A compact guide to organisational knowledge that remains understandable, verifiable, and reusable by a team. AI did not create these principles, but it makes their absence more expensive by becoming another fast and literal consumer of corporate knowledge.
AI does not create corporate memory. It becomes useful when that memory distinguishes sources, assertions, evidence, inferences, decisions, and unknowns.
The arrow is a reading path, not a storage pipeline. In practice, knowledge forms an acyclic graph of typed relationships.
source ──evidence_for──▶ canonical assertion
│ │
└────derived_from────────▶ report
│
assertions + constraints ───▶ decision ──implemented_by──▶ action or code
Do not force epistemic state, document role, authorship, and lifecycle into one status. Apply each field only when it is meaningful.
| Field | Axis | Values | What it describes |
|---|---|---|---|
epistemic_state |
Epistemic state | verified, claim, hypothesis, unknown |
What is known and how strongly it is supported. verified always implies named verification criteria. |
record_kind |
Record role | source, assertion, decision, instruction, report |
What responsibility the record owns. Evidence is a typed relationship, not a competing record role. |
generation_mode |
Generation mode | captured, human_authored, machine_generated, mixed |
How the content came into existence. Derivation is recorded separately through derived_from. |
lifecycle |
Lifecycle | document: draft, active, archived; decision: proposed, accepted, superseded, rejected |
Publication state and entity state are separate. A draft cannot silently override active knowledge. |
Every material assertion has exactly one canonical owner within a declared scope. The owner is the record or system that maintains the current assertion; source evidence and the human steward are separate.
A source preserves what happened or was said. A canonical assertion records what the organisation currently accepts. A derivative transforms upstream material. None silently substitutes for another.
Provenance answers who produced a record, when, from what, and through which transformation. Verification separately states what the evidence proves, under which criteria, scope, and date.
Do not encode truth, role, authorship, and lifecycle in one label. Use independent fields so a machine-generated report can also be derived, active, and based on claims without contradiction.
Link to the canonical record using an explicit relationship such as evidence_for, derived_from, implements, or supersedes. A useful link explains what is there and why it should be followed.
Decisions, agreements, and instructions belong with the correct canonical owner; chat and email carry links and notifications. Code owns implementation, operational systems own transactional state, and documentation owns intent, rationale, and contracts.
Change meaning at the canonical owner first. Then inspect direct dependants and update, invalidate, or explicitly leave them unchanged. Conflicting active knowledge is a defect, not an alternative truth.
Repeating record kinds use a small schema with conditional fields. Separate document status from decision or delivery status; record as_of, review dates, and supersedes only where time or replacement is meaningful.
The useful atomic boundary is an independent owner, lifecycle, or change boundary—not every sentence. Split monoliths without creating a maze of fragments.
Every durable record is reachable from a relevant index. Start readers with a compact map and annotated links, then disclose deeper context only when the task requires it.
A human or AI answer, report, and decision identifies supporting evidence, separates quotation from inference, and preserves unknowns. Confident language is not a substitute for verification.
Scope, effective date, provenance, and access restrictions survive indexing, summarisation, and retrieval. Search and AI must not turn a local record into context-free common knowledge.
This example records exactly what the evidence proves: a customer made a statement. It does not silently promote that statement into proof that the money exists.
id: AST-042
record_kind: assertion
status: active
canonical_for:
- customer_budget_status
steward: sales-lead
as_of: "2026-08-14"
epistemic_state: claim
generation_mode: human_authored
evidence:
- path: correspondence/client-message.md
supports: customer_stated_that_budget_is_approved
derived_from: []
supersedes: AST-037
access: internalFor a longer Russian explanation derived from this registry, see «Корпоративная память, пригодная для AI».
- W3C PROV Primer — provenance through entities, activities, and agents.
- FAIR Guiding Principles — findable, accessible, interoperable, and reusable data for people and machines.
- Pro Git: About Version Control — change history as part of the working object.