CAFE / PRESS SENTIMENT / METHOD
How AIN reads the news.
CAFE means Current Affairs & Further Exploration. It brings together news links, an estimate of the tone of press language, and source-backed discoveries through RABBIT.
The number is a reading of the language AIN collected—not a verdict on the world. This page explains what is measured, how the readings are combined, and where interpretation begins.
1. What is VADER?
VADER stands for Valence Aware Dictionary and sEntiment Reasoner. Developed by C. J. Hutto and Eric Gilbert, it combines a dictionary of sentiment-rated terms with rules for how wording modifies their emotional intensity. It was introduced for short social-media text. [1]
It accounts for features such as negation, intensifiers, capitalization, and punctuation. For example, “good,” “very good,” and “not good” need not receive the same reading. This is rule-based text analysis, not a generative model composing an opinion about an article. [2]
AIN applies sentiment analysis to the headlines and summaries it collects. That is a limited view of reporting, not a reading of every article or every page an outlet publishes.
2. What do positive and negative mean?
Positive means the analyzed language has a more favorable emotional tone. Negative means it has a more unfavorable emotional tone. These describe language, not whether AIN approves of the story.
A report about suffering may have negative language and still be accurate, necessary journalism. A cheerful announcement may have positive language without its claims being true or its consequences being beneficial.
Zero is the center of the scale: little or no net directional sentiment. A reading near zero can also result from positive and negative contributions balancing each other. It does not prove that each story was neutral, balanced, or free of bias.
VADER's raw compound score ranges from −1 to +1. AIN expresses its aggregated readings on a −100 to +100 scale. [2] A value such as −14.7 is a sentiment index reading—not “14.7% of stories are negative,” a probability, or an accuracy rating.
A rise from −15 to −5 means the measured tone became less negative; it did not become positive. Small movements should not be treated as precise changes in public mood.
3. How AIN builds a reading
For each monitored source, AIN collapses feed-visible duplicates by URL—or by headline when a URL is absent—then uses the eligible headlines and summaries from the rolling previous 24 hours. The outlet score is not based on one randomly selected article or only the links chosen for Brief.
Each eligible item receives a VADER compound score. Newer items receive more weight, using an 18-hour half-life. The weighted mean is bounded to VADER's −1 to +1 range, multiplied by 100, and stored as a rounded whole-number outlet reading.
The combined Core reading is an equal-weight mean of eligible current Core outlet readings. Each included outlet contributes one stored outlet score; an outlet does not receive extra weight in the final mean merely because it published more items.
Edge, analysis/discovery sources, and non-VADER reference signals are not included in that Core mean. Missing or stale outlet readings are not replaced with zero. The source register shows the available coverage; a change in coverage can affect the combined reading.
Calculation
These describe the processing order. Displayed precision does not make the estimate more certain.
4. Press Oracle: the current reading
Press Oracle presents the current accepted Core reading. Its as-of timestamp belongs to that reading, not to the moment a visitor opened the page.
The word inside the sphere is AIN's editorial shorthand for a numerical range. It is not another measurement, an emotion reported by the public, or a prediction. The artwork and color are presentation cues, not additional evidence.
The word comes from a predefined mapping rather than a newly generated response on every visit. Changing the chart's visual scale must not change what the same score means. Missing or stale data should be read through its availability status, not mistaken for a neutral reading.
5. Press Tone: the 30-day view
Running 30D Tracking shows recorded daily tone over the displayed calendar window. The zero line separates net-negative and net-positive readings. The vertical scale adapts to the available values; it does not change the underlying score.
Daily values average the genuine Core aggregate observations recorded during that America/New_York calendar day. The current day can remain partial until it ends.
The Oracle's latest rolling-24-hour reading and the chart's daily average answer different questions. They can legitimately differ: one is the latest snapshot, while the other summarizes snapshots recorded throughout the day. Click, tap, or focus a point to inspect its date, value, and provenance.
A missing observation is not zero. If a chart includes reconstructed, estimated, or illustrative material, its status remains identifiable. A value recovered from retained observations differs from an estimate; a display illustration is not historical evidence. Display-only values never determine the current Oracle reading or enter measured history.
6. Press Band: individual sources
Press Band places each measured outlet at its own numerical position. Publisher names link to their websites; evidence links, where supplied, identify reporting associated with the observation. One evidence link is not the entire sample used to score that outlet.
The dotted Edge panel groups those monitored sources separately. It is not an Edge average and does not contribute to the Core score.
The standalone Press Band provides the source register and compact source histories. A source's latest score and its recorded daily averages may differ for the same reason the Oracle and daily chart differ. Mini-chart scales may adapt by source, so compare numerical readings rather than assuming equal line heights mean equal values.
7. Brief: why a link is here
Brief is an editorial reading list, not a ranking of outlets by positive sentiment. Its labels explain three selection considerations:
National Relevance — public scope and consequence.
Broad Consensus — coverage across independent source families. Agreement or repetition does not by itself prove a claim true.
AIN Selection — AIN's editorial interest in a specific story or discovery worth encountering.
The list aims to show up to twenty distinct eligible events without manufacturing entries to fill space. The labels explain selection; they are not VADER scores or guarantees of factual accuracy.
An outlet may be monitored without being eligible as an editorial link. AIN's current editorial policy excludes Fox News links from Brief and Fox News originating reports from RABBIT; that rule is separate from the Press monitoring roster.
8. RABBIT: from a report to a discovery
RABBIT means Relational Aperture for Branching Beyond Immediate Topics. It begins with a reported event and follows relevant details into substantive, source-backed discoveries.
Each continuation should develop something found in the preceding destination rather than repeatedly rephrasing the original headline. The full-trail review standard is an originating event followed by four researched destinations. This is the current editorial requirement, not a claim that every older published trail already meets it.
Source reading, evidence review, and owner publication approval are separate steps. A successful evidence check does not itself publish a draft. Readers encounter stored approved material; opening a public trail does not commission a new paid investigation.
9. What this does not tell you
These readings are not a measure of truth, political neutrality, public opinion, journalistic quality, or how good the world is. Topic choice, quotation, irony, ambiguous language, incomplete summaries, and uneven source coverage can all limit interpretation.
A small or changing sample deserves more caution than a well-covered one. The same numerical result can arise from different mixtures of stories. Use the score as a prompt to look closer, not as a substitute for reading the sources.
The former eight-axis CAFE display is no longer the current reader interface. Historical methodology and implementation records remain part of the project history; they should not be read as descriptions of the present screen.
10. Sources and implementation
[1] C. J. Hutto and Eric Gilbert (2014), VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text. Proceedings of ICWSM, 8(1), 216–225. Read the original paper ↗
[2] VADER reference implementation and scoring documentation. View the reference implementation ↗
The VADER references explain the underlying model. AIN's collection window, source selection, aggregation, display vocabulary, and editorial rules are separate implementation choices. The current source roster and available observations can be inspected in Press Band.