The shortest version: jAIn searches model space for three tabular tasks on one table, and it costs an hour or so of compute to find out what it finds. It is worth that when the model you have is not near the ceiling of what the data supports, when you have enough rows for a held-out score to mean anything, and when a few points of accuracy or several times less serving cost are worth a deployment. The nine sections below are the cases where at least one of those is false.
1. Your problem is not one of the three
jAIn takes one rectangular table — CSV or TSV — with one target column, and searches for binary classification, multiclass classification or regression. Everything else is out of scope today, and out of scope means not offered rather than unpolished.
Specifically: no time series and no forecasting — there are no horizons, and the leakage story for chronological splitting is unfinished, so it is not offered at all. No text, images, audio or video: a free-text column is read as a category, not as language. Nothing that is not one table with one target — joined tables, ranking, recommendation, clustering, survival analysis, causal effects and optimisation are all outside it. None of these is on a dated roadmap, and whichever arrives first will arrive with results rather than with an announcement.
This is the most common reason an evaluation ends, and it is the cheapest one to discover. The scope band on the product page is the authoritative version of it.
2. You have a few thousand rows or fewer
A search applies selection pressure to a held-out split, and a small split is a target that can be hit by luck. Under a few thousand rows the search starts fitting the evaluation rather than the problem: the reported score improves while the deployed model does not. A regularised linear or logistic model usually wins outright at that size, trains in seconds, and is easier to defend to whoever has to approve it.
There is no hard row count below which we refuse to run — the threshold depends on the number of columns, the class balance and how noisy the target is. The honest guidance is that under a few thousand rows you should treat any lift jAIn reports as unproven until it survives a repeated split you ran yourself.
3. A linear model is already the right answer
If the relationship you are modelling is close to additive, or the coefficients are the product — a pricing factor, a risk weight, an elasticity someone will argue about — then the model is not the deliverable, the explanation is. A searched pipeline can win on held-out error and still be the wrong artefact to hand over. Fit the linear model, look at the residuals, and only reach for a search if the residuals have structure in them.
4. Your existing model is already near the ceiling
Every dataset has a level of irreducible noise, and no method gets past it. When a well-tuned gradient-boosted model is already close to that level, what is left to win is small: on those problems our median lift is under a point. That is real and it is measurable, and it is rarely worth a procurement cycle, a security review and a deployment. The tell is that your last three modelling efforts each bought a tenth of a point.
The corollary is where the search does pay: a first model nobody has tuned, a model that has not been revisited since the data changed, or a model that is accurate enough and far too expensive to serve.
5. You already know the governing relationship
If the physics, the actuarial structure or the business rule is known, encode it. jAIn has no way to take domain knowledge as an input yet — no constraints, no monotonicity requirements, no known-terms to start from — so it spends its budget rediscovering what you could have written down, and it may well rediscover it imperfectly. This is a real gap in the product and not a property of the method; until it closes, a constrained model you specified beats a search you did not.
6. Coefficient-level interpretability is a requirement
If a regulator, a model risk function or a customer-facing adverse-action notice needs reasons at the level of individual coefficients, a searched ensemble is a harder story to tell than a GLM. Feature attribution ships with the model, and attribution is not structure: it tells you what moved a prediction, not what the model believes in general. Some review boards accept that and some do not. Find out which yours is before the evaluation, not after.
7. The compute is not worth the fraction of a point
A search is an hour or more of compute per run, and it has to be re-run when the data shifts. That trade is easy when the model sits under something that earns per prediction, and it is a bad trade for a model that is scored once a quarter by three people, or for a table where a point of accuracy changes no decision anybody makes. Work out what one point is worth to you first. If the answer is "not much", the answer to this page is "not us", and that is a perfectly good outcome.
8. You need a benchmark result today
This is the limitation we are least comfortable with and it is the one most worth stating. jAIn does not yet run your incumbent model as a baseline or write an auditable record of a run, so today it cannot hand you a receipt that says this model, this data, this budget, this score, reproducible by you. Until it can, every comparison to your existing model is one you have to run yourself. The protocol we will publish results under is written and public already, deliberately in advance of the results, so it can be argued with before it is convenient.
If your evaluation requires a vendor benchmark against a named competitor, we do not have one, and the ones you will find elsewhere were written by the competitor. That is a real reason to defer looking at us.
9. You cannot put the data through a hosted preview
jAIn is a free Research Preview with no SLA (terms), and uploading is not live in the hosted product yet. A private deployment is designed and not shipped. If the data cannot leave your environment, the route today is running it yourself — locally or in Docker — and if that does not work for you either, then the answer is to wait rather than to start.
What this page is not
It is not a negotiating position, and it is not a list we intend to shorten by rewording it. Four of these conditions are on the product page too, under "Where jAIn loses", because a limitation that only appears on a page nobody links to has been hidden rather than published. When the proving run produces a win, tie and loss distribution across real datasets, the losses in it will be published here and the conditions above will be rewritten against evidence rather than against experience.
If your problem is not on this page and you want a second opinion on whether it is worth an hour of search, write to us and say what the table looks like. We will tell you when the answer is no.