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.
1 comment
Show HN: VOR – A verified runtime with 0% hallucination via observations | Heykuki News