Understand MagPunk

This page does not promote MagPunk; it explains the problem MagPunk is trying to solve and the method it uses.

The page progresses in concept–response pairs. First, a concept of journalism or measurement is explained — echo chamber, gatekeeping, Jaccard similarity, precision and recall. Immediately following is MagPunk’s response to that concept.

At the very end, the things the app does not and cannot do are listed in a single block, along with measured error rates.

The goal is for you to leave having learned something, even if you don’t download the app.

23 concepts · 34 responses · 57 cards · approx. 44 minutes

If you prefer to read the whole thing on a single page instead of proceeding card by card, the Document button on the top bar provides the same content on a single printable page. You can change modes at any moment; you will stay at the heading you are on.

News and Headline

News is the communication of a new event that concerns the public; the headline is the way of highlighting the most striking one among these events.

In traditional journalism, the limited page space forces editors to choose the most important event of the day. The headline is the result of this choice and directly reflects the publication’s editorial policy. By looking at the headline, the reader quickly understands what is considered critical in the world that day. In the digital environment, although the space limit is lifted, since attention is limited, the headline logic continues by changing shape.

Example: The main news given in the largest font on the front page of a newspaper is the headline.

New Media

It is all of the communication tools where content is produced and distributed in the digital environment, and the reader is not only a consumer but also a participant.

Unlike traditional media, new media offers two-way communication. While printed newspapers or television broadcasts provide a one-way flow of information, new media platforms center on interaction with mechanisms for commenting, sharing, and immediate reaction. This structure increases the speed of information dissemination, while also paving the way for unverified content to quickly enter circulation.

Example: Comment sections of news sites, news shares on social networks, and citizen journalism are parts of new media.

Internet Journalism

It is a continuous broadcasting model where the processes of gathering, writing, and distributing news are carried out entirely over digital networks.

Internet journalism provides an uninterrupted and continuous flow of information throughout the day by completely eliminating physical boundaries such as printing times or broadcast schedules. The ability to update news instantly allows events to be followed in real time. However, this pressure of speed makes in-depth research difficult, which is a structural problem that leads to the production of superficial texts focused solely on attracting attention.

Example: Announcing an event with a few-sentence breaking news note the moment it happens, and continuously updating the text afterwards is typical internet journalism.

RSS and RSS Readers

It is a data feed technology that allows collecting and reading up-to-date content on websites in a standard format without entering the site.

RSS (Rich Site Summary) is a simple but powerful protocol that makes it possible to follow hundreds of different sources from a single place. Unlike algorithmic feeds, there is no content filtering or sorting in RSS readers; every published content flows chronologically. In this way, the reader decides what to see and ensures information consumption completely independent of the directions of platforms.

Example: Instead of visiting ten different news sites one by one, seeing the new headlines of all sites at the same time in an RSS reader application.

Source Selection: Who Determines What You Read?

It is the choice of which platforms, publications, or intermediaries to trust in order to learn about a news item or event.

In modern information consumption, the content read is generally determined not by the reader themselves, but by the algorithms of the platform used. Source selection is the filter that draws how the outside world will be perceived. If news is taken only from networks defending a single view or from a single social platform, the worldview becomes trapped within the boundaries of those sources. Diversity is the most important component of an information diet.

Example: Someone who follows only newspapers with a specific political leaning remaining completely unaware of the arguments of other views.

Echo Chamber

It is a closed communication environment where individuals only encounter views that confirm their own thoughts and different voices are excluded.

An echo chamber is formed by social networks connecting like-minded people to increase interaction. In this environment, the same ideas are constantly repeated, creating a sense of unquestionable reality. Different views are either not seen at all or are marginalized by being made a direct target of attack and ridicule. This situation is one of the main causes of social polarization.

Example: A user who follows only people with the same political view believing that the entire country thinks the same way.

Framing Theory

It is the direction of how the reader will perceive the subject by highlighting a specific aspect of an event, rather than its entirety.

Framing is concerned with how the news is understood rather than what it tells. Even if the event remains the same, the words used, the photos chosen, and the highlighted details completely change the meaning. Calling an action a ‘protest’ or a ‘riot’ is a framing preference that directly determines the reader’s perspective on the event. This situation is the most powerful influence tool in the hands of the media.

Example: A tax increase being framed as an ‘economic measure’ in some news and as a ‘burden on the citizen’ in others.

