Automatic Classification
Automatic and Statistical Classification
MagPunk analyzes and classifies content with a statistical model running directly on your device. There is no human editor or large language model reading the content in the background. All of the labels are based on predefined mathematical formulas.
Not an Editorial Judgment
The “Critical”, “Low Quality”, or Sentiment (Positive, Neutral, Negative) labels assigned to content are merely a mathematical probability calculation. These labels do not reflect MagPunk’s editorial opinion, endorsement, bias, or an objective truth about the accuracy of the content.
Statement of Reputation
Our algorithm attaching a label like “Low Quality” or “Critical” to a piece of content or publication is merely an automatic estimation about the syntax of a single title. This classification in no way constitutes a judgment or evaluation about the general quality, reliability, vision, or corporate reputation of the institution, author, or publisher publishing the content.
Expectation of Margin of Error
Because the system is statistical, it is inevitable that it will produce wrong decisions. A truly important and urgent critical event may not be labeled (false negative), or ordinary, harmless content may be mistakenly shown as if it were critical (false positive). MagPunk should never be seen as a single and absolute source of information.
Title Limit
The classification model reads and evaluates only the title of the content, it does not examine the body or the entirety of the content. In the event that the content title is incomplete, misleading, ironic, or clickbait, the decision the model gives will also inevitably be misleading.
Not Advisory in Nature
No label, scoring, score, or algorithmic output displayed in the MagPunk app constitutes any legal, medical, financial, or professional advice. Always refer to primary official sources or experts for your decisions requiring liability.
The Teacher and Its Limits
The labels in the training pool were produced by a large language model; a single consistent teacher was preferred over inconsistent labelling from many sources. The model itself, the language-specific decision thresholds and the “prefer missing over pressing a wrong label” policy are ours; hand-curated trap examples were added to the pool and the labels were reviewed in rounds. The teacher itself does not produce an absolute or flawless reality; it may contain inconsistencies within itself. Our labels are framed by the accuracy and limits of this standard.
Transparency
The background of every automatic decision is transparent. In the app, opening an item’s details shows word by word why the model assigned that label and how much each word contributed to the result.
To read step by step how a decision is produced, see Understand MagPunk; for the measured accuracy figures, how the measurement was done, the labelling criteria and the downloadable verification data, see Data & Transparency.