AI Output Is Not Knowledge: Add a Candidate Layer and Human Review

Do not write every AI output into a knowledge base. Use a candidate layer that preserves sources, scope, uncertainty, sensitivity, and human review.

# AI Output Is Not Knowledge: Add a Candidate Layer and Human Review

An AI system can converse, research, create documents, and retain memory. None of those outputs should automatically become validated knowledge. A safer approach adds a candidate layer before the knowledge base: record sources, scope, uncertainty, sensitivity, and a human reviewer before deciding whether to accept the material. That prevents guesses, stale sources, task notes, and private content from being institutionalized together.

Governance boundary: knowledge-outbox and the main-only gate are candidate mechanisms proposed in this article. They are not built-in Hermes knowledge-base features and do not automatically sync content into an accepted knowledge base.

Conversation, task output, memory, and knowledge are different things

A Session is a history of conversation and tool activity. It can help trace sources, reasoning, and process, but it is not organized knowledge that is ready for reuse. Hermes Sessions documentation describes saved messages, tool calls, model settings, and timing information; a preserved record does not automatically become correct. [Sessions documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/sessions.md)

Memory is limited durable context for future Sessions. A Task output answers a particular work problem and may be useful without being reusable elsewhere. Validated knowledge needs a higher standard: traceable sources, a clear scope, visible unknowns, an assessed sensitivity level, and an explicit human acceptance decision.

Use a candidate gate before acceptance

In the model proposed here, a specialist returns task outputs, sources, and uncertainty. Only the Work Chief of Staff can create a knowledge-outbox candidate packet. Before doing so, it checks:

  1. Does this have value across more than one task?
  2. Where are the original sources and task artifacts?
  3. Which parts are facts, inferences, and untested assumptions?
  4. Does it contain private, sensitive, or non-public information?
  5. Who will review, accept, narrow, revise, or reject it?

Outbox means a candidate awaiting human review. It is not the knowledge base and does not trigger synchronization. The separation between producer and acceptor prevents speed from replacing judgment.

What a candidate knowledge packet can contain

This structure is fictional. It is not a real Hermes runtime packet:

candidate_type: reusable-principle
claim: For status updates that do not require real-time collaboration, start with an asynchronous written update before deciding whether to meet.
source_task: research-task-example
source_artifacts:
  - literature-summary.md
evidence:
  - Add official or primary-research links here
scope: Routine status coordination in knowledge-work teams
uncertainties:
  - Not tested for emergencies or highly interdependent work
sensitivity: public
review_status: pending-human-review
proposed_destination: team-collaboration-guidance

A reviewer may accept it, narrow its scope, request stronger evidence, or reject it. If the source supports only some teams under certain conditions, the claim cannot become “every remote team should do this.” The candidate layer makes that reasoning boundary visible before formal retention.

Why not let every worker write directly?

Direct writing removes more than a step: it removes a quality responsibility. Specialists are closest to details and can be most likely to generalize a local conclusion. A researcher's source list, a writer's draft, and an operator's temporary workaround should stay in their Task evidence context first. Only material that passes reuse and boundary checks deserves to become a candidate.

When the thing to retain is a repeatable procedure, a Skill may be a better form than a memory entry or a knowledge-base document. Hermes' Creating Skills guidance explains how to package an expressible procedure as a Skill; it does not mean every work output should become institutional knowledge. [Creating Skills documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/developer-guide/creating-skills.md)

Limits and next step

A single gate may become a bottleneck or overlook specialist judgment. It needs a standard candidate format and regular review, or it only moves noise to another folder. Before adopting the model broadly, require a reusable claim, sources, scope, uncertainty, sensitivity, and reviewer for a small set of real candidates, then measure the friction and quality.