirmed, and the corrected version.
One of that test's authors is a data vendor 2, so vendors already make expert-built spreadsheet tasks. Nobody was found selling real audit files.
Who holds it and who may share it
A few firms do most of the work. Forvis Mazars says it has given opinions on more than 2,000 3 project models and employs over 100 3 model auditors. Its Sydney team was formerly an independent firm, bought in 2016 4. Operis says it audits over a hundred 5 models a year and backs each opinion with a liability cap of up to £20 million 5. BDO in Britain says it has supported over 2,500 6 projects.
The model belongs to the project's backers or the borrower. The issue list and working papers are the auditor's own product. The opinion is addressed to named lenders.
The published 2013 report states that it is confidential, is not to go beyond the addressees of the engagement letter, and may not be disclosed to others without the firm's written consent 1. That points to at least two consents for a real file, the auditor's and the owner's, with the lenders' confidentiality terms on top. No engagement letter for a project-loan audit was opened, so the exact terms were not established.
Who is already there
The 2026 spreadsheet test has 321 2 tasks and a debugging section built, its authors say, from real business data with errors marked by experts. Its models are company accounts with errors planted in them, not project-loan models with real findings.
A second test, published the same year, draws 172 7 working sequences from old corporate files; the best system passed 38.4% 7.
One data vendor publishes a finance question set and says its tasks are written by people from large investment banks 8.
The auditors themselves sell review software and training. None was found selling data.
What no public set holds is a real project-loan model with its real numbered findings and the corrected version. The wider idea, spreadsheet debugging with expert-confirmed errors, is taken.
What would have to be stripped
There is no personal information. The sensitive content is commercial: tariffs, margins and returns that can identify a project even with the names removed. A model rebuilt closely enough from a real deal to be recognised carries the same problem as the original.
What stands in the way
- The real file sits behind the owner's rights, the auditor's liability and the lenders' confidentiality terms.
- A version rebuilt by experts is lawful but is the same kind of goods vendors already sell, and labs mark down contrived tasks.
- The pool of people who have done this work is small, and most are employed by the firms that hold the files.
- A published low score invites every vendor to aim at it, so the gap may not last.