The check
A small program with one job, such as matching freight bills to contract rates. It reads a folder of exports on your computer and writes a worklist and a report.
Every check is plain code that applies written rules to a file you export. There is no AI model inside a check: the same file gives the same result every time. An AI assistant is optional, and it only reads the results.
A small program with one job, such as matching freight bills to contract rates. It reads a folder of exports on your computer and writes a worklist and a report.
Most dates, rates, thresholds and rules that change with the law or a contract are settings in a file you can read and edit. The few that are fixed in code are named on the product page. Every result lists the settings it used.
Your AI assistant can call a check as a tool (MCP). It sees the result, with people's names and identifiers masked in most products, and helps you work through it. Through a check it cannot pay, send, file or change anything.
When an assistant calls a check, the check reads your file locally and returns findings. In most products, people's names, account numbers, birth dates, addresses and similar fields are replaced with stable pseudonyms or removed before anything is returned, and some have a stricter mode that masks more. Fairbooks and Fairlease mask less today: tax IDs are masked, but people's names reach the model. Amounts, dates and codes stay in clear, because that is what the finding is about.
The results then go to whatever model your assistant uses. For sensitive files, use a local model or one covered by your own agreements. Masking is a design choice, not a legal determination, and we make no compliance claim. Each product page says field by field what is masked and what stays in clear.
Every product carries its own unit tests and scored eval cases, and a demo that runs on made-up files. On 21 Nov 2026 all 18 code repositories were downloaded fresh and their unit tests were run again on Node.js 22: 15,355 of 15,355 passed.
14 of the 17 products have also been through a final pass, on 19 Nov 2026. It adds a hostile pass that feeds the checks poisoned files and arguments, a 10,000-row speed test, the same cases in eight languages, and a release check run from a fresh copy with no network. Accuracy was then measured three ways on made-up files: the problems planted in each check's test files, the same files with every problem removed (anything reported there is a false alarm), and one-change copies that should each produce one finding. Last, the checks were run on public, real-world data, which found real bugs in most families; they were fixed before these numbers were taken.
For Fairbooks the final pass covers 7 of its 30 checks, and for Fairhaul 26 of its 30. Fairclaim, Fairlease and Fairstay have not had it yet, so they have no accuracy numbers, and their rows below say so.
| Product | Planted found | False alarms on clean rows | One-change tests | Public data runs |
|---|---|---|---|---|
| Fairhaul (26 of its 30 checks) | 174 / 174 | 0 on 197 | 46 / 46 | 2 |
| Fairpractice wave 1 | 207 / 207 | 0 on 2,297 | 78 / 78 | 3 |
| Fairpractice wave 2 | 177 / 177 | 0 on 407 | 86 / 86 | 4 |
| Fairwire | 270 / 270 | 0 on 424 | 66 / 66 | 4 |
| Fairbuild | 170 / 170 | 0 on 67 | 46 / 46 | 2 |
| Fairbrief | 187 / 187 | 0 on 204 | 20 / 20 | 2 |
| Fairtable | 217 / 217 | 0 on 490 | 64 / 64 | 2 |
| Fairshop | 158 / 158 | 0 on 380 | 82 / 82 | 5 |
| Fairmake | 230 / 230 | 0 on 440 | 84 / 84 | 2 |
| Fairpolicy | 122 / 122 | 0 on 442 | 79 / 79 | 4 |
| Fairdrive | 185 / 185 | 0 on 205 | 63 / 63 | 2 |
| Fairroom | 175 / 175 | 0 on 373 | 107 / 107 | 2 |
| Fairclinic | 180 / 180 | 0 on 468 | 96 / 96 | 4 |
| Fairgive | 185 / 185 | 0 on 575 | 90 / 90 | 3 |
| Fairbooks (7 of its 30 checks) | 44 / 44 | 0 on 82 | 18 / 18 | 4 |
| Fairclaim | Not yet measured this way. 865 of 865 unit tests and 498 of 498 eval cases pass. | |||
| Fairlease | Not yet measured this way. 932 of 932 unit tests and 594 of 594 eval cases pass. | |||
| Fairstay | Not yet measured this way. 858 of 858 unit tests and 512 of 512 eval cases pass. | |||
| All measured | 2,681 / 2,681 | 0 on 7,051 | 1,025 / 1,025 | 45 |
The test files and the checks were written by the same team, so these numbers show that the rules do what they are written to do. They do not yet say how often a check is right on a real business's files: that is measured at each setup.
The 14 measured products know 314 named export layouts from the systems they read, counted per product. 31 have been run on a real published file or a vendor's public sample. The rest were written from knowledge of each vendor's export and have not yet been tested on a customer's own file. That is why every setup starts by confirming the layout on your files, and why any column can be mapped by hand. Each product page lists its formats one by one.
Most rules that change with the law, a standard or a contract (rates, limits, deadlines, code lists) are dated settings, listed in each check's README with their source. You can edit them yourself in a plain settings file. The few fixed in code are named on the product page and change only in a new version. On a support plan we update them, and the export layouts, when they change, and tell you what changed.
Fair Agents is new. We are looking for our first design partners: a fixed-price setup on your own files, with your feedback shaping the formats and the rules.
The first three clients of each product get 40% off the setup. How the offer works.