Open rules, paid receipts — the game theory of a public rulebook
We publish our tax rules and their tests, and we plan to charge for running them. A competitor can copy the files. Here is why we publish anyway, and what we will and won't charge for.
As at 12 November 2026.
We write tax law as rules in Logical English, publish the rules with their test cases, and pay practitioners to try to break them. We also intend to earn money when those rules are run. People ask the obvious question: if the rules are public, what stops a competitor copying them?
Nothing does. This post explains why we publish anyway, and the rules we have set for ourselves.
What a copier can take
A rule file is text. Its tests are text. The reasoner that runs them is published openly and maintained upstream. Anyone can download all three and get the same answers we do. That is the point of a deterministic rule: same facts, same file, same result, whoever runs it.
So the price of simply running a public rule falls towards the cost of computing it, which is close to nothing. We don't think that can be fought, and we don't intend to try.
It goes further. A practitioner's signature is recorded against a file's fingerprint, not against whoever hosts the file. When our files carry signatures, someone running an identical copy will be able to point at our registry and say, truthfully, that this exact file was signed. The review we paid for is a public good.
What a copier cannot take
- The receipt. Evidence that a particular answer came from a particular file, on a particular reasoner version, from particular inputs, on a particular date.
- Standing behind the run. Someone accountable for the run having been computed as the file specifies.
- Currency. Knowing that a rule changed, when, and which earlier answers it affects.
- The record. Who reviewed which file, what they found, and what was fixed. A changed copy has a new fingerprint and starts with no review at all.
- The names. Our tier labels and our reviewers' names are not ours to lend and not anyone else's to borrow.
A practitioner does not need a number. They can get a number anywhere. They need to be able to show, later, how they satisfied themselves that the number was right.
The moves, and what we do about each
| Someone else's move | What happens | What we do |
|---|---|---|
| Hosts our files unchanged and charges less | Same answers, without the receipt | Count it as adoption |
| Changes the rules | New fingerprint; no review carries over | Nothing. They start from the bottom tier |
| Embeds the rules in a platform for free | Running becomes free everywhere | Supply the reviewed, receipted version |
| Copies once and never updates | Stale answers after the law changes | Keep change and defect notices public |
| Claims "verified" with no review | Buyers cannot tell the difference | Publish the test scenarios, and never overclaim our own status |
| Pays our reviewers more | Reviewers work for others too | No exclusivity. Reviewers are free agents |
The rules we hold ourselves to
- Rules, tests and defect notices stay public. We will not close something after opening it.
- We charge only for what we alone supply: the receipted run, the accountability for it, and telling you what a change means for your past answers.
- We do not charge on the tax outcome. A run costs the same whatever the answer is.
- We do not claim a status we have not earned. A file with findings and no signature is "under review", and says so on every answer.
- We do not lock reviewers in.
Our intention is to charge by status: nothing for a run on a file nobody has signed, and a price for a run on a reviewed and signed file. That is a design, not a price list. No payment rail is live today.
One thing is not final, and we would rather say so: the licence on the rule files and tests. Until it is, an interim notice says what you may and may not do, and a lawyer is being asked to settle the rest under Bounty #16.
A challenge, not a claim
If you are building AI for Australian tax, we would like you to joust with us. We are opening the Division 36 challenge on our loss-sequencing rules.
There are two ways in. Find a case inside the stated scope where our rules give the wrong figure. Or run our published scenarios through your own system and publish what it says, including the cases that separate an engine from a language model: a fact that is missing, and a question that is out of scope.
One condition applies to both: no guessing. A challenge states the facts, the figure, and the provision that produces it. An opinion that our answer “looks wrong” is not a challenge. A derivation is.
We expect to lose some of these. Round one of our own review returned five verdicts of “fix” out of five, and we rebuilt the file. That is the process working.
Where this is heading: a shared marketplace
The long-term aim is a marketplace of rules, not a product with our name on every answer. Tax and accounting law, written as rules anyone can read, each file with its tests, its review record and its fingerprint, published by whoever is best placed to write it, in any jurisdiction that has its own law and its own register of practitioners.
That is why the challenge is an invitation and not a dare. If you are building AI for tax, you do not have to choose between copying our files and competing with them. There is a third option: publish your own rule files in the same form, have practitioners review and sign them, and list them in the same registry. Your product stays yours. Your rules earn their standing the same way ours do, and a practitioner who reviews for you is free to review for us.
A marketplace needs more than one seller. We would rather share a large one than own a small one.
What exists today is the first part: published rules, public tests and paid review. The registry for outside publishers, the listing for hosts and the metered interface are designed, not built. Ways to take part sets out each role and what it gets today, and From bounties to ownership describes the intent. Logical English itself is maintained upstream; it is not ours, and anyone can write in it.
For practitioners
Reviewing a rule against the Act is technical work, and it is paid: a flat fee whatever your verdict, and a bonus for each accepted defect. We are also adding a record of the activity that you may be able to put towards your professional education hours. Until our review portal issues it automatically, ask and we will send one. The details are on Your review time and CPD.
Where we are
As at 12 November 2026:
- Sixteen bounties published. Ten calculator reviews are paid and reviewed.
- Division 36 loss sequencing. Round one: five verdicts, all "fix". Revision 2: one of three verdicts in, "fix" with eight findings. Revision 3 was built in response on 9 November: 48 scenarios, 321 tests, all passing.
- CGT event A1. 114 rules, 62 scenarios, 163 tests, all passing. Open for review in four family-sized pieces. No verdicts yet.
- Signatures: none. Every rule file is "under review".
- Tax Genii runs the rules in test, with that label on each answer.
- Payments for runs: not live.
Where we are going
- A review portal, so reviewers claim, review and sign on web pages and no longer send files by email.
- A first signature, then a second from a different practitioner, on a fixed file.
- A public profile for each reviewer, built from the registry.
- A record of CPD activity with every accepted review.
- A second jurisdiction.
- A metered, receipted interface other software can call.
- A marketplace: rule files from other publishers and other jurisdictions, in the same registry, on the same terms.
Whether paying a signer for each use of a file is the right legal shape is with a lawyer, as Bounty #15. Nothing here is legal advice, and nothing here is an offer.
The rules, their tests and the review record are at lodgeit.org/bounties. The earlier post Logical English — statute is code explains the medium.