HSI runs a five-stage pipeline. At every stage, one rule holds: every claim must be traceable back to its source.
Public posts, articles, transcripts, and campaigns are captured and archived with timestamps and cryptographic hashes.
Models extract claims, entities, and rhetorical structures from raw material — always preserving a link to the exact source passage.
A structured classifier interprets framing, technique, and influence patterns, attaching a confidence score to every inference.
Trained reviewers verify, challenge, or reject each interpretation. Models assist; humans make the published call.
Only verified, fully-traceable findings are published — with the evidence chain available for public inspection.
This is the non-negotiable foundation of HSI. If a conclusion cannot be walked backward to a verifiable source, it is not published.
Every AI inference carries an explicit confidence score so reviewers and readers can weigh it appropriately.
We publish what our models can and cannot do, including known failure modes and category boundaries.
Archived evidence and versioned models mean findings can be re-examined and independently audited.