Paddock

A Champions Meeting preparation tool for Uma Musume Pretty Derby — Global.

Everything on the site is scoped to what is actually live on the Global server, and every name uses the Global (EN) wording.

What it does

Planner — set the race: racecourse, direction, distance and surface, running style, going, weather, season and field size. Optionally set who is running it — a specific umamusume, her unique at the level you actually have it, her aptitudes, and the skills you expect to finish training with. Then set the rest of the field: either a headcount per running style, or every rival built out individually. It returns:

• a course profile — corners, straights, slopes, phase boundaries (the last-spurt phase starts at ⅚, not ⅔), the final corner, the home straight, where your last spurt starts, and every stretch where acceleration actually pays
• the Stamina you need, solved from the HP model for a full-length last spurt on that exact course, going and running style — plus how much of the last spurt your current stats cover, your stamina pool and an estimated finishing time
• stat targets, and what the next 100 points are worth in each stat, measured by re-running the race with 100 more of it and converting the time saved into lengths. That same number prices every green skill
• the field model: where your running style sits in this field mix at each of the four phases, so a skill gated on placing is priced at the phase it fires in rather than at the finish
• the pace-up (kakari) risk your Wit and style aptitude buy you, and what it is expected to cost in lengths on this course
• what your own run is worth — every skill you plan to have, valued on this race as a deck rather than as a pile of independent skills, with the unique scaled to its level and each row saying what it is worth alone and what it is worth next to everything above it
• the best deck for an SP budget you set, picked by marginal lengths per SP
• every skill ranked for that race, split into Speed & accel, Recovery, Green / passive, Debuffs and Positioning, sortable by lengths or by lengths per 100 SP, with every factor printed next to it
• the best uniques that can fire with that running style — scored under the style you picked rather than the one their owner prefers, each with the same skill read under all four styles, and gated on whether the owner can run this race at all after the aptitude sparks set in the rail
• the best unique to inherit from a parent, ranked as the weaker white copy you would actually be carrying, on your own runner's style and aptitudes
• the best umamusume for this race, added up from her unique, her own skill list, and what her aptitudes cost against the clock and the activation roll — her unique priced against the skills you said you are training, so one gated on your own skill history is counted off that list rather than guessed at
• the support cards that hand out the top-ranked skills

Race sim — the whole field run forward at 1/15 s, a few hundred times. Your win rate, top-3 rate, average margin, place spread and stamina at the line, for you and for every rival. A step-by-step replay of one race showing every runner's gap to the leader with your skill activations marked on it. And a check against the ranking: pick your skills and the app runs the field with each one and without it, on identical seeds, and prints what the simulation says next to what the ranking said — with a 95% confidence interval on every row, so a result the sample size cannot separate from zero is labelled noise rather than ranked. Because every variant runs on the same seeds, that interval is the paired one and is far tighter than the gap between two independent win rates.

It also answers the question a Champions Meeting actually asks. The field-shape sweep races the same build against five plausible field compositions on identical seeds and reports the spread: a build that wins 80% against a closer-heavy field and 25% against a front-heavy one is a worse pick than a flatter one, and the worst case is the number to plan against when the round is not announced yet. Build against build does the same for two saved builds, head to head, with the interval that says whether the difference is real.

Courses — every course on the Global server read against one build: what your kit is worth there priced as a deck, which of the four running styles each course wants, the Stamina it asks for at this going, how much of the last spurt your stats can pay for, and the pace-up risk. Sort by fit, by stamina need, by distance or by track. It is the inverse of the Planner: not what wins on this course but where does this runner belong.

Team — the actual Champions Meeting entry: three umamusume, a six-card deck each, all on one page and all planned against the course above. Per uma: aptitude check for this course, running-style override, stats, its own stamina requirement, the complete pool of skills that run can end with (its unique, its own skill list, guaranteed card events and card hints), and what each of those is worth in lengths. Plus a priority skill list you write yourself — every deck is then scored on how much of it you cover, what is missing, and for anything covered, which rank closed it and off which card. Team level: style spread, total expected lengths, and which cards you are running in more than one deck.

