BS Score measures how dense the tracked manipulative patterns are in a piece of content or in an aggregated peer-group record.

• For a single analysis run, BS Score is the average of the percentages for all tracked reasoning attributes in that run, including attributes that appear 0 times.
• Each attribute percentage is calculated as attribute count divided by analyzed word count. If one pattern appears 57 times in 1,024 analyzed words, its percentage is 57 / 1,024 = 0.055664, or 5.5664%.
• Quoted language is not treated the same way for every peer type. Article, video, publication, topic, and keyword scores use the full analyzed text, including quotations. Author scores use only the author or narrator portion of the text and exclude words attributed to named speakers. Speaker scores do the opposite: they use only the words attributed to that speaker.
• For article and video rankings, antiBSmachine then averages the available run-level scores and the overall aggregate score, and that final average becomes the ranked article or video BS Score.

Example: Birth trauma is common. One Utah mom explains why better support matters⁠24%. In the gpt-5.4-nano analysis for that article, the system analyzed 1,024 words across 55 tracked attributes. Each attribute gets a percentage equal to count / wordCount. For example, Confirmation Bias was flagged 57 times, so its percentage was 57 / 1,024 = 0.055664, or 5.5664%. Anchoring Bias was 38 / 1,024 = 0.037109, or 3.7109%. Availability Heuristic was 65 / 1,024 = 0.063477, or 6.3477%.

The gpt-5.4-nano stats file for that same article contained 55 attribute percentages whose total sum was 1.740233. antiBSmachine divides that by all 55 tracked attributes, including the ones that were 0, producing 1.740233 / 55 = 0.0316406. That is the gpt-5.4-nano run’s BS Score, or 3.16406%.

The article-level BS Score used for ranking is then averaged across the available scores for that item. For this article, the stored inputs were: overall aggregate score 0.0191672409, gpt-5.4-nano 0.0316406, claude-haiku-4-5-20251001 0.0105291091, and grok-4-fast-non-reasoning 0.0028586545. Their average is (0.0191672409 + 0.0316406 + 0.0105291091 + 0.0028586545) / 4 = 0.0160489011, or 1.60489%.

That quote-handling rule matters most once the same article is rolled up into entity rankings. The publication score keeps the full article, including quoted passages. The author score keeps only the author or narrator track from that article. If a named speaker is quoted, that speaker’s words move into speaker stats instead of being counted as author words.