A compact model program for testing deliberate reasoning, instruction following, and deployment tradeoffs on accessible hardware.
Scope of the model line
Aurora Proelia is North ML’s compact-model research line. It is used to study deliberate instruction following, bounded reasoning, and deployment behavior under tight parameter and compute constraints.
The work is exploratory. Model records describe design intent and evaluation method; they should not be read as independent benchmark validation or a claim of parity with frontier systems.
Evaluation before release language
A useful compact model must be stable across prompt phrasing, know when to escalate, and preserve requested structure. North separates those behaviors into small, reproducible evaluation sets rather than collapsing them into one headline score.
Runs record prompt templates, decoding settings, evaluator version, and known contamination risks. Changes are compared against the same fixtures so an apparent improvement can be traced to a model change rather than a grading change.
System value, not model mythology
Compact models become valuable when the surrounding system gives them a narrow job, clean context, and a reliable fallback. They are not replacements for every general model task.
Aurora Proelia is therefore evaluated as a component in retrieval and research pipelines. The practical question is whether the complete system becomes faster, cheaper, or more controllable without hiding a quality loss.
Current limitations
Long-context synthesis, unfamiliar domains, and ambiguous multi-step instructions remain high-risk areas. Public model records will keep those limitations attached to any future benchmark results.
North will distinguish internal evaluations, third-party results, and production observations. Until independent replication exists, target language remains target language.