How it works
The teacher reasons once. The student answers every time.
HeartCut has an expensive model do its analysis once, during the build, then checks that analysis against your data. What survives trains a model small enough to run where the data lives.
The pipeline
From raw data to a working model
Steps 1 to 4 use existing AI models that HeartCut calls. After the freeze, steps 5 to 7 train the new small model, called the student.
-
Present each record two ways Two views
Every record is turned into two versions: text for the teacher to read, and arrays of numbers (tensors) for the student. Both come from the same underlying record, so the two models see the same facts.
-
Have a leading AI model analyze it
A capable general-purpose model, run by a provider or on your own hardware, answers questions drawn from your question bank. Your skills document (your domain knowledge, written down) and your style guide steer it. Every number it writes carries a citation pointing into the record.
-
Pass it through the validation gate Required
Three checks run in order, and an answer must pass all of them.
- Citations: every cited number is looked up in the source record and compared with what the answer says, either exactly or within a set tolerance.
- Format and wording: rules detect problems. A rule or a model rewrite repairs them, and the rules check again.
- Logic: catches answers whose numbers are all correct but whose claims contradict each other or the record.
-
Rank, label, and expand Optional
A model or a person ranks the answers that passed. Rule-based generators write the answers that ordinary code can produce, for every record, with no AI cost. A rewriting step restates a validated answer in another voice, length, or structure. It never sees the record, so it cannot introduce a new finding.
The freeze: settings locked before training
-
Lock the vocabulary and the input format
A tokenizer (the model's word list) is built from the validated answers, or taken from an existing pretrained model. An encoding (how the numbers in your records are scaled and labeled) is fitted to your input data. Both are locked and travel with the finished model.
-
Train a series of models, large to small
An optional mid-size helper model comes first, then the student, then one-question specialists that learn from the helper's detailed confidence scores. Writing style is a setting, so one student can answer in any voice or level of detail, switchable at answer time. Optional extra tuning reuses the gate as a hard rule on the student's own answers.
-
Test on real records, then deploy
Every record stays labeled real or generated through the whole pipeline, and testing uses real records only. The finished model goes onto a CPU beside the data, and can be used from the command line, from Python, or over HTTP.
Citations
Every number points into the record
A citation is a small tag wrapped around a number. The tag's name says what kind of thing the number is, its path says where to find it in the record, and its content is what a reader sees.
You define the addresses. units.42.s11.c197 means something in one project and
nothing in another. HeartCut only requires that a path always leads to the same value and that you
supply the code that follows it.
We had <signups path=metadata.users.new>10</signups> new users
from <val path=units.42.s11.c137>47.2</val> to <val path=units.42.s11.c197>48.9</val>
The validation gate
No model ever rules on a number
A number either appears in the record or it does not, so no judgment is needed. Asking an AI model to decide would bring back the problem the check exists to catch.
| Check | How | Why |
|---|---|---|
| Does the cited number match the source? | Automatic: follow the citation into the data and compare, exactly or within a tolerance | No judgment is needed, so no AI model is involved |
| Does it follow the format and wording rules? | Rules detect problems; a rule or an AI rewrite fixes them, then the rules check again | Spotting a problem is mechanical. Fixing it is rewriting, which an AI model does well |
| Is it consistent with itself and the data? | Logic checks, one for each rule you declare, run against the looked-up values | Catches answers whose numbers are correct but whose claims contradict each other or the record |
Without a gate Hard to detect
The most common failure is invented evidence under plausible reasoning. The writing is fluent, the numbers are wrong, and the answer looks the same as a correct one unless someone redoes the analysis by hand.
With the gate Seconds to detect
The evidence is always correct, so the only failure left is questionable reasoning. A domain expert spots it quickly, because the numbers are there and they check out.
The gate confirms the evidence, and an expert judges the conclusion. Every number behind a recommendation is checked against your data, so a reviewer weighs the reasoning without redoing the arithmetic. HeartCut promises grounded answers; it does not promise that every conclusion is correct.
Why the student is small and fast
Two reasons it is so much lighter
It states the conclusion directly
The teacher reasons once, while the training examples are being built. The student learns to give the conclusion without working it out again, and it holds one domain instead of all of human knowledge.
It reads numbers, not prose
A large model receives a table as text. Each decimal is spelled out as several tokens, and most of the length goes to separators and repeated column headers, all read before the first word of the answer. The student reads the same record as compact numeric input, at a fraction of the length and with no preamble. See a worked example.
Division of labor
You bring the knowledge. HeartCut brings the structure and the checking.
You know your domain, your data, and your questions. HeartCut supplies the format each piece of your setup must follow, the tools that check it, and everything between the raw data and the finished model.
You supply
- The question bank: the fixed set of questions worth asking
- The data, and the instructions for turning it into records
- The citation format, and the code that looks a cited value up
- A skills document: your domain knowledge, written down
- A style guide, and the logic checks your answers must pass
- The choice of teacher model
HeartCut supplies
- A required format for every configuration file, and checks that catch mistakes an author would miss
- The validation gate
- The training sequence, from locking the settings to the finished small models
- Testing on real records, and a build report of what every check found
- Deployment from the command line, Python, or HTTP
Next
Why train a small model, when an API call is one line of code?
Cost, speed, running on-site, and checkable answers: what each one changes, and where an API call is still the better choice.