Agenda-Setting Theory

It is the approach explaining the media’s power to dictate to people not what to think, but what to think about.

An event appearing in headlines for days creates the perception that the event is the most important problem for the society. The media draws the boundaries of public discussion by deciding which news to highlight and which to ignore. The reader tends to see the topics heavily covered by the media as the main issues of the country.

Example: Despite a statistically low crime rate, the creation of a widespread feeling of insecurity in society by the media constantly reporting public order news.

Gatekeeping

It is the process of deciding which among millions of events will become news and which will go to the trash.

Countless events take place in the world every day, but a very small portion of these finds a place in a publication. Gatekeepers; as editors, editors-in-chief, and algorithms, filter the flow of information. This filtering process takes place in line with the priorities, political stance, and economic interests of the publication.

Example: An editor’s decision to headline a short statement by a local politician instead of news of a disaster in a distant continent.

Clickbait

It is a deceptive headline technique aiming to ensure the content is clicked by intentionally provoking the reader’s curiosity with incomplete information.

Clickbait looks for the value of the news not in its content, but in the click-through rate of the headline. Instead of giving a summary of the event, an information gap is created with expressions like ‘Here is that detail’, ‘It shocked those who saw it’. The reader is forced to click the link to complete the missing information. (Note: Clickbait is the name of a concept. MagPunk’s ‘low quality’ label does not map onto it one to one: it catches measurable patterns and often misses headlines that merely leave a curiosity gap — see 2.3.3.)

Example: Giving the result of a match as ‘Surprise score in the giant derby!’ and explaining the score only in the last paragraph of the news.

Noise (in the context of communication)

It is all of the distracting, repetitive, or low quality content that makes it difficult to reach the actual information.

In the information age, the real problem is not being unable to reach data, but sorting out what is meaningful from a massive pile of data. Noise consists of duplicate news, unverified rumors, distracting ads, and overproduced content. Too much noise makes it extremely difficult for the reader to see the signal, meaning the real and important news.

Example: An important law change getting lost among thousands of messages posted about a celebrity’s outfit on the same day.

Data Privacy · case: Cambridge Analytica

It is the problem of digital footprints belonging to the user being collected and used for other purposes without their consent and knowledge.

Personal data is the fuel of the digital economy. The value truly targeted in services offered for free is often the human data itself. Permissions given for a simple personality test or survey can turn into massive profiling networks in the background. This collected data is used to secretly map people’s vulnerabilities, fears, and tendencies.

Example: A seemingly innocent application secretly collecting the data of all the friends of the person who gave permission and selling it for political ad targeting.

Backend (and the importance of its absence)

It is the back face of an application that is not seen by the user, processes data, stores it, and runs on remote servers.

Most modern applications use a backend, meaning a centralized server architecture. In these systems, every click made on the screen, every text read is sent to a server over the internet, processed, and recorded there. The absence of a backend architecture means that the application can work entirely on its own smoothly without exchanging any data with the outside.

Example: A cloud-based note application constantly sending the written texts to the company’s servers is an example of backend usage.

Edge Computing (on-device computation)

It is the carrying out of data processing processes not on a remote cloud server, but where the data is produced, meaning directly on the user’s own device.

While data is sent to remote servers to be processed in cloud computing, heavy operations take place on the phone’s or computer’s own processor in the edge computing approach. This method prevents data from going out, minimizes latency, and eliminates the necessity for a continuous internet connection.

Example: The facial recognition process in photos being done within the phone’s own chip, not on a remote server.

Ad Algorithms (and the importance of their absence)

It is the set of mathematical rules trying to show the user the most suitable ad by analyzing their interests, behaviors, and vulnerabilities.

The main goal of ad algorithms is not to inform people, but to convert all of human attention into commercial money. To keep people on the platform longer, these systems constantly highlight provocative content that often arouses anger, fear, or excessive curiosity. The value of the text is determined entirely by the probability of it being clicked, not its accuracy.

Example: After a shoe search, shoe ads constantly catching the eye on every site visited.

Staying Offline

It is the ability of the software to continue fulfilling its core functions without missing anything even when the internet connection is lost.

Almost all modern applications continuously demand an uninterrupted internet connection. However, text-based data can easily be downloaded to the device beforehand. Being able to work offline is an extremely vital feature that always secures access to information even in extraordinary situations where there is no internet reception or it is censored.

