Chapter 02 of 11

The governed runtime

EKA is the first implemented agent-style product built on a reusable technical foundation. It is not intended to become a one-off assistant whose foundation must be rebuilt for every new agent.

The shared foundation

                 Metrolane Agent Foundation
                           │
          ┌────────────────┼────────────────┐
          │                │                │
         EKA        Future Agent A    Future Agent B

The foundation owns

identity / viewer context
access enforcement
tool definitions and execution
Queryable Subject Registry
semantic query validation
authority handling
feature resolution
evidence and citations
temporal semantics
deterministic analytics
result sets and task state
historical replay security
budgets, retries and observability

EKA adds

mission
planner instructions
subject / tool exposure
EKA-specific orchestration
answer behavior
evaluation cases

Access is part of the answer

EKA operates as a governed viewer, and every tool is viewer-aware. The access check happens before evidence reaches the model — not by reading everything, handing it to the model and asking the model not to reveal it.

NOT_PRESENT

The requested feature is not represented there.

PRESENT_BUT_UNREADABLE

The feature exists, but the current viewer cannot use it.

Authority is part of the answer

Enterprise data exists at different authority levels: canonical, source, document, scenario, data journey and derived calculation. EKA does not assume one authority is universally superior. The relevant question is which authority is appropriate for this claim.

canonical amount unavailable≠permission to silently substitute an ERP amount

policy passage≠recorded rejection reason

scenario≠live current state

The separation

What the LLM decides — and what it does not

The LLM decides

what the question means
how to decompose it
which governed tools are relevant
what dependencies exist
whether more evidence is needed
how to explain the result

The deterministic runtime owns

viewer / access
schema validation
tool validation
subject / feature validation
query execution
calculation
budgets
result-state mechanics
evidence normalization
citation validity

The current implementation

EKA v0.4.1 is headless, read-only, server-side, on live Postgres, and CLI-driven for development and evaluation. The conversational UI is the next product layer; it does not require redesigning the headless core.

13 governed read-only tools
viewer-aware reads
Queryable Subject Registry
feature resolution
typed evidence
deterministic aggregation
document retrieval
business-time semantics
conversation / result-set persistence
classified saved-answer replay
prompt provenance
40-case golden evaluation set

Chapter takeaway

The foundation is reusable; EKA is its first consumer.
Access is enforced before evidence reaches the model.
Authority is chosen per claim, never substituted silently.