Knowledge System of Record

AI agents can find
information.
But which page is real?

KSoR is an open-source SDK for building governed, authoritative knowledge systems for humans and AI agents. It solves the authority problem — not just the search problem.

View on GitHub →
The Authority Problem

Three AI helpers.
Three different answers.

A hospital uses three AI systems. A nurse asks all three the same question. All three sound confident. All three disagree. The problem isn't that AI can't find information — it's that no one told it which source the hospital considers official.

🏥 "How much medicine is safe for a six-year-old?"
AI AGENT 01
5 mg/kg
Source: old hospital wiki
✗ Ungoverned
AI AGENT 02
7.5 mg/kg
Source: LLM training memory
✗ Ungoverned
AI AGENT 03
4 mg/kg
Source: external web page
✗ Ungoverned
A computer can guess what looks relevant. It cannot guess what your organization has officially approved.
⚠️
Why AI agents disagree
Four root causes
LLMs are probabilistic — same question, different answer each time
Training data can be wrong, outdated, or contradictory
Models hallucinate when they lack information
Your org's private knowledge is never in the training data
✅
What KSoR provides
The governed layer
One canonical governed record — humans decide what is authoritative
Governance metadata: ownership, approval, version, audience
Abstention — the agent says "I don't know" instead of guessing
Citations — every answer traces back to its approved source
The Mental Model

Eight concepts behind KSoR

Each concept answers one question. Together they define how governed knowledge works for both humans and AI agents.

1
📖
Which page is real?
One official record
The hospital selects one governed record and declares it authoritative. Every AI helper must answer from that source. If the AI remembers something different, the authoritative record wins.
"A computer cannot choose which page is real. A person has to."
2
🗂️
What belongs here?
Governed knowledge only
Dosage rules → KSoR. Available beds right now → operational system. The distinction is not text vs database. It's governed knowledge vs operational state.
"The book holds what we decided. Other systems hold what is happening."
3
🔖
Can this be trusted?
Governance stamps
Written by: Dr Sana. Approved by: Medicines Committee. Version 4. Effective: 3 March 2026. Audience: Nurses. Without these stamps you have text, not institutional knowledge.
"No stamps, no trust. A page nobody signed is only a note."
4
🚪
Should this answer leave?
The governance boundary
Finding information and being allowed to disclose it are different things. The system may know something without being permitted to share it. No knowledge crosses a boundary without passing a governance decision.
"The book decides who may hear the answers."
5
🤐
What if there's no answer?
Abstention
If the KSoR has no approved dosage rule for infants, it says "The Knowledge System of Record does not contain enough information." A guess that sounds right is more dangerous than no answer at all.
"Abstention is a first-class feature, not a failure mode."
6
📋
How can I verify this?
Citation & provenance
Instead of "here is the safe amount," the system says "here is the safe amount — from Medicine Rules, page 7, version 4." The trace runs: AI answer → passage → document → git commit → approved source.
"Citation before confidence."
7
✅
How does knowledge become official?
Governance lifecycle
Dr Sana writes it Monday. Head nurse reviews Tuesday. Committee approves Friday. Only after approval may AI helpers use it as governed knowledge. Storage does not create authority — governance does.
"Knowledge becomes authoritative through governance, not through storage."
8
🔑
How do different users access it?
Many doors, one book
One governed record, multiple projections: a website for humans, llms.txt for AI discovery, an MCP interface for agents, a portable package for exchange. Fix an error once — all projections update.
"Govern knowledge once. Project it many ways."
Architecture

One governed record.
Many open projections.

The architecture separates concerns so institutional knowledge does not depend on any single vendor. Markdown is the authoritative medium. Git provides version control. Every other layer is a replaceable reference binding.

Markdown corpus
knowledge/
→
governance.yaml
.ksor/
→
KSoR Core
governed record
→
Fumadocs site
llms.txt
MCP server
OKF package
📄
Knowledge as Code
Stored in Markdown, versioned in Git. History, diffs, pull requests, approvals, rollback — all apply to knowledge governance.
🔐
Fail-closed access
No audience listed means no access. Quiet never means yes. Identity via OAuth/OIDC; governance rules enforced before any response.
🔗
Traceable to source
SLSA/Sigstore proves what was published. OpenTelemetry tells you what happened. Every answer is verifiable back to a git commit.
⚖️
Vendor-neutral
Postgres + pgvector for retrieval. Named technologies are replaceable reference bindings. The essential part is the governance semantics.
🔍
RAG
Retrieval-Augmented Generation
Asks: "What is relevant?"
Finds similar chunks via vector similarity
No ownership, approval, or lifecycle
Cannot abstain — will always retrieve something
Can be a retrieval component inside a KSoR
🏛️
KSoR
Knowledge System of Record
Asks: "What is the org authorized to operate from?"
Retrieval is governance-filtered before ranking
Ownership, approval, version, audience are mandatory
Abstains when governed knowledge has no answer
RAG, wikis, CMSs can all live inside it
Use Cases

Where KSoR fits

AI agents need governed rules + current facts
An accounting agent receives a $42,000 software cost. It needs to know if it can be capitalized. From the KSoR it gets the capitalization policy, definitions, and approval criteria. From the traditional System of Record it gets the invoice, ledger, and vendor data. It combines governed rules + current facts, then cites the rule it applied. If the policy doesn't cover the situation, it abstains — never invents a policy.
FROM KSoR (governed knowledge)
✓Capitalization policy
✓Approval thresholds
✓Exception rules
✓Accounting methods
FROM Traditional SoR (operational state)
→Invoice & ledger data
→Current balances
→Vendor information
One curriculum. Many AI tutors.
Without a KSoR, two students taking the same course get different AI tutors teaching different concepts at different depths. Personalization has quietly turned into inconsistency. A KSoR governs the canonical curriculum, learning objectives, approved sources, and assessment rules. Multiple tutors can then personalize how they teach while staying inside the same academic truth. Panaversity's Agent Factory is the working example.
KNOWLEDGE SoR governs
✓Canonical curriculum
✓Learning objectives
✓Approved sources
✓Assessment rules
LEARNER SoR records
→Student progress
→Completions & grades
→Current study position
Domain-specific governed knowledge
A Vertical KSoR is simply one whose authoritative scope is a particular profession or industry. It's not a different technology — just a different scope. A Method KSoR governs a reusable way of working. Agents can compose multiple KSoRs: the Method KSoR tells it how to work, while the Vertical KSoR tells it what is authoritative in the domain.
Example vertical domains
✓Healthcare — dosage, clinical protocols
✓Accounting — capitalization, compliance
✓Legal — procedure, jurisdiction rules
✓Government contracting — policy, approval
✓Banking — credit analysis, thresholds
✓Insurance — coverage, underwriting rules