Example: News headlines downloaded before getting on the subway being readable and analyzable even without an internet connection.

Jaccard Similarity

Jaccard similarity is the ratio of the common items of two sets to the total of distinct items in those two sets.

Two titles are first converted into a set of words. The number of common words is divided by the total of distinct words appearing in the two sets. The result is a number between 0 and 1: 0 means no common words at all, 1 means the two sets are exactly the same. The measurement does not look at the order of the words, it only looks at which words are present. Its calculation is cheap; it can be run among thousands of titles.

Example: In the {earthquake, region, damage} and {earthquake, region, team} sets, the common word count is 2, the total distinct word count is 4. Jaccard = 2 / 4 = 0.50.

Classical ML (Machine Learning)

They are software models that make predictions by extracting statistical patterns over large datasets, learning the rules from data instead of humans.

Classical machine learning does not understand text or context like humans at all. Instead, it makes linear probability calculations by converting the frequencies of words and letters into mathematical weights. For this reason, although it cannot resolve implicit subtle ironies, it can detect certain patterns within its statistical limits.

Example: A model of approximately 7 MB in size running on the device making a prediction just by looking at the word frequencies in the title without reading the body of a news item.

Precision, Recall and Threshold (P / R / F1)

They are the metrics that measure the success of a detection system; they express the balance between how few errors are made (precision) and how many of the targets can be found (recall).

Precision is the rate of the system being right when it says ‘this is so’. Recall is how many of the ones that are truly so it can find. The threshold value is the balance point between these two. While fewer but more accurate results are obtained when the threshold is raised, mistakes begin to mix in when the threshold is lowered.

Example: In a spam filter, if precision is high, real emails never fall into spam. If recall is high, no spam escapes.

Media Sentiment Index

It is the mathematical calculation of the general tone created by the words used in news texts as positive, negative, or neutral.

How an event is exactly presented to the outside is completely directly related to the word choices and writing style passing in the text. Sentiment analysis successfully and quickly measures the overall atmosphere of the news by mathematically summing up the emotional loads carried individually by the words in the text. This calculation aims to show the statistical tone of the words at a basic level.

Example: A headline containing ‘crisis’, ‘collapse’, ‘disaster’ statistically giving a highly negative tone.

Speed, Trend and Viral in Media

It is the dynamics of a topic exploding simultaneously across multiple sources, spreading rapidly, and entering circulation.

In digital media, short-lived passing trends are generally followed much more than fundamental events. A news item being copied quickly and spreading with great speed (becoming viral) without hitting any obstacles shows only its surface attractiveness, not its structural importance. When topics become a trend, the feed changes completely.

Example: A celebrity statement being given as breaking news with the exact same sentences across all platforms within 3 hours.

Content Density in Media

It is the volume taken up in the feed by duplicate news produced on the same topic, mostly with the same words.

When an event takes place, different publishers constantly using agency texts exactly by copying them instead of doing their own original research creates a massive content density. This annoying situation persistently causes the different headlines that constantly appear before people to actually be copies fed from a single source.

Example: Ten sites, thought to have different views, publishing a text consisting of the exact same words by copying it.

Bias, Neutrality and Objectivity — and the limit of measuring them

It is the worldview (bias) of the publication, its effort to remain impartial while conveying events (neutrality), and its goal to present the truth without bending it (objectivity).

These concepts are very clearly the most fundamental and complex natural elements of human communication. No text can be wholly impartial because even which word is carefully chosen reflects a hidden ideological bias. It is not feasible for software to fully formulate these deep human concepts.

Example: The human interpretation in presenting the same event as a ‘struggle for freedom’ in one newspaper and a ‘security threat’ in another.

Has No Server — No Account, No Login

MagPunk has no central server; your data is not stored or processed on a remote computer.

The moment you download the app, everything starts working. You do not need to provide an email address, set a password, or open an account. All your settings, reading history, and favorites are kept securely directly in your phone’s storage. This architecture structurally prevents your personal data from being transferred to another company’s servers.

Example: Putting your phone in airplane mode and opening the app to see your saved data load completely.

Everything is Calculated on Your Device

Heavy operations like analyzing and labeling news are done not in the cloud, but on your phone’s own processor (edge computing).

Instead of sending news titles to a remote server and waiting for the result, MagPunk uses small-sized analysis models integrated into your device. The moment you download a news headline, your phone scans this data internally, performs the calculations, and prints the result on the screen. This prevents unnecessary data sharing with the outside.

