KipScores — About

The one thing worth being clear about

We used AI to read the rulebook. We did not use AI to score your routine.

Those are two different jobs, and only one of them is a good use for a language model.

The scoring is arithmetic. Your routine and the ruleset go in, and a score comes out. Nothing else goes in: no server call, no stored state, no randomness, and no model in the path while you type — there is nothing anywhere in it that can form an opinion about your gymnast. The same routine scored twice gives the same answer, on any machine, on any day. The engine reads the date exactly once, to check that the ruleset is still inside the cycle it was written for; that check can decline to score, and it cannot change a score. When it takes 0.50 off for a Special Requirement, that is because a rule in the compiled ruleset says that requirement costs 0.50 when it is not met, and this routine did not meet it. You could do the same arithmetic by hand; it would take you an evening, and you would get the same number.

What AI did was the reading. USA Gymnastics publishes the event rules charts and the appendices as free PDFs. Turning those into structured data — rule by rule, each one tagged with the level and event it applies to, the document it came from, and how confident we are that we read it correctly — is a large, dull, careful transcription job. That is the part a machine that reads is genuinely good at, and that is the part we used it for.

So if you want to be sceptical of “built using AI” — and you should be — be sceptical of the right thing. Be sceptical of the transcription, because reading is where a machine can be wrong. The addition is just addition.

Which is why we then checked it. The system was built by referencing the published materials, and the parts a mistake would hide in were afterwards verified against the newly published Code of Points. Two in particular. The element numbering — the scheme that lets the engine read a difficulty letter straight off a Code element number — holds against the published list of elements whose values changed, every one of which carries a number that changed with it, and against the structure of the published charts themselves. The element combinations each level requires were checked against the Code the same way, and matched exactly.

That is a claim about a process, which is the only kind worth making here: we did the work, and then we checked it. It is not a claim that anybody approved the result.

And the engine is entirely ours. The scoring machinery — the engine itself, the compiler that builds the ruleset, the schemas, the element icons, the test suite — is AEXTech’s own work, end to end. What we read from the published charts are the rules. The machine that applies them, we built.

How a rulebook becomes data

This is the part we think is genuinely interesting, and it is the reason the numbers are worth anything at all.

The obvious way to build a scoring tool is to sit down and type one big rules file. It works, and it produces something nobody can ever audit. A number sits in the file, a coach disputes it, and there is no answer available except to go back to the PDF and hope.

So we did not do that. Internally the rulebook is a corpus rather than a configuration: one rule is one record, and each record carries the rule itself, what it applies to, which published document it was read from and where in that document to look, and how sure we are of the reading. For the 2026–2030 cycle that is 531 records covering optional Levels 6 to 10, and every single one of them carries its source. Three were read from a partial extract and are permanently marked as such; the rest came from full extracted text. Among them are 207 vault Start Values, the four Special Requirements for each of fifteen event–level combinations, and the 115 records behind the 200 rows of deduction table the app loads.

The file the app actually loads is compiled from those records by a script, and never edited by hand. That is the discipline the rest of it hangs on: if a number in the app has no record behind it, the build fails rather than shipping. Correcting a rule means correcting the record and rebuilding — so a correction can never quietly become a different sort of thing from the rule it corrected, and a convenient hand-tweak in the app’s rules file is not a shortcut, it is a bug by definition.

We call that shape an ergastorion — a workshop — and it is how this company builds every body of rules it touches, whether they come from a sports federation or a state legislature. The Code of Points turned out to be an unusually good test of the idea: it is published, versioned, effective-dated, internally cross-referenced, and it expires on a schedule someone else sets. Rules that expire on a published schedule are rules somebody has to keep current — and keeping them current, correctly and visibly, is the actual product.

What it says when it does not know

If you read one section of this page sceptically, make it this one. The rule the whole tool is built around is in its own configuration file:

Where the corpus cannot price something, this page says so in words rather than showing a zero. A zero means nothing was earned; silence means we do not know.

That is enforced, not aspirational. In practice:

• Start Value is a ceiling, never a prediction. Execution deductions belong to the judge on the day, and we do not model them. This page scores composition. It will not tell you what she will score.
• A requirement the engine cannot check is attested, and an unattested one is unknown. Of the 60 Special Requirements across the fifteen event–levels, 21 can be machine-checked from the routine data and 39 are ones you confirm yourself. Unknown is a third state: never assumed met, and never scored as missed.
• Where the published material gives a budget but no schedule, the answer is “cannot price” and the reason. The Level 10 Difficulty Value bonus is the live example — the chart states a total and does not say how to earn it element by element, so the page refuses rather than returning a zero you would reasonably read as “you earned nothing.” At Level 9, where the chart does state the schedule, it is computed.
• Three rules are flagged as needing a second reading — two Level 10 vault values whose numbers sat across a column break in the extracted text, and one summary-block rule whose scope below Level 9 we have not established. The flags are in the data and they travel through to the scorecard. We would rather you saw the flag than trusted the number.
• Some things are our reading of the rules rather than a transcription of them, and those are labelled separately all the way to the scorecard. The Code states its requirements in prose; turning a sentence into a test a computer can run is an interpretation, and a defensible interpretation can still be wrong. You should be able to tell which is which, so we keep them apart.

Coverage here is a list, not a percentage — what is in the corpus, what is not, and what the engine therefore will not price. There are six areas it knows it does not cover and it names all six. The coverage banner on the Check is generated from the ruleset itself rather than written by hand, so it cannot drift away from what the engine really does. We know that matters because an earlier hand-written version drifted inside a day.

What we can show you, and what we cannot yet

Every rule in our corpus carries its source — the published document, its revision stamp, and where in it to look. That is what makes a disputed number answerable at all, and it is why the corpus is built the way it is.

What is free, and what we expect to sell

The calculator is free, and it is meant to be. Applying published rules to a routine is education — it is what a coach already does on paper with the chart in front of her, and there is no good reason to charge anyone for arithmetic. Everything described on this page is the free part.

What we expect to sell is a different thing: optimisation strategies for an individual gymnast, checked against the published guidelines. The free tool answers what is this routine worth. The harder question is what should this gymnast change — given what she already has, what she is close to, and what the season looks like. That is where the real work is, and that is the part we think is worth paying for.

We are not promising it, and we are not going to describe features that do not exist. It is what we are building next.

Privacy

There is no account and no sign-up, because there is nothing to sign up to. A routine is labelled with a reference you choose — not a gymnast’s name — and is saved in your browser and nowhere else. The site is a static page and its ruleset, with no server behind it to send anything to. Nothing about a child’s training leaves the device.

Who makes this

KipScores is built by AEXTech Holdings. It came out of our work on keeping bodies of rules current as data — the same problem, in a domain where the rules happen to be published, dated, and genuinely used by people who need them to be right.