The priority list holds one entry per skill group, so aiming at both ranks of the same skill cannot be counted as two goals; picking another rank moves the target rather than adding a second row. A better rank always satisfies an entry — finishing with Right-Handed ◎ when you asked for Right-Handed ○ is not a miss. A weaker rank counts only if you tick it. The × rank never counts: it lives in the same group but has the opposite effect, so a card handing out Corner Recovery × does not cover Corner Recovery ○.

Cards that list the same skill both as their event skill and as a hint — 72 of the Global ones do — are counted once, at the event's full weight, rather than twice.

Both pickers are ranked, not alphabetical. Umas are sorted by what their own unique and skill list is worth on this course, discounted for missing aptitude. Support cards are sorted by what they would add to that specific deck — priority skills first, then expected lengths — and every card is listed with all of its skills, tagged event or hint, with priority hits ringed and skills the deck already has dimmed out.

A What to fix next panel reads the whole entry and says what to do: stamina shortfalls with the number of points or the percentage of recovery that would close them, aptitudes below A, empty slots, unreachable priority skills together with the card that would cover the most of them, which stat the next 100 points should go into, and whether the three decks have collapsed into the same list.

Builds are kept automatically between visits, and can be saved by name and switched between.

Collection — tick the umas and support cards you actually own. With the restriction on, a deck is five of your own cards plus one borrowed from a friend. That borrowed card is not pinned to a fixed position — any of the six may be it, there just cannot be two. Cards outside the collection therefore stay visible in the picker, badged friend, and are ranked by what they would add like everything else; once the borrow is spent they remain listed but disabled, so it is clear what taking the friend's card cost. An All / Mine / Friend's switch narrows the list either way.

Click any skill anywhere in the app for the full breakdown: what it does, the exact condition from the game data, its complete valuation on the current course, and every uma and support card that can give it to you — split into unique, uma skill list, guaranteed card event and card hint.

Umas — all 64 Global umas / 97 outfits, each with its full skill list (unique, gold and normal kept apart), the 10-way aptitude grid and running style. Filter by minimum aptitude in any of the ten categories, by running style, by rarity, and by which skills they have. Sort by any aptitude, rarity, gold-skill count or total skill score.

Support cards — every Global card indexed by the skills it teaches, with each skill tagged event (guaranteed, from the card's training event) or hint (from a random hint click). Filter by one skill or several, require all or any, and restrict the match to event-only or hint-only.

Also match the other rank — on by default. Search for a gold skill and cards that only carry its normal version are matched too: Determined Descent ⇄ Straight Descent. This works off the skill group id in the game data, so it covers every rank pair, including the ◎/○/× families. It is a search filter, deliberately symmetric and deliberately including ×, which is not how the Team page's priority coverage treats ranks — that one only ever counts equal-or-better, and never ×.

Skills — all 591 Global skills as a sortable table: effect, the real activation condition rendered from the raw game expression, SP cost, score, score-per-SP, duration, and where the skill can be obtained. Filter by rank, effect type, race phase, running style, distance, surface, terrain and source.

Data

What	Where from
Ids, names, aptitudes, skill conditions and effects, course geometry	Dump of the Global client master database, published by alpha123/uma-tools (umalator-global/)
Release dates, canonical Global naming	GameTora, best-effort

Because the dump is generated from the Global client and not the Japanese one, anything in it is by definition already released on Global. That is what the release filter is built on:

• skills and courses need no filtering — they come from the Global dump directly;
• support cards are published as one combined list, so a card counts as Global when every skill it teaches exists in the Global skill set. Cards that pass but sit past the current release frontier are flagged unverified instead of trusted silently;
• the Global client ships card data slightly ahead of the banner, so an outfit can appear before it is playable.

data-overrides/supports.json and data-overrides/characters.json pin any card or outfit either way and win over everything else; GameTora's release dates take priority over the inference whenever that pass succeeds. If something shows up that is not on Global yet, those two files are where to correct it.

.github/workflows/refresh-data.yml re-runs the whole pipeline daily and commits only when something changed. The Data tab in the site reports the last rebuild, the counts and whether the GameTora pass applied.

Rebuilding locally

npm run build:data              # fetch + rebuild docs/data/*.json
npm run build:data:offline      # same, skipping the GameTora pass

# icons (needs a checkout of alpha123/uma-tools and Pillow)
python3 scripts/build_icons.py /path/to/uma-tools

npm run serve                   # http://localhost:8080 (no-store, so edits show up)

The site is plain static files — ES modules, no bundler, no dependencies.