Example: The Critical label being given by your phone’s own processor without a response coming from an external server.

No Popularity Signal and No Ads

Items are highlighted not according to how much they are clicked or how much ad revenue they will bring, but solely based on their information value.

There are no ad spaces or sponsored content inside the app. MagPunk does not see user behavior-based popularity signals like ‘most read’ or ‘trending’ as a determining factor in the feed. The system evaluates only the news itself according to its statistical rules, so your attention is not sold to provocative content or ads.

Example: A news item scoring solely based on the information value it carries in the MagPunk feed, even if it is clicked millions of times on social media.

Works Even When You Have No Internet

Since data is kept on your device, you maintain full access to the items you have previously downloaded even when your internet connection is lost.

When you open the app, the current feed downloads to your phone. Later, when you get on the subway or pass into a region with no internet reception, you seamlessly continue reading all the downloaded titles, summaries, and analysis labels. The internet is only used to pull new data, it is not required for the app to function.

Example: Being able to read news summaries and see their labels in the feed even after turning off your internet connection.

The Source List is Yours — The 435-Source Catalog is a Start

In MagPunk, it is not algorithms but entirely you who determines what to read; you are only presented with a curated starting point.

There is a ready-made catalog of 435 sources inside the app, but this is not a ranking of quality. You create a wholly personal reading list by selecting what you want from this catalog or by adding your own favorite sources that are not on the list. Algorithms cannot secretly put any source you did not select in front of you.

Example: Adding the address of a local newspaper not featured in the catalog and being able to read only the news of that newspaper.

Inventory of Data: What is Kept, Where Does It Go

It is the full list showing clearly what data of yours stays where, how long it is stored, and under what circumstances it leaves the device.

Your reading habits never leave the device. When you use the feedback form, it (only the text you write) is sent voluntarily. The only exception that goes out of the device is the ‘Search web’ function; and this does not run automatically, it only transmits the searched word directly to the provider. No data indicating who you are is collected.

Example: Not a single request going automatically even to the search engine unless you press a button.

Score Instead of Chronology (10–100)

It is the ranking of news not from newest to oldest, but according to an information value score between 10 and 100.

A continuously flowing chronological list causes the news that is actually unimportant but entered most recently to appear at the very top. MagPunk sorts the feed not by time order, but by the score of the news. Thus, a valuable item entered a few hours ago does not get buried under an unimportant item entered just now.

Example: A tabloid news item entered five minutes ago staying at the bottom with a score of 20, while a news item from two hours ago with a score of 85 stays at the top.

What Raises the Score, What Lowers It

It is by what mathematical rules the score, starting from the base score, changes according to the qualities of the news (headline, authority, quality).

The score starts with a base score of 60. If the news is a headline it receives +10 points, if it is not a headline it receives minus points according to its scan order (order×2). The news being covered in a large number of sources brings a +(8 × authority) contribution, while being a low-quality item (for example, clickbait) brings a -25 penalty point.

Example: The score of a news item that is a headline (+10) but is labeled low quality (-25) dropping from 60 to 45.

Decay: No Title is Permanent

No matter how high the score of the news is, it is the gradual dropping of this score with a statistical formula as time passes and the news moving down in the feed.

Information gets old over time. A very important news item with a score of 90 should not still sit at the very top hours after it went on air. The decay function gradually pulls the scores between 10 and 99 downwards over time. Critical items (operates in the 10-100 range) start at the very top but decay at the same speed; their privilege is their starting point, not durability. The badge stays for 3 hours; but the score does not freeze, it decays from the first moment.

Example: A news item that was at the top with a score of 90 in the morning decaying and moving further down towards the evening hours.

Equal Opportunity: Narrow Interests Are Not Crushed

It is the supporting of niche (special interest) news that might get lost in the density of mainstream media with extra points if there are few sources in its category.

A news item coming from just a single local newspaper is normally crushed against a mainstream news item transmitted by hundreds of agencies. MagPunk gives a +8 equal opportunity score to that item if the item is not multi-sourced and there are less than 3 sources in its category (for example, science or local news). Thus, minority voices preserve their visibility.

Example: An important article coming from a single technology magazine being able to climb to the top ranks among hundreds of politics news items.

