Source verification for defense and national security. Sentinel checks a claim against a typed source corpus, scores how strongly the evidence supports it, and seals a verdict with an artifact. Built so an analyst can rely on it and an evaluator can reproduce it.
Sentinel finds the sources that speak to a claim, weighs how strong and how current each one is, and returns a ruling with a confidence score and an artifact. Same claim, same sources, same ruling, every time. No AI sits in the verdict. It follows fixed logic, phase by phase, and every phase is instrumented and traceable.
Each phase is timestamped and snapshotted. The trace is the product. It is what lets an evaluator see exactly how a verdict was reached.
Breaks the claim into what has to be verified and identifies the claim type, from factual and temporal to quantitative, geospatial, causal, and more.
Searches the source corpus for matches against the claim and scores how well each source speaks to it, returning the strongest matches in ranked order.
Weighs each source by age and type, and by reliability and credibility on a standard scale, so stale or weak sourcing is surfaced rather than hidden.
Projects the verdict through the scored evidence, applies verdict precedence, and runs contradiction detection before any ruling is formed.
Checks that the engine's own reasoning stayed stable across every phase. In a clean deterministic chain this reads zero. Any drift is flagged.
Runs the detectors, computes the determinism hash over inputs and verdict, and assembles the full reasoning chain into a sealed, hash-sealed result.
How strongly the sourced evidence supports the claim, combined from correlation, currency, and source weight. One number an analyst can stand behind.
The distance between the engine's external reading and its internal reasoning. In a deterministic chain it should read 0.00. Anything higher is a flag, not a footnote.
Every ruling seals with a determinism hash over its inputs and output. The same claim and sources reproduce the same verdict and the same artifact.
Sentinel is designed to align with the direction of the DoD Responsible AI pathway and the intelligence community's risk-management posture, for environments where a wrong answer is expensive and every call may be reviewed. It produces the one thing those environments demand and probabilistic tools cannot give. A verdict that is the same every time, anchored to its sources, and reproducible by the people who have to trust it.