By design, the agent's judgement is semantic/rhetorical, not ideological
What Agent 2 does: scores how things are said (framing, loaded language, vague attribution, causal leaps, apples-to-oranges, missing context cues), with guards for quotes/attribution/hedges/satire and the flag vs. effect split.
Caveats (honest ones): any curated ruleset carries subtle choices, and regex/guard vocab can interact with specific beats. But the detector isn’t checking “who’s right,” only linguistic patterns.
We will however put this to the text: add symmetry tests (swap actors, keep syntax) to check it stays ideology-agnostic.
Tech notes at the bottom.
Loading...
Notes
This pass was classic Infowars: heavy framing and loaded language, plus vague attribution ("critics say," "sources claim") doing a lot of work.
Agent 2 threw reds when allegations were presented as facts without sourcing, when causal leaps were asserted from thin evidence, and when apples-to-oranges comparisons were used to imply motive or hypocrisy.
We downgraded mere tonal jabs to flags, but kept effects for stacked cues (e.g., absolute certainty + comparison lexicon + no guard tokens).
Net result: fewer cheap reds for bluster, consistent reds for unverified claims framed as truth.