Thresholds Are in Your Hands — Your Own Gatekeeper

It is being able to set the limit values yourself that determine the difficulty level of the items that will enter the feed, that is, showing what proportion of doubt will be accepted.

MagPunk does not impose fixed rules on you. You can adjust the overlap and similarity thresholds, the decay speed, the low quality penalty, and the reference and trend contribution. If you raise the threshold, the system makes fewer mistakes but might miss some items (high precision, low recall); if you lower it, it catches more items but wrong predictions might mix in.

Example: Choosing to get notifications only for truly massive events by setting the notification score threshold very high.

On-Device Classifier

It is the built-in classical machine learning system of about 7 MB in size that measures the quality, criticality, and sentiment of the news on your device.

Instead of using massive models in the cloud, MagPunk runs small-sized (≈7 MB) classical machine learning algorithms integrated into your device. Trained with L2 normalization in a space consisting of letters and words with around 16,000 (8,000 + 8,000) dimensions, this lightweight model rapidly calculates the statistical tone of every news item that comes to you.

Example: Your phone being able to calculate the negative word density in the news within seconds even when you have no internet.

Critical Label and Three-Hour Badge

It is the high importance mark given to disasters, wars, or very vital policy changes that could deeply affect society.

The critical label prevents the most vital news in the feed from getting lost. These items go above the standard 99 ceiling and directly receive 100 points. A news item with a critical label is displayed in your app with a protected badge for 3 hours. A critical item starts at the very top but decays at the same speed; its privilege is its starting point, not durability.

Example: A major earthquake news item being published with a red critical label and carrying a badge for three hours.

Low Quality Label

It is the system that recognises content of low information value by the measurable patterns it carries, and marks it.

The label reads the form of a headline, not its intent: listicle and service headlines (‘7 practical methods’), horoscope, fortune-telling and dream-interpretation content, and question patterns such as ‘who is’, ‘when is’. Content that receives the label instantly loses -25 penalty points from its score. The machine learning model has an F1: 0.75 in low quality TR detection, while the rule engine has an F1: 0.41 success by looking at specific phrases (‘ne zaman’, ‘kimdir’ etc.) or question mark patterns (2 or more ‘?’).

Headlines that carry no such pattern and merely leave a curiosity gap mostly do not get this label. A headline like ‘Here is that detail’ or ‘It shocked those who saw it’ steals the reader’s time just the same, but leaves no measurable trace; the system cannot see it.

Example: The headline ‘What does it mean to dream about water?’ receiving the low quality label — while ‘Here is that development concerning millions!’ does not.

Sentiment and Status Quo Test

It is the system that does not look at whether words are individually positive or negative, but what the action in the content changes compared to the situation yesterday.

Sentiment analysis does not look at word tones; it gives a negative decision if there is a new harm, positive if a harm is ending, and neutral (status quo) if the wheel turns and there is no clear result. For this reason, contrasting situations containing the same words are successfully separated.

Example: The title ‘A theft occurred’ separating as negative, while the title ‘The thief was caught’ separating as positive.

Political Neutrality Rule

It is the classification model intentionally avoiding labeling statements supporting or opposing a specific political figure, party, or ideology.

The goal of MagPunk is not to say that a news item is politically right or wrong. The model does not take sides in political texts and does not make implications like ‘this news is biased’ or ‘this news is objective’. The app’s job is solely to make technical (low quality) or statistical (negative) measurements.

Example: The model not stamping an ideological label even though newspapers with two different views present the same political news differently.

“Why This Label?” — Explainability

It is the transparency screen showing exactly with the contribution of which word the machine learning model on your device gave that label to a news item.

When the system assigns a label, there is no black box behind it. When you press the ‘Why this label?’ button, you clearly see from which word the model received how many points while making that decision. As a requirement of the transparency invariant; the sum of the mathematical contributions of the words and the base value of the system (Σ(word contribution) + base = score) is always equal to the final score of the label.

Example: Seeing transparently on the screen that the word ‘earthquake’ in a headline provided a direct contribution of +0.42 points to the model.

Rule Engine: Fallback System if the Model Fails to Load

It is the backup and open dictionary-based system that kicks in if the machine learning (ML) model cannot be loaded.

This engine is only a fuse; if the ML model is loaded it does not run at all. The two do not run simultaneously and contradict each other. The rule engine does not use complex math; it looks directly at word lists, strict logical conditions (like having consecutive question marks in the title), or dictionary scoring.

