The epistemic governance layer

Three primitives, one principle, a living system

The middle layer is not a database with better metadata. It is a governed substrate where every claim carries provenance, confidence, temporal validity, and a lifecycle that determines what actions it can support.

The three governed primitives

Primitive 1

Knowledge

The governed substrate. A bi-temporal semantic graph where every entry is a claim with provenance, confidence, and temporal validity. Not documents indexed. Not data collected. What the organisation currently has reason to believe, under explicit epistemic governance.

Not a knowledge base. Not a document store. Not a vector index.

Primitive 2

Context

The governed projection. When an agent or human needs to decide, the system constructs a bounded decision package - filtered by authorisation, scoped to the task, gated by epistemic quality. With a full manifest of what was included, excluded, and why.

Not "relevant documents retrieved." Not a prompt. Not a context window.

Primitive 3

Memory

The governed faculty. Bidirectional. Encoding writes operational experience back into Knowledge. Retrieval constructs Context for the next decision. Memory is the mechanism through which the system learns from its own operation.

Not a conversation buffer. Not chat history. Not a session log.

Intelligence emerges when all three operate with sufficient depth and governance. It is not a fourth primitive - you cannot build it directly. It appears. Three levels: Basic gives governed retrieval with provenance. Anticipatory gives proactive flagging - the system surfaces decaying claims and contradictions before they reach decisions. Initiative is the frontier - the system identifies goals worth pursuing through immersion, not instruction.

The atomic unit

Everything is a claim

Every piece of knowledge in the graph is a claim - a discrete, verifiable assertion carrying structured metadata. If you cannot express something as a claim with these properties, it does not enter the graph.

Provenance - where it came from, who validated it

Epistemic tier - earned, not assigned

Temporal validity - bi-temporal: world-time and graph-time

Confidence - five dimensions, computed

Decay class - how this knowledge ages

Dependency chain - what breaks if this changes

Scope - where this claim applies

Authority - who can promote or revise

The mechanism

Six steps. The circuit breaker is the spine.

The system is circular, not linear. Step six feeds step one. Every operational cycle deepens the substrate - making the next decision better informed than the last.

Step 1

Claims enter

Four acquisition modes: Harvest (automated), Extract (tools), Capture (expert-led), Emerge (graph self-extension). Everything enters at Provisional tier.

Step 2

Tiers are earned

Through validation, structural checking, expert review. Four tiers: Provisional, Emerging, Validated, Foundational. Computed from five confidence dimensions.

Step 3

Context is assembled

A bounded decision package: filtered by authorisation, scoped to task, quality-gated. The manifest logs what went in, what was excluded, and why.

Step 4 — The spine

Circuit breaker gates

Consequence classes map to minimum epistemic tiers. If supporting claims don't meet the threshold, the action halts or escalates. The system does not silently degrade.

Step 5

Memory writes back

Outcomes, feedback, observations encode back into the graph as new claims or evidence modifying existing ones. Every cycle enriches the substrate.

Step 6

The substrate deepens

Denser, better calibrated, more interconnected. The next context is higher quality. The next circuit breaker decision is better informed. This is the compounding loop.

The distinction that matters

Epistemic governance is not data governance

Data governance answers: who can access what, where is it stored, how long is it retained, which jurisdiction applies. It governs access and compliance. It says nothing about whether the content is still true.

Epistemic governance answers: is this still valid, what is its confidence, what contradicts it, what happens when it decays, and what actions is it currently supporting. It governs truth status and consequence.

You can have perfect data governance and still let an agent act on a claim that expired last week. The two governance types are complementary and both necessary. They are not the same discipline, and they require different organisational owners.

Epistemic tiers

Claims earn their authority

Tiers are not assigned by humans. They are computed from five confidence dimensions and earned through the lifecycle. Each tier determines which consequence classes of action the claim can support.

Provisional

Exists, untested. Cannot drive any consequential action.

Emerging

Structurally checked, passing automated validation.

Validated

Expert-reviewed. Eligible for operational use.

Foundational

Primary source. Highest tier. Eligible for autonomous action.

Knowledge has a shelf life

Managed decay

Knowledge rots at different rates. A regulatory interpretation decays on regulatory schedules. An operational assumption decays when business conditions change. A market assessment decays when the source updates.

Four decay classes govern this: Regulatory (expires on regulatory schedules), Operational (business-condition dependent), External (third-party source dependent), and Structural (rarely changes, massive cascade impact when it does).

When a claim decays past its validity window, its tier is downgraded. If that drops it below the minimum tier for any active consequence class, the circuit breaker fires. The system catches decay before it reaches decisions - not after.

Architectural properties

What this architecture produces

Not feature claims. Structural consequences of the three primitives and the circuit breaker principle.

Player-agnostic intelligence

Intelligence lives in the substrate, not the agent. Swap any model, upgrade any framework - the knowledge layer doesn't change. The graph is the asset. The Player is replaceable.

Bidirectional memory

Not a conversation buffer. Encoding writes experience into the graph. Retrieval constructs governed context. The system learns from its own operation under governance.

Operational compounding

Each cycle deepens the substrate. First domain is most expensive. Every subsequent domain is cheaper because the graph already carries transferable knowledge.

Contradiction preservation

Contradictions are maintained as typed information, never silently resolved. Two conflicting claims coexist with their respective provenance until human authority resolves them.

Bi-temporal audit

Two timelines: when something was true in the world, and when the graph learned about it. Reconstruct exactly what was known, by whom, at the moment any decision was made.

Go deeper

The full intellectual foundation

The architecture described here is specified in detail across a five-paper research programme, five public manifestos, and a set of named contributions that define the vocabulary.

Read the research →