Hosting on GitHub Pages

The whole site lives in docs/. Two ways to publish it:

• Deploy from a branch (simplest): Settings → Pages → Source: Deploy from a branch → pick the branch, folder /docs. No workflow needed, and it works from any branch — it does not have to be the default one.
• GitHub Actions (what .github/workflows/pages.yml does): Settings → Pages → Source: GitHub Actions. That workflow deploys on every push to main touching docs/, so it only fires once main exists.

Either way the site ends up at https://<user>.github.io/umatool/.

If the folder is left on / (root) instead of /docs, the root index.html in this repository redirects to docs/ so the site still opens — but /docs is the correct setting and avoids the extra hop. The .nojekyll files stop Pages from running the content through Jekyll, which it has no reason to do here.

How skills are valued

There are two engines, and they agree.

The ranking (docs/assets/js/model.mjs) is closed form, because it is asked about 600 skills every time a slider moves. It reads each skill's real condition string, intersects it with the actual course geometry, and works out expected ground gained on the field.

The simulator (docs/assets/js/race/) runs all nine runners forward at 1/15 s with the real conditions checked per tick. It is the reference. Over a basket of thirty-odd skills on Kyoto 2200 m the two correlate at r ≈ 0.94, with the closed form running about 15 % high in absolute terms. Any row can be re-checked against the simulator from the Race page, which prints both numbers side by side.

Underneath both of them the physics is checked against a third thing: an independent transcription of RaceSolver.ts and HpPolicy.ts from alpha123/uma-skill-tools. npm run verify:model runs every course, running style, going and a spread of aptitudes through both and fails on any drift — 5712 combinations, currently zero.

Three things the ranking gets right that a plain “m/s × seconds” score does not, and which is why it used to disagree with every published Champions Meeting list:

• Acceleration only pays on a ramp. A +0.4 m/s² skill fired on the back straight, where you are already at target speed, is worth almost nothing; the same skill fired into the last-spurt ramp is one of the best things you can carry. The model finds every stretch where you are below target speed — the gate, the phase steps, the top of each hill, the run into the last spurt — and asks whether the skill can reach one.
• Debuffs count. Roughly a fifth of the obtainable skill pool does nothing to you and something to everyone else. Scored as “metres gained” they come out at zero and vanish from the list; scored as ground gained on the field they do not. Slowing the runners ahead of you is worth more per head than slowing the whole field, and slowing the ones behind you is worth almost nothing.
• Green skills are stat changes, priced from the same finite difference as the stat table — so a racecourse ○ is two lengths when the last spurt is short and exactly nothing when it is already paid for, and Sunny Days ○ is worth nothing in the rain because weather is now a hard gate.

And one thing a per-skill score cannot do at all: a deck is not the sum of its skills. Skills compete for finite things — there is one stamina hole, each ramp is one climb, and a stat only crosses a regime boundary once. Four recovery skills scored independently on Tokyo 2400 m at 900 Stamina come to 2.74 lengths; scored in sequence they are 1.40, and the fourth is worth exactly nothing. Four acceleration skills that each reach the run-up into the last spurt come to 3.63 independently and 2.89 as a deck, because that ramp is 6.97 seconds long and there is only one of it. Every total on the site is now a deck valuation: skills are scored greedily, each one against the state the ones above it leave behind, and each row reports both what it is worth alone and what it is worth in place. Greedy is exact rather than approximate here, because both resources are convex — the next skill onto a ramp always saves less than the last, so no later pick can overtake an earlier one.