Example: When the model file fails to load for any reason, the rule engine kicking in and continuing to classify the news at a basic level.

Fingerprint and Jaccard Comparison

MagPunk reduces each title to a set of phrases and compares these sets with the Jaccard ratio to find items telling the same event.

The title is first simplified: converted to lowercase, apostrophes and suffixes are deleted, punctuation is discarded (Turkish letters are preserved), standalone numbers and words shorter than three letters are eliminated. The root of the remaining words is taken; stopwords and roots shorter than four letters are weeded out. One, two, and three-word phrases are extracted from what remains. This set of phrases is the fingerprint of the title.

Example: If the Jaccard ratio is 50% and above, two items are considered reference, if it is between 30-49% they are considered similar, if it is below 30% they are considered unrelated.

Reference, Similar, Unrelated

It is the similarity grading showing how much two news texts overlap over their words and semantic roots.

The system matches the news with each other by taking their fingerprints. If similarity is ≥ 50% the content is considered reference, if between 30-49% similar, if < 30% unrelated. The reference label is not an accusation of copying or plagiarism; high-quality publications that use the same official agency text and process the topic with exactly the same words naturally receive high similarity too.

Example: All of them being linked to each other as references upon a statement made by the Ministry being published with the same words on ten different sites.

Pioneer or Follower? (Authority Multiplier)

It is the authority multiplier calculated in a 3-day window determining whether a source is generally the first to break the news or gives the news later.

If your publication time is always the earliest among all references of a news item in the same group, the source gains authority. In the authority window of the last 3 days, if the source’s first publication rate is ≥60%, it receives the PIONEER (1.5× authority multiplier) label, if <30%, the FOLLOWER (0.5× authority multiplier) label. If there is not enough data in the system (less than 3 items), the multiplier is fixed at 1.0×.

Example: A news site grabbing the ‘Pioneer’ multiplier by being the source that always announces an event first within the last three days.

Viral: The Arithmetic of Spread Speed

It is the spreading label indicating a news item exploding simultaneously across multiple independent sources within a short time.

The viral label measures not the importance of the event but how fast it spreads. To earn the viral badge, the item’s age must be ≤ 3 hours and it must have at least one reference and two similar (or at least two reference and one similar) connections. This label allows you to see the momentary storms in the feed and stays on the item card for 4 hours.

Example: The item receiving the viral badge upon a statement made by a famous actor making headlines across four different platforms within two hours.

Spam Cluster: The Same Source Repeating Itself

It is the automatic filtering mechanism preventing a source from manipulating the system by publishing the same news repeatedly within itself.

Different sources breaking the same news shows the magnitude of the event (authority); however, the same source constantly breaking the same news again with very minor changes is spam for the system. The system groups consecutive references coming from the same source into a cluster and does not grant an authority bonus.

Example: A site not being able to gain points when it republishes the same traffic accident news with five different links for SEO.

Reports and Compare: How Did Two Sources Look at the Same Event?

It is the analysis screen calculating which source produced how much news and with how many minutes of difference and with what headlines they gave the same event compared to each other.

MagPunk does not secretly analyze your data in the background, it compares the news you downloaded upon your request. While you can see the overall performance of each source in the Reports tab, in the Compare tab exclusive to the Pro version, you can select two different sources and directly match their speed of giving the same news (for example, as “NTV agenda avg. 23 min behind”).

Example: Seeing side-by-side on the Compare screen how two different news channels looked at the same event, their headlines, and who broke the news first.

The Seven Gates the Notification Passes Through

It is the seven different check stages where a news item is silently tested inside the device before vibrating as a notification on your phone.

MagPunk does not disturb you with continuous notifications. For a news item to fall into your notification bar, it must pass through exactly 7 different gates within the system. All items (including critical ones) must exceed the 80 score threshold. The only privilege of critical items is that they bypass the category, language, and source filters. Thus, your valuable attention is not divided by ordinary developments.

Example: An ordinary accident news item entering the feed but not vibrating your phone because it could not pass the notification gates.

Daily Brief and Flash Duration

It is the summary presenting you the 10 most important news of the last 24 hours in the evenings and the metric indicating the hotness duration of the events.

Every day at 19:00, your device extracts the summary of the last 24 hours and informs you. At most 3 of the 10 items entering the brief can be critical news. As for the flash duration in the system, it indicates the active time news that directly crossed the exact 80 threshold was able to stay above that value.

