Show HN: VOR – A verified runtime with 0% hallucination via observations
1 points
1 hour ago
| 0 comments
| HN
Hi HN — sharing a system I’ve been building that takes a hard line on hallucinations. VOR (Verified Observation Runtime) is a minimal, auditable runtime for QA and tool-based reasoning that refuses to answer unless the answer is provably grounded in observed evidence. Key properties: 0% hallucination (enforced, not tuned) Explicit ABSTAIN / CONFLICT / ANSWER outcomes Deterministic evidence chains (hashable, replayable) Strict parity: baselines and verifier see identical observations Witness-green: anyone can reproduce results from a clean clone What this is not: Not AGI / not alignment theory Not probabilistic “confidence scoring” Not a chatbot replacement What it is: A runtime you can put under LLMs, agents, or retrieval systems A way to turn “trust me” outputs into “prove it or abstain” The public repo is intentionally small: kernel + CLI + demo packs. Private components (corpus builders, generators) are explicitly excluded. Repo: https://github.com/CULPRITCHAOS/VOR Witness instructions: in docs/WITNESS_RUN_MESSAGE.txt I’m specifically looking for: Repro attempts on other machines Adversarial inputs I missed Critiques of the proof/abstention model Happy to answer technical questions.
No one has commented on this post.