How we build this

We believe you should see exactly how the information here is made — that’s what being independent means. No black box.

From a real decision to a plain-English summary

1. 1. We fetch the actual decision

   Every entry comes from a public Board of Veterans’ Appeals decision, published by the Board itself.

2. 2. AI reads and summarizes it

   A language model extracts the structured facts (what was claimed, the outcome, the deciding factor) and writes a plain-English summary. It reports only what the decision says.

3. 3. A trust gate hides anything uncertain

   A decision only becomes visible if the model’s confidence clears a bar and the outcome is clearly determinable. Low-confidence or ambiguous extractions stay hidden rather than risk misinforming you.

4. 4. We check the AI against itself

   We compare different models on real decisions (below) and spot-check against a stronger reference model, so we know where the summaries are reliable and where they’re weaker.

What’s in the library right now

1,290 vetted decisions visible. Outcomes: 204 denied · 402 remanded · 134 granted · 49 dismissed · 211 mixed. This grows continuously as more decisions are processed.

The models we evaluated, on real cases

We ran several open models over the same real Board decisions and measured how often they agreed with a stronger cloud reference model on the key facts, plus how fast they ran. We extract most of the corpus on our own hardware (private, no per-decision cost), reserving cloud checks for verification.

“Match” = agreement with a stronger cloud reference model on the same decisions, not a measure of absolute truth. Outcome (granted/denied/remanded) is the most reliable field; service-connection theory is the weakest and is the one we backfill with cloud checks. Speed is seconds per decision on local hardware.