← Blog

· Hive AI

Knolo Query Receipts: What a Receipt Proves, and What It Doesn't

A Knolo query receipt binds an image, a statement and its hits so anyone can replay them. It is an integrity check, not proof an answer is true. What changed on 9 Nov.

Knolo is the open-source, local-first Knowledge Image runtime and format, maintained by Hive AI under Apache-2.0 and published as @knolo/core on npm. A Knowledge Image is a bounded, content-addressed artifact. One of the things Knolo's open contract has to define is what a query leaves behind, and the answer is a receipt.

This note sets out what a Knolo query receipt proves, what it deliberately doesn't, and what changed in the public source on 9 November 2026.

What a Knolo receipt binds

Plan and result roots make a query inspectable against one committed state. A receipt ties three things together: the image that was mounted, the statement that was run, and the hits that came back. Replaying it answers one narrow question. Do the same image and the same statement produce the same plan and the same hits?

The developer side is in the Knolo query receipts docs and roots and receipts. On the command line, knolo query --receipt writes one and knolo explain reads it back against the pack or image you name.

What a Knolo receipt does not prove

A receipt does not say the hits are enough to answer a question, and it does not say an answer built from them is true. The documentation states this directly: a receipt is an integrity check. Retrieved text stays untrusted evidence. Keeping that line clear matters more than any single feature, because a record that claims more than it checks stops being useful to an independent reviewer.

The same principle runs through the Knolo specifications page: unpublished is better than invented.

Knolo receipts fail closed

The host-agnostic service in @knolo/tools writes an openclaw-knolo-query-v1 receipt, with a qr_ id, before search results return. If the audit directory can't be written, the call fails with AUDIT_UNAVAILABLE and the host returns no hits. A search that can't be recorded doesn't happen.

Caller identity comes from host configuration, not from tool arguments. A model talking to the tools cannot rename itself on the record.

What changed in the Knolo source this week

Pull request #139 in the Knolo repository, merged on 9 November 2026, adds two fields to new openclaw-knolo-query-v1 receipts written by the local knolo-box work:

  • hitCount, the number of allowed hits, which verify now checks.
  • durationMs, which covers the query and the policy check but not the audit write. Verify does not replay the clock.

Receipts written before those fields still verify. That is the compatibility rule in one line: a new field may add a check, but it may not break an old record. The change shipped with no version bump, and @knolo/box is not published to npm. Published @knolo/core, @knolo/tools and @knolo/mcp stay as they are.

Why this belongs in the open Knolo record

Receipts are where a runtime's claims meet an outside check. When their meaning is written down in public source and documentation, a second implementation can be held to the same contract. That is the work the Knolo compatibility page and Knolo KIPs exist to carry.