Sentinel
Six signals, one verdict, and the working shown.
A claim is weighed across six independent signals. No single signal can make it verified, and a check that could not run is reported as not run, never counted for or against.
- Six independent signals
- Verified, or flagged
- Events, never individuals
The signals
What weighs a claim.
Cross-source corroboration
How much coverage from outlets graded credible carries the claim. Thin coverage is reported as thin; a search that could not run is reported as unavailable, never as evidence against.
The fact-check record
Published fact-checks of the claim. A fact-check never lifts a claim toward verified, and a published false rating is the one input that decides a verdict on its own.
Spread velocity
How fast the narrative is spreading. A story accelerating before anyone has verified it counts against the claim.
Prior-narrative correlation
Whether the claim matches a narrative already flagged, including one pushed across several channels at once.
Known manipulation patterns
Known manipulation patterns and their historical precedent. A match counts against the claim.
Models weigh, humans verify
A model’s own read of the text for manipulation. It is one signal of six: on its own it cannot make a claim verified.
The human step
A person decides what reaches the feed.
Verify
An analyst can verify a claim only with a confidence score attached.
Dismiss
Dismissing a claim takes a written reason.
Every decision is recorded with the analyst’s name and written to the audit log. The customer feed carries only claims a person verified; a request for anything else is refused, not quietly narrowed.
Physical observations
Measurements are kept apart from claims.
The ethic
Events and risks, never individuals.
Author names and handles are dropped before a scan is stored, and airspace is kept as counts, never identities.
Next