Current shipping contract
V4 Knowledge Images
Portable .knolo files with deterministic lexical evidence, receipts, and local execution. Open-source contract lives in knolo-core.
OPEN SOURCE · knolo-core
github.com/HiveForensics-AI/knolo-core →
Platform / Knolo
How this company implements Knolo.
Knolo is a verifiable knowledge runtime for agents. Hive Forensics AI is the builder that images it, permissions it, evaluates it, and runs it inside HIVE systems. Product storytelling lives on HIVE.
We build AI systems on Knolo, a verifiable knowledge runtime for agents. Today that starts with V4 Knowledge Images: portable .knolo files with deterministic lexical evidence, receipts, and local execution. V5 extends the same file into a transactional runtime — one ledger for knowledge, policy, memory, authority, and agent runs.
This page is the implementation view: what we actually do when a corpus has to survive counsel, a CISO, and a Monday morning operator. We do not treat Knolo as an SDK sticker. The Knowledge Image is the unit of work.
Product pages, open-source V4, and docs are one runtime — not three products. This domain explains how the company implements and operates it.
Knolo contract
V4 ships. V5 is the runtime upgrade.
We build AI systems on Knolo, a verifiable knowledge runtime for agents. Today that starts with V4 Knowledge Images: portable .knolo files with deterministic lexical evidence, receipts, and local execution. V5 extends the same file into a transactional runtime — one ledger for knowledge, policy, memory, authority, and agent runs.
Upgrade in progress
V5 Verifiable Knowledge Runtime
Architecture proposed and being implemented — not GA. The same file becomes a transactional runtime: one ledger for knowledge, policy, memory, authority, and agent runs.
NOT GA · SAME .KNOLO FILE
V4 evidence path
What a Knowledge Image has to prove.
01 / LEXICAL GROUNDING
The default retrieval path is deterministic and lexical. Embeddings, if used, rerank behind that contract. They are not the trust root.
02 / SPANS
A hit points at a location in the image — source, region, passage. An answer that cannot point is a failure, not a style.
03 / PLAN HASH
The retrieval plan is hashable. Same question, same image, same policy: the same evidence set, replayable later.
04 / READ RECEIPT
A query leaves a receipt. Counsel can ask what was read. The receipt is part of the system, not a chat log export.
V5 runtime path
Same file. Transactional runtime. Not GA.
Architecture proposed and being implemented. We will not describe it as generally available.
01 / STATE ROOTS
V5 pins runtime state so a run can be named, compared, and resumed. Proposed and being implemented — not GA.
02 / WRITE RECEIPTS
Mutations to knowledge, policy, or memory are receipted. An unreceipted write is not a write we will operate.
03 / RUN RECEIPTS
An agent run is a record: what it read, what it attempted, what was allowed. Fluency is not the log.
04 / DEFAULT-DENY AUTHORITY
Tools and side effects require an approval or capability record. No implicit toolbox because a demo looked busy.
05 / RESUMABLE RUNS
A run that suspends for a human can continue from the same root. Abandoning the trace is not an architecture.
What we will not do with it
Embedding-first retrieval as the trust root
Approximate nearest neighbors are not a permission model and not an audit trail.
Agents that act without a record
If there is no approval or capability record, the tool does not run.
Knowledge that cannot be replayed
If it cannot be mounted, pointed at, and replayed, it does not ship.
Broader stack: architecture. Data handling: security. Product language: hiveai.tech/knolo.
Engage
Work starts on HIVE Bootcamp.
Unscoped work starts at hiveai.tech/bootcamp. This domain is for engineering already constrained. Work is scoped privately.
Scoped privately · no public rate card on this domain