Example: Being able to read the summary of only the 10 most important ones out of hundreds of developments occurring throughout the day when you return home in the evening.

Pulse: Source by Source High-Scoring Content

It is the screen filtering the main feed and combining only the news that exceeded the ≥ 85 score threshold in the last 24 hours into a simple list.

The Pulse view is an ‘elites’ summary for those who do not want to read all the news. It consists of only the strongest items within the last 24 hours. Pulse is ordered directly source by source; it groups the highest scoring items of each source and the listing ceiling is a maximum of 15 items per source (source and threshold settings can be changed in the Pro version).

Example: When checking the news after a long journey, skipping the unimportant details and merely browsing the developments with a very high score.

Watched Words — and Why They Don’t Change the Score

It is the feature allowing you to define the words you want to track, so you can see the news containing these words colorfully or prominently in the feed.

Watched words are a visual marker and increase permanence (exempting the item from archive cleanup); however, they do not grant your device the authority to change the score of that news directly. If a specific word you chose added points to the news, you would have created an artificial echo chamber with your own hands. For this reason, the score bonus belongs only to the trending words detected by the system.

Example: When you watch the word ‘space’, the news containing this word being underlined in the feed but not jumping to the very top.

Honest Report Card — Numbers Measured in a Real Feed

The honest report card is the version of labeling performance measured not in a laboratory environment, but on a natural distribution real feed.

The measurement was taken on July 17, 2026, on 1110 titles that never went into training, with calibrated thresholds on the device. Values are given separately by language, because the model behaves differently in every language. P shows how many of the ones we labeled were correct, R shows how many of the ones that should have been labeled we caught. The weak rows are also staying in the table; there are no cut or rounded numbers.

Example: In Low Quality TR, precision is 0.99, recall is 0.61. When it stamps the label, it almost never errs, but it catches only a portion of the low-quality items.

Only the Title is Read: Information Ceiling

It is the limit of the model and the rule engine looking only at the title of the news, not the full text of the news, while doing their analysis.

The machine learning model and rule engine on your device do not analyze the content of the news. Since the system makes its decisions by focusing solely on the news headline, details not clearly reflected in the headline cannot be measured and included in the score. This design makes it impossible to detect the fine details or context carried in the body text.

Example: The system being able to mark it as low quality if the headline is inadequate, even if the text of the news is highly detailed.

Does Not Do Fact-Checking

It is that MagPunk lacks the ability to research whether a news item actually happened or whether the information is correct.

The system does not confirm the reality of the data (names, numbers, or events) in the outside world. The app only measures the statistical tone of words and the spreading speed of news in reference to each other. When a news item is produced entirely as a lie or fiction by the publisher source, the system cannot perceive this.

Example: The system counting a fabricated event as critical or viral in case it spreads with capital letters across ten different sources.

Does Not Measure Neutrality

It is the inability of the app to detect the ideological stance, political position, or how objective a news source is.

MagPunk cannot claim that any content is politically impartial or neutral. The system only measures negative, positive, and neutral sentiment, this statistical emotional tone is not the equivalent of neutrality. A news item being intensely partisan can also be hidden by using statistically ‘neutral’ words.

Example: A group producing a text that is actually wholly one-sided and biased by choosing its words very carefully and neutrally.

Labels Are a Statistical Estimate, Not a Verdict

It is that no badge or label you see on the screen signifies a definitive truth or a final verdict.

All of the labels like low quality or critical stamped by the app consist of a simple probability calculation. The machine learning model (≈7 MB in size) and the threshold values behind it reflect a mathematically determined probability onto the screen. While the system operates, it can sometimes be wrong and produce incorrect labels from time to time.

Example: The system statistically making a mistake in one out of every ten decisions even in a setting where it operates with ninety percent (0.90) precision.

Did You See a Wrong Label? → Feedback

It is being able to report a piece of content the model incorrectly labeled to us, after personally seeing and confirming what will be sent.

You cannot manually change the label inside the app; this is a conscious choice, because a correction on a single device does not train the model, it only breaks that content. Instead, you tap the “report error” link in the reading panel, mark which label is wrong or which one is missing. It is copied to your clipboard and the submission page opens in the browser. Nothing leaves your device until you say “Submit”.

Example: An ordinary news item receiving the critical label because of the words inside it, and you reporting this after reading the text to be sent.