Launching today
PrismShine
Self-hosted anti-hallucination verdict for LLM apps
4 followers
Self-hosted anti-hallucination verdict for LLM apps
4 followers
PrismShine is a self-hosted anti-hallucination verdict engine. It runs cause-side forensics on runtime evidence, then grounds the answer against your preload — one auditable ShineVerdict with a named resolution_gate. Apache-2.0. pip install prismshine.
The resolution_gate concept is genuinely useful for tracing back why an answer got flagged. One thing I'd love to see is a way to export the verdict history as a structured log so I can plug it into my existing observability stack and correlate spikes in hallucinations with specific deploys or prompt changes.
@rhastekkes51150 Thanks Rabia — that’s exactly the use case we designed resolution_gate for.
Today every ShineVerdict is already a structured record (decision + named resolution_gate + evidence_hash + signatures/spans). Pattern most people use: log the verdict JSON next to your request metadata (deploy SHA, prompt version, model), then spike on decision!=pass or on specific gates in whatever you already run (Datadog / Loki / BigQuery / etc.).
What we don’t ship yet is a built-in “verdict history → observability sink” exporter — your note is on the short list. If you drop which stack you use (OTel? Datadog? something else), that helps us prioritize the shape.
Demo if useful: https://insightitsgit.github.io/PrismShine/demo.html
tried it on a flaky agent log and honestly the resolution_gate label made debugging way less painful than usual. basically exactly what i needed
@berkecakal35596 Appreciate that, Berke — glad resolution_gate earned its keep on a real flaky log. That’s the whole point: name the gate, don’t leave you with a mystery score.
If anything felt confusing in the verdict fields, shout — happy to clarify.