Guide takeaway
EKA is not trusted because an LLM sounds confident. It is trusted because the LLM operates inside a governed runtime.
The runtime knows who is asking, what subjects and features exist, which authority is being used, which tools are permitted, what evidence was actually returned, what calculations were performed, what is absent, what is inaccessible, and what the evidence still does not establish.
This chapter is the same at every reader level — it records no run anatomy to reveal.
These limits are part of the product's trust model, not hidden implementation details. EKA V1 does not provide:
- business mutations / commands
- external web evidence
- implicit currency conversion
- arbitrary SQL generated by the LLM
- arbitrary generated code as its normal calculator
- historical effective-policy determination without explicit effective-date data
- canonical losing / competing survivorship contributions
- long-term personalization
EKA currently
- READS
- RETRIEVES
- REASONS
- ANSWERS
A future enterprise agent may
- OBSERVE
- REASON
- PROPOSE / DECIDE
- ACT
Actions should not bypass enterprise commands and handlers.
Human UI ─────┐ Simulator ────┤ Agent ────────┼──→ governed command / handler Integration ──┘
EKA therefore serves two purposes: a useful enterprise knowledge assistant, and the first proving implementation of reusable agent-runtime foundations.
The written guide no longer depends on missing technical run evidence. Future enrichment is visual, and the recordings should reuse these same canonical questions rather than inventing prettier showcase scenarios.
- written guide + real run inspectornow
- UI screenshotsnow — chapters 05–08, replayed from these runs in the owner-only Workspace
- short embedded videosnow — one silent recording per run, with a transcript
- optional full demolater
When the internal UI arrives it should expose the same run anatomy rather than inventing a different mental model: final answer, citations, evidence drawer, authority indicator, feature-resolution notice, result status, calculation details, conversation history, and an optional technical run inspector.