A rebuilt statistical engine, a rebuilt completer, and one slider fewer(everything that changed, and why)
You may have noticed that k2dex looks and feels much different than it did previously. This update touches nearly every part of it: the model picker, the completer, the analysis page, and the meta page. The short version is that the model underneath is now fit with a fundamentally better method, and almost every visible change follows from that. Here's the tour, starting with what you'll notice first.
The old picker offered four models: Species, Species @ Item (recommended), a weighted experimental variant, and an experimental Regulation M-B model. That menu is gone. There is now exactly one model per regulation, and it folds in everything the variants used to trade off against each other for a simplified and improved experience without any loss in model quality.
If you preferred the old species-only view, it didn't disappear: it became a toggle. In the completer's advanced options you can exclude the item attribute entirely, and results aggregate by Pokémon set. Crucially this is not a separate, coarser model: the item dimension is collapsed from the full model exactly, so the species-level answers are consistent with the item-level ones and the old species-only model.
The completer's input used to be a single search box of full builds like "Charizard @ Life Orb". It's now a proper six-slot roster editor. Each slot has a Pokémon picker and an optional item picker, and the item being optional is a new capability: leave it blank and you've locked in the Pokémon while letting the model choose the item as part of the completion. Set it and you've pinned the exact build, as before. The same editor now drives the analysis page, so moving a team between the two surfaces feels the same.
Constraints grew a second half. Alongside the existing exclude list ("never suggest these") there's now an include list: give it a pool of Pokémon and completions are restricted to that pool plus your pins. This new constraint is perfect for newer players who may be limited by the Pokémon they have recruited thus far, creating teams from the Pokémon they have available instead of the full regulation roster.
Results changed shape too. Instead of a ranked table of the completed members, each suggestion is a full six-Pokémon team card, with your pinned slots marked and the team's observables and actions underneath. The default stats strip now reports Variety (how different the shown teams are from each other) and Novelty (how far they sit from teams already in the tournament data), while the sampler diagnostics that used to occupy that space are still available behind a toggle. And anywhere a Score or Coherence value appears, hovering it shows the corpus percentile: how that number ranks against every real tournament team the model was trained on.
The old completer had a problem that power users will certainly recognize: pin a niche Pokémon and ask for the rest of the team, and you'd often get a perfectly coherent five-Pokémon core with your pin bolted on as a passenger. This was the mathematically likeliest team containing your pick, but not a team built around it.
The new Anchor Strength slider fixes this. It tilts the sampler toward teams whose members actually interact with your pins, by amplifying the couplings between pinned and free slots during the search. At the default of 1.0 nothing changes; slide it up and the suggestions become increasingly pin-centric. The math used to accomplish this retains the same level of scientific rigor that k2dex has always prided itself on.
Gone, on purpose. The slider existed because the old model needed correcting: run as fitted, it leaned too hard on individual popularity, so the app shipped with the bias term dialed down (0.5 by default) and let you tune it. The split between "raw" and "adjusted" scores existed for the same reason.
The new model doesn't need that correction. Its fitting procedure guarantees the model's statistics match the tournament data (more on that below), so the webapp now runs it exactly as fitted: no bias correction, temperature 1. This was validated directly by sweeping the old knobs against the new correctionless model, and no hand-tuned setting beat it. So the slider is gone, there is a single Score, and the temperature control (now defaulting to 1.0) has moved into the advanced sampler options where it belongs: a diagnostic tool, not a setting you should ever need to touch.
The old models were fit by pseudo-likelihood: fast, standard, and subtly wrong for this use. Pseudo-likelihood matches the model's conditional distributions ("given these five, how likely is this sixth?") but not its overall statistics, and in practice it over-concentrated: the worst-case gap between how often the model used a build and how often tournaments actually did was about 0.34, which is enormous for a probability. This is what the old Bias Adjustment slider was built to address, and why it's no longer necessary. The new models are fit by Boltzmann learning, which iteratively samples the model and adjusts it until its usage rates and pairing rates match the data directly. That worst-case gap drops to about 0.05. When the completer says a team is likely, that now means likely under a model whose statistics agree with real tournament results.
The second change is structural. The model now treats each team slot as a Pokémon whose held item is one of several states of that Pokémon (in physics terms, a Potts model: the site is the species, the state is the item). The sampler moves accordingly, swapping which Pokémon occupies a slot and, separately, rerolling which item it holds. That factoring is exactly what makes the new features possible: locking a Pokémon while the model picks its item is just freezing the site and sampling the state, and the item-free toggle is just summing the states.
The Potts framing also makes the model easily extensible. The old representation treated every distinct build as its own vocabulary entry, which meant adding another attribute would have multiplied the vocabulary: every Pokémon crossed with every item crossed with every ability, and so on, with most combinations too rare to learn anything about. The new model instead stores each attribute as its own dimension of a Pokémon's state, so adding one grows the model additively rather than multiplicatively, and Boltzmann learning keeps the fit correct as the state space grows. That's not a hypothetical: items are the first attribute on this footing, and the plan is to bring abilities, natures, and eventually moves into the model the same way.
The meta page and the Pokémon detail modal were rebuilt on the same view. Synergy is now reported between Pokémon, with each pairing expandable to show how specific items strengthen or undermine it, instead of one flat list of build-pair couplings. The synergy rankings also now weight each build by how often it's actually used, which removes a bias that previously favored Pokémon with only one common item.
This update addressed many of the issues with the old model and sets up k2dex to continue to improve. That said, this is an enormous update and likely to have some rough edges. If something on the new site looks wrong, behaves strangely, or is simply worse than what it replaced, or if you have an idea that would make k2dex more useful to you, come say so in the k2dex Discord server. Bug reports, feature requests, and questions about how any of this works are all welcome. This is a true passion project for me, so I genuinely enjoy answering questions and appreciate any feedback you might have.
Happy teambuilding!