Outrage
Strongly condemnatory and angry wording.
Methodology 1.1.0
How MediaLens evaluates language and editorial work
We describe observable wording, framing, sourcing, and context and support findings with excerpts. We do not decide whether the author is lying, what they truly think, or how they feel.
What scores mean
An intensity scale, not a verdict
A value from 0 to 100 describes the intensity of observable language patterns in the analyzed text. It is not a probability, a media quality rating, or a measure of emotions, intent, or truthfulness.
Frustration
Language of helplessness, irritation, and repeated complaints.
Sarcasm / irony
Ironic or mocking contrast in the wording.
Suggestiveness
Leading questions, insinuations, and unsupported implications.
Confrontation
Personal attacks, sharp confrontations, and denigration.
Factual style
Concrete, verifiably phrased information; not factual correctness.
A second analysis layer
Observable editorial concerns
We separately examine whether the headline and composition push a conclusion, whether a disputed issue relies mainly on one side, whether claims are clearly attributed and supported within the article, and whether context needed to understand the dispute is missing.
Framing pressure
The headline, lead, or composition pre-frames an emotional or evaluative conclusion.
Source imbalance
A disputed issue relies mainly on one side without a relevant response.
Weak attribution
The origin or status of important claims is insufficiently clear.
Unsupported claims
Significant claims lack a concrete basis stated within the article.
Context gaps
Information needed to interpret the reported dispute is missing.
Combined indicator
How rhetorical intensity works
Rhetorical intensity summarizes how strongly five tracked rhetorical patterns appear in the article. A lower number means fewer of these patterns; it is not a poor rating.
The value is the arithmetic mean of outrage, frustration, sarcasm or irony, suggestiveness, and confrontation scores. Factual style is shown separately and is not included in the mean.
Indicative descriptive bands
Low
Few tracked rhetorical patterns.
Moderate
Some patterns are more noticeable.
Elevated
Patterns appear prominently in the text.
High
Patterns are very prominent in the text.
The bands are currently indicative and have not been calibrated against a human-annotated corpus. The indicator does not verify facts, source credibility, or the article's overall quality.
From URL to result
Six auditable steps
- 01
Safe retrieval
The URL, DNS results, and redirects pass SSRF checks; downloads have time and size limits.
- 02
Article extraction
The main text, headline, and available metadata are separated from HTML without running page scripts.
- 03
Attribution
Segments are heuristically labeled as editorial voice, quoted person, headline, or unclear attribution.
- 04
Structured analysis
The provider returns six language indicators, five editorial concerns, explanations, short supporting excerpts, and confidence levels.
- 05
Technical validation
The backend checks the schema, value ranges, offsets, and exact match of each excerpt against the extracted text.
- 06
Versioned result
Provenance, model, methodology version, time, and uncertainties are stored; the complete article is not publicly archived.
Technical design
What happens inside an analysis
The current version combines safe retrieval, full-article availability checks, normalized extraction, a structured model call, and several independent validations.
Retrieval and access
A lightweight HTTP request is tried first; dynamic content may fall back to an isolated browser.
- The URL, DNS results, and every redirect pass SSRF checks and network limits.
- A clean browser context prefers rejecting optional cookies; explicit consent is a fallback.
- An unresolved banner, login requirement, or incomplete paywalled text stops the analysis.
Extraction and preparation
Readability-v2 selects the main content and the system checks for a complete article before analysis.
- Navigation, advertising, consent UI, and unrelated content are removed.
- Normalized text receives stable offsets; editorial voice and quotations remain distinct.
- The canonical URL and content hash support deduplication and later change detection.
Structured AI analysis
The model returns only predefined JSON; article content is treated as untrusted input, not instructions.
- Output contains six language indicators and five separate editorial concerns.
- Every finding requires a category, explanation, short excerpt, attribution, offsets, and confidence.
- The analysis does not rate outlet reputation, external truth, or the author's hidden intent.
Verification and publication
A result is published only after schema, range, and textual evidence checks.
- Excerpts must exactly match normalized text at the supplied offsets.
- The deduplication key includes content, methodology, provider, and model; concurrent copies must not be billed twice.
- Cookies and browser storage are discarded, BYOK keys are not stored, and the full article is not publicly archived.
Quality assurance
How we test the methodology
We separate automated technical checks that already run from methodological validation that is still being prepared.
✓
Implemented
Technical checks
- Validate the complete output with a shared runtime schema.
- Check excerpts and their offsets exactly against normalized text.
- Display editorial voice, quoted people, and unclear attribution separately.
- Security tests for internal addresses, redirects, timeouts, and oversized responses.
- Deduplicate by content, methodology, and model so the same analysis does not run concurrently more than once.
→
Validation plan
Human evaluation and calibration
- Build a genre-diverse Czech reference corpus with news, opinion pieces, interviews, headlines, and negative examples.
- Create annotation guidelines and have every sample labeled independently by at least two people.
- Measure annotator agreement, false positives, attribution accuracy, and extraction success.
- Calibrate the scale and its verbal interpretation only after reviewing the human evaluation results.
- Use the same corpus as a regression test whenever the prompt, model, extractor, or methodology changes.
- Publish the dataset size and composition, metrics, measurement date, and known failures without redistributing full copyrighted articles.
Interpretation boundaries
What the result cannot tell you
MediaLens stays focused on the language of a specific text. It does not replace fact-checking, psychological assessment, or consideration of the wider social context.
Truth and credibility
MediaLens does not verify factual correctness or source reputation. It evaluates only support, attribution, balance, and context visible in the article.
Personality, motives, and emotions
We do not infer personality, mental state, hidden motives, or genuine emotions of the author or speaker from the text.
Irony and wider context
Irony, indirect speech, and missing context can produce an uncertain or incorrect result.
Extraction quality
Automatic extraction can fail on paywalls, dynamic websites, or unusual page structures.
Result status
Technically checkedThe result passed automated technical checks but has not been independently reviewed by a human under the methodology.
Transparent development
The methodology is versioned
Changing the model, prompt, or rules can change the result. Every analysis therefore shows its technical provenance, and older results are not silently overwritten.