Source Reliability

How we evaluate our sources

Every Telegram channel monitored by OSFeed is profiled across 5 dimensions. These profiles determine how each source is weighted in event summaries, contradiction detection, and corroboration scoring. No source is trusted blindly.

Five Profiling Dimensions

Reliability

HighMediumLow

Historical accuracy of the channel's factual claims. A 'high' rating means the channel consistently reports verifiable facts. 'Medium' means the channel is sometimes accurate but has a track record of exaggeration or premature claims. 'Low' means the channel frequently publishes unverifiable or demonstrably false information.

Pipeline impact: High-reliability sources take precedence for factual claims in event summaries. Low-reliability sources are included only when corroborated by at least one independent source.

Bias

Pro-RussiaPro-UkrainePro-IranPro-IsraelPro-WestNeutralMixed

The channel's consistent editorial lean. This is not a value judgment — it describes the framing pattern. A channel with a known bias can still be high-reliability: it reports facts accurately, but selects and frames them through a consistent perspective.

Pipeline impact: Bias tags are visible alongside source attribution in event cards. When sources with opposing biases report the same facts, corroboration confidence increases. When they contradict on facts, the event is flagged as contested.

Content Type

Breaking NewsAnalysisPropagandaAggregatorOfficialMixed

What the channel primarily publishes. Breaking news channels report events as they happen. Analysis channels provide context and interpretation after the fact. Propaganda channels have a clear persuasion objective. Aggregators repost from other sources. Official channels represent government or military institutions.

Pipeline impact: Breaking news sources are weighted for speed of detection. Official sources are weighted for authoritative claims. Aggregators are used for corroboration counting but not treated as independent sources if they repost from channels OSFeed already monitors.

Region Focus

e.g., Middle East, Ukraine, Sahel, Global

The channel's primary geographic coverage area. Channels are assigned to the topic that matches their regional focus.

Pipeline impact: Ensures event detection uses contextually relevant sources. A channel focused on Ukraine is not expected to provide coverage of Sahelian developments — so its absence from a Sahel event is not considered a coverage gap.

Language

Detected primary language

The channel's primary publishing language, detected through multi-stage analysis: character-set heuristics, lingua library classification, and Telegram metadata fallback.

Pipeline impact: Drives the translation pipeline. Messages in the channel's primary language receive optimized translation with domain-specific glossary injection.

How Profiling Works

Channel profiling is automatic. When a new channel is added to OSFeed, the following process runs without human intervention.

Step 1Message Sampling

When a channel is first added, OSFeed fetches the last 50+ messages and selects at least 5 substantive messages (>20 characters) as a representative sample.

Step 2LLM Profiling

The sample is sent to an LLM with a structured profiling prompt. The model evaluates factual language vs. sensationalism, identifies consistent framing patterns, classifies content type, and detects geographic focus.

Step 3Profile Generation

The LLM returns a structured JSON profile with all 5 dimensions. This profile is stored alongside the channel record and used in all downstream pipeline decisions.

Step 4Periodic Refresh

Profiles are not static. As channels evolve — a formerly neutral channel adopts a clear bias, or a breaking news channel shifts to analysis — profiles are refreshed from recent message samples.

How Profiles Affect Event Synthesis

Source profiles directly influence how events are summarized, which claims are prioritized, and when contradictions are flagged.

High-reliability neutral sources

Take precedence for factual claims in event summaries.

Opposing-bias corroboration

When sources with known opposing biases report the same facts, confidence increases significantly.

Low-reliability or propaganda sources

Included in event source lists but only surface in summaries when corroborated by at least one independent source.

Aggregator deduplication

Aggregators that repost from channels OSFeed already monitors are counted for corroboration but not treated as independent sources.

Regional relevance

Sources covering their primary region are weighted higher than out-of-region channels reporting on the same event.

Source Count as Confidence Proxy

The number of independent sources reporting an event serves as a confidence indicator. But it is not a raw count — it is contextual:

  • If you follow all source channels, the event card shows “N sources”
  • If you don't follow all channels, it shows “N of your sources · M others” — without revealing which
  • Higher count = higher corroboration = higher confidence
  • Used in event ranking: Breaking (>10 sources) → Significant (5–10) → Emerging (2–4)

Current Coverage

OSFeed monitors 44+ channels across 4 geopolitical topics. Coverage is continuously expanded as gaps are identified.

Ukraine-Russia Conflict

Channels~15
LanguagesRussian, Ukrainian, English
Military channels, milbloggers, official government accounts, independent OSINT, diplomatic sources

Middle East Tensions

Channels~12
LanguagesArabic, Hebrew, Farsi, English
Resistance-aligned channels, Israeli military/government, independent conflict monitors, diplomatic sources

Sahel & West Africa

Channels~8
LanguagesFrench, Russian, English
Francophone African channels, Russian state media, independent security analysts, regional news

USA & Global Power Shifts

Channels~9
LanguagesEnglish, Russian, Chinese
Economic/financial channels, policy analysis, state media, diplomatic sources

Evaluate our sources yourself

Every event card in OSFeed shows source attribution, reliability ratings, and bias tags. Join the beta to see the full picture.

Join the Beta