A relevant passage is not enough evidence

Most AI assistants find the passages that look most relevant and answer from them. That works until the right-looking passage is missing the exception, the newer policy, or the document that says the opposite. Hive AI has published a draft research paper, KAR-1, about a stricter question: do we have enough of the right kinds of evidence to answer, and if we don't, can we show that the system refused to guess?

Read the preprint: KAR-1: Frontier-Constrained Minimum Evidence Retrieval over Committed Knowledge Images. It is a non-peer-reviewed preprint (October 2026), not a finished, reviewed paper.

Why this matters to a business owner

Picture a chatbot that answers refund questions from your policy documents. A customer asks whether a sale item can be returned. Search finds the general 30-day return policy, which is highly relevant, and the bot says yes. But a newer page says clearance items are final sale. The answer sounded grounded, and it was still wrong.

Relevance tells you a passage is about the topic. It does not tell you whether you have the passage that qualifies it, the one that contradicts it, the one that is current, or the one that has the authority to decide. KAR-1 is our attempt to make those requirements explicit instead of hoping the top search results happen to cover them.

What KAR-1 proposes, in plain terms

• Declare what an answer needs. Before answering, the system states its evidence requirements: something that supports the answer, a check for anything that opposes or qualifies it, evidence that applies to the right date, and a source with the right authority.
• Find the smallest set that meets them. Instead of a long ranked list, KAR-1 looks for a minimum evidence set that satisfies every declared requirement.
• Fail closed. If the requirements can't be met, it says so. The preprint states it "does not silently substitute a heuristic top-k fallback."
• Leave a receipt. Each result comes with a certificate that someone else can replay and verify against the same committed knowledge.

One honest limit, in the paper's own framing: a result marked SATISFIED means the declared evidence requirements were met. It does not mean a claim is proven true.

A short technical section

KAR-1 (Knolo Adversarial Retrieval) runs over a committed Knowledge Image and a Committed Evidence Graph. It keeps support, opposition, qualification, temporal/applicability and authority as separate evidence frontiers rather than folding them into one relevance score, then decides whether a declared retrieval plan can be satisfied by a minimum applicable evidence set. Outcomes are cryptographically committed, so a certificate can be replayed and checked independently.

The preprint describes its own validation work: TypeScript/Rust parity checks, 50,000 differential cases, 10,000 exhaustive exact-cover comparisons, and 10,000 tamper mutations in which no tampered certificate was accepted. It also notes that automatic Domain Pack extraction still needs review pending human-gold validation. These are the authors' internal tests reported in a draft, not independent results.

KAR is available in Knolo 6.0.0 on npm as @knolo/core/kar (package). The Python and Rust packages don't include KAR yet. Knolo is owned and maintained by Hive Forensics AI Inc. More technical detail is on hiveai.tech/knolo.

What this means for the systems we build

You don't need to read a research paper to get the benefit. The idea behind KAR-1 shapes how Hive AI builds chatbots, voice agents and AI integrations: decide what evidence an answer needs, refuse to bluff when it isn't there, hand off to a person instead, and keep a receipt you can check later.

• Want your team to learn to build AI this way? Join the Hive AI Bootcamp.
• Want a chatbot or voice agent that knows when not to answer? Talk to Hive.