User-Provided Knowledge

Goal

Describe how a user or project can add a knowledge base for use by agents and skills without that addition turning into a document dump in the prompt.

This is the human-side surface of the Knowledge Governance Layer.

What “adding knowledge” means

When a user attaches a framework, policy, persona pack, research report, or design guideline to a project, they are doing four things at once — even if the UI hides it:

declaring a source (file, link, internal doc)
claiming a purpose (what tasks it should inform)
classifying it (sensitivity, authority, owner)
constraining its use (citation, exposure, export)

If only step 1 happens, the system has a document. It does not have governed knowledge.

What user-provided knowledge is not

It is not a system prompt extension.
It is not raw paste into agent context.
It is not a memory store.
It is not a fine-tuning dataset.
It is not an authority override for compliance, security, or accessibility sources.

The minimum a user provides

To register a base, the user supplies (or is asked for) at least:

a short name and id
a type (see source taxonomy in knowledge-source-contract)
an owner
a sensitivity level
an authority level
a scope (which task families this source is allowed to inform)
a retrieval mode preference (capsule-first, excerpt-only, etc.)
a usage policy (citation, exposure, export)

This is captured in a knowledge pack manifest.

Why not just paste the document

Pasting bypasses every guardrail this layer exists for:

no version, no rollback when the document changes
no owner, no review cycle
no scope, so the document influences tasks it should not
no authority level, so a persona suggestion can outrank an accessibility rule
no audit, so no one can answer “what did this source change?”
no exposure limit, so restricted text can land in logs, traces, and handoffs

A document without manifest is a leak waiting for a trigger.

Progressive maturity

A user does not need a full manifest on day one. The system should accept:

Level	What the user provides	What the system can do
0 — Unregistered	nothing	refuse to use as governed source
1 — Minimal	id, owner, sensitivity, scope	use only in low-risk tasks, no citation, no export
2 — Operational	+ authority level, retrieval mode, capsule	use in declared scope with capsule-first
3 — Governed	+ usage policy, review cycle, audit fields	full participation under the resolver

A skill that requires governed knowledge should be unable to bind to a level-0 source. A skill that accepts minimal sources should mark its output with an uncertainty / lower-authority signal.

Failure modes to make visible

When user-provided knowledge cannot be used safely, the system must say so — not silently swap to a generic mode:

“This source is not registered. Either register it or proceed in generic mode.”
“This source has authority interpretive but the task requires a mandatory source. Cannot proceed without a higher-authority anchor.”
“This source is confidential. It cannot be cited verbatim in an external-facing deliverable. Capsule-only mode applied.”

Loud refusal is a feature.