The distinction is not a matter of interface. It concerns what kind of model sits underneath, what that model was trained on, what it gives back, and how much you can rely on it. A chatbot returns text. Alfraido returns predicted values for measured properties, with a range showing how confident the prediction is, calculated from your own laboratory records.

Alfraido's three components are ordered by dependency:

1. Data structuring. Laboratory records are converted into a consistent form: harmonized units, resolved raw material identities, declared test methods, recorded provenance.
2. Machine learning engine. Models are fitted on those structured records to predict properties, report how confident each prediction is, and suggest which experiment to run next.
3. Collaboration. Several contributors write into the same structured record set, so the models see the laboratory's shared dataset rather than a collection of private spreadsheets.

The modeling layer only works because the structuring layer exists. This is why the platform is not reducible to an algorithm.

1. Two families of models

Machine learning models fall into two families. The difference matters because it decides how much data you need before the model becomes useful.

A parametric model has a fixed internal structure decided before it sees any data. Think of fitting a straight line through a set of points. The line has two adjustable quantities, the slope and the intercept, whether you fit it through 10 points or 10 million. Once fitted, the original data can be thrown away, because everything the model retains is contained in those two numbers. Modern neural networks are the same idea at a much larger scale: they have a fixed structure with billions of adjustable quantities, set before training begins.

A non-parametric model has no fixed structure. It adapts to however much data you have. These models keep the data itself and use it directly when making a prediction. A prediction for a new formulation is built from the measured formulations nearest to it, weighted by how similar they are. Nothing is compressed away and discarded.

The name is slightly misleading. Non-parametric does not mean the model has no settings. It means the shape of the model is not fixed in advance; it is determined by the observations.

1.1 Why parametric models need a lot of data

A model with many adjustable quantities has to determine all of them from the data available. If you have 200 experiments and a model with a billion adjustable quantities, the data does not come close to pinning them down. The result is a model that changes drastically if you add or remove a few experiments, and that will confidently give different answers depending on which subset it was trained on.

This is why large neural networks work well in domains where data is abundant, such as images, text, and web traffic, where training sets contain millions to billions of examples. Take away the very large dataset and the approach does not degrade gracefully. It stops being usable, not because the software is bad but because the question being asked of the data has no stable answer.

Two consequences for formulation R&D:

• A typical formulation or process development programme produces tens to a few hundred trials. Each one costs raw material, equipment time, and technician time.
• That is thousands to millions of times smaller than what a large parametric model needs. The gap cannot be closed by better software.

1.2 Why non-parametric models work with small datasets

Three reasons.

They start from a reasonable assumption instead of from nothing. These models assume that similar formulations give similar results, and that properties change smoothly rather than jumping erratically. That assumption is usually correct in chemistry, and it is doing a great deal of the work. It means the model does not have to learn the general shape of the response from scratch; it only has to learn where your particular system sits. This is why useful predictions are possible from a few tens of trials.

They stay as simple as the data justifies. In parts of the formulation space where you have run many trials, the model captures fine detail. In parts where you have run few, it stays conservative. You do not have to decide in advance how complicated the model should be.

They tell you what they do not know. Every prediction comes with a range, and that range is calculated from how close the new formulation is to formulations you have actually measured. Ask for a prediction well outside anything you have tested, and the range comes back wide. This is the property that makes it possible to plan experiments rationally: you can choose the next trial to reduce the largest remaining uncertainty, or to test the most promising untested region, instead of choosing by intuition.

The trade-offs should be stated plainly. These models slow down as the dataset grows very large, because every prediction consults the whole dataset. They rely on the similarity assumption being reasonable for your system. They lose accuracy when there are very many input variables and few trials. These are the right costs to accept when data is scarce and expensive. They would be the wrong costs if you had millions of records.

1.3 Summary of the contrast

2. Generative AI and a non-parametric machine learning engine

Large language models are the extreme case of the parametric family. The differences below are structural. They are not a question of the technology being immature.

They answer a different question. A language model predicts which word comes next in a sentence. Alfraido's engine predicts what value a property will take for a given formulation or set of process conditions. One produces text. The other produces a number with units.

They are trained on different data. A language model is trained on text collected from the internet before your laboratory was involved. Alfraido's models are fitted on your trials. A model that has never seen your raw materials, your test methods, or your process conditions cannot make a quantitative statement about them, however fluently it discusses them.

They handle confidence differently. A language model has no mechanism for knowing whether it has any basis for a number it states. There is no interval and no signal that the question falls outside what it can support. Alfraido returns a range with every prediction, derived from the trials it has.

They fail differently. A language model with no relevant information still produces a plausible, well-written answer, because producing plausible text is exactly what it was built to do. A non-parametric model with no relevant data returns a wide range. The first failure is invisible, the second is visible.

They are reproducible to different degrees. Given the same dataset and settings, Alfraido returns the same prediction every time, and the calculation can be traced and audited. Text generation does not work this way.

More data helps in different ways. Generative models improve when more general text is added to their training. Alfraido's models improve when you add your own trials. There is no external dataset that can substitute for them.

None of this means language models are useless. They are good at text: summarizing reports, drafting documentation, pulling fields out of unstructured records. That last one is close to Alfraido's structuring layer. It is not the modeling layer, and it does not produce predictions.

3. The structuring layer as a precondition

Everything above assumes the data exists in a form a model can use. In practice it usually does not. Laboratory data tends to sit in individual spreadsheets, with inconsistent units, raw materials entered under whatever name was on the drum that week, test methods left unrecorded, and no way to tell which trials are comparable.

Structuring is therefore not administrative housekeeping. It decides whether the dataset has any value at all. In practice it means:

• Consistent units. A property recorded in mixed units is not one variable. It is noise.
• Resolved raw material identities. The same product entered under three supplier designations looks like three different materials to a model.
• Declared test methods. Two results from different test methods are not measurements of the same thing and must not be pooled.
• Provenance. Operator, date, batch, and equipment are what allow a systematic effect to be identified as such, rather than wrongly attributed to the formulation.
• Failed and out-of-specification trials kept, not deleted. Removing them leaves a dataset covering only what already works, which is the region where there is least left to learn.

A small, well-structured dataset supports a model. A large, unstructured one supports nothing.

4. The collaboration layer

The same reasoning applies. If the number of trials is what limits model quality, then it matters whether trials are pooled. One researcher's dataset is smaller than a team's, which is smaller than the organization's across sites.

Collaboration here means shared access to a common structure, with attribution and history, so that trials accumulate in one place instead of fragmenting. The reason is statistical rather than organizational: it increases the amount of data available to every model on the platform, and it means the data stays when people change roles.

5. What the system does not do

Stating the limits is part of describing the system.

• It does not replace experiment. It reduces how many experiments are needed to reach a target, by indicating which ones to run.
• It does not reliably predict far outside the region you have tested. It reports that the region is untested.
• It does not explain mechanism. A model that predicts a property accurately is not a chemical explanation of why.
• It inherits the limits of your history. If past trials cover only a narrow part of the formulation space, the models are limited to that part, and the confidence ranges will show it.
• It does not converse.

6. Summary

Alfraido sits in a different place from generative AI because it solves a different problem. Generative models are large fixed-structure models that need volumes of data no laboratory produces, and they return text rather than measured quantities. Alfraido uses models that adapt to the amount of data available, start from the reasonable assumption that similar formulations behave similarly, and report a confidence range with every prediction. That combination is what makes it defensible to learn from datasets of tens to hundreds of trials.

The structuring and collaboration layers exist because the modeling layer depends on them. Structured data is what the models need; collaboration determines how much of it there is.