Accuracy notes

• Race conditions are hard gates. Running style, distance band, surface, handedness, track, going, weather and season are read from the skill's own condition string. Fail one and the skill is dropped, not discounted.
• Aptitudes are modelled. Distance aptitude scales the Speed term in the final leg; surface and running-style aptitude scale acceleration. Below A costs real time and the model charges for it.
• Aptitudes are also sparked. The card's aptitude grid is not the grid an uma runs a Champions Meeting on — inheritance factors raise distance, surface and running-style aptitude, and running style is the one everybody fixes. The rail has an Aptitude sparks switch: Card reads the grid exactly as shipped, Style (the default) assumes the running style is factored up to A, All to A assumes distance and surface as well. Reading the card as final is not a conservative choice, it is a wrong one: on Kyoto 2200m as an End Closer it cut 57 of 97 outfits out of the ranking on a grade that costs one factor line to change, including the outfit the round is actually about.
• A count gate is read off your deck, not off a table. A handful of skills are gated on your own skill history — Festive Miracle wants three recovery skills to have fired, Dazzl'n ♪ Diver two middle-leg ones. Those used to get a flat entry in the odds table (activate_count_heal>=3 was a hard 15%), which prices a unique you deliberately build a deck around as though the deck were a coincidence. They are now counted: every candidate in the planned deck contributes its own activation probability and the answer is the Poisson-binomial P(N ≥ v), with a contributor that fires after the gate is checked counting at half weight. On Kyoto 2200m with seven recovery skills the closed form puts Festive Miracle at 85%; the simulator fires it in 88% of 300 races, at 1151m against the closed form's 1100m.
• …and Best umamusume ranks the gate as met, because that panel is read before the deck exists. Counting the gate is right for a deck you have written down and wrong for the panel you use to decide who to build: an empty skill list makes the gate impossible, which scored Christmas Oguri Cap at 0.05 and dropped her to 65th — a different wrong answer from the 15% it replaced, and one no amount of changing the race conditions can escape. So that one ranking is on the gate-satisfied value, marked ?, with the gate written out in the row (“only if your deck fires 3 recovery skills first, and the list you have planned holds 2”) and the as-planned number printed under it. Two umamusume in the whole Global set ever reach a top 15 that way, and both say why. Everywhere a concrete deck exists — the deck valuation, the SP optimiser, the skill ranking, Team — the counted number is the one used.
• Position keep is why early skills are cheap. For the first two thirds of the race every runner behind the leader is holding a slot, so ground stolen there is largely handed back. Measured against the simulator, an early speed skill keeps about half its nominal value and a last-spurt one keeps all of it. It is also why a front runner needs the field to be stamina-tight to win: with stamina to spare the closers simply out-spurt it, and the Race page shows that happening.
• Unique skill level. The dump ships uniques at their base value, so level 1 is the game's own number. Levels above that are applied as +10 % of base each — the community reading — and everything that uses it says which level it used.
• Stat ceilings are not hardcoded. Set the cap your scenario gives you and the inputs and target ranges follow it.
• Recovery is scored against how tight your stamina actually is. Once the last spurt is fully paid for, healing buys almost nothing and drops down the ranking by itself — and inside a deck, each heal is priced against the hole the ones before it already filled, so the third one is charged what it is really worth.
• Position keep in the opening leg is a per-style nudge, not the game's four-mode state machine. This is the largest remaining fidelity gap and it is deliberately still open: the reference implementation is not vendored here, and verify:model covers the speed, acceleration and HP formulas rather than position keep, so a hand-written state machine could not be checked against anything. A calibrated approximation that is disclosed beats an unvalidated guess that is not.
• The closed form runs high on uniques, and worst in the middle leg. The 15% quoted above is measured over a basket of ordinary skills. Uniques are long-duration and large, and on Kyoto 2200m as an End Closer six of them come out at 0.99–1.99× what the simulator pays, averaging about 1.5×. The overshoot tracks where the skill fires: positionWeight charges only 5% against a skill landing anywhere between ⅙ and ⅔ of the race, while the position-keep note above says ground stolen there is largely handed back. A skill firing at the halfway mark would need a weight near 0.58 to match the simulator, not the 0.95 it gets. Relative order inside the last spurt is fine; a middle-leg skill ranked against a last-spurt one is not, and the Race page's per-row check against the simulator is the way to settle any row that matters.
• What the simulator does not model is listed on the Race page itself: lanes are one-dimensional, skill ordering inside a tick is list order, position keep is a per-style nudge rather than the full four-mode state machine, and a handful of conditions (post number, favourite, lane side) are rolled at a fixed probability and named on the skill rather than simulated.
• Support card training effects (friendship bonus, specialty rate, stat gains), character growth bonuses and training event choices are not in the dump this build reads, so they are not shown.

Credits

Game data belongs to Cygames. This is an unaffiliated fan tool. Master-database dump by alpha123/uma-tools; release data cross-checked against GameTora.