Time, conversations and provenance
Three things decide what an answer means later: which clock it used, what a conversational reference still refers to, and which configuration produced it.
Metrolane has an enterprise clock, so “the last ten days” may refer to Metrolane business time rather than the wall clock when the user asks. The subject's governed temporal semantics determine which one applies.
BUSINESS_NOW
The enterprise clock. In these captures, 2026-07-31 18:30:00+00.
REQUEST_TIME
When the question was actually asked.
A future guide example can demonstrate this visually once the UI exposes the time interpretation clearly.
Show the open orders. Which of those came from CRM? What are their statuses now?
A result set preserves what “those” referred to. But the result set is not a frozen copy of current truth. On reuse, historical membership is re-authorized for the current viewer and current values are re-read, so the conversation keeps its meaning while still respecting current access and current enterprise state.
A saved answer is what EKA established at that historical run, not automatically current truth. Asked what the status is now, EKA performs a current read. Asked what changed since that answer, EKA compares the historical observation with current governed evidence, subject to current access. Saved value-bearing content remains classified and replay is governed by current authorization.
EKA's behavior is not identified by the model name alone. A run is influenced by more than a source file called “the prompt”, so the canonical runs record the exact static configuration that produced them.
The effective planner prompt includes
- mission
- planner instructions
- rendered subject registry
- rendered tool catalogue and JSON schemas
- page / task context
- runtime feedback on re-plan
- user question
The synthesis prompt includes
- mission
- authorized evidence
- runtime notes / feature resolutions
- user question
Recorded fingerprints
- Prompt bundle
- PROMPT-EKA-0.4.6
- Planner hash
- sha256:0fb08227cf1e04d7a3687b2e6af6d6e84ab3b7918f1520bae94c26684a899485
- Synthesis hash
- sha256:2e56d7b0237722db332278e6e89e0b8fd70703b1677a1d8e8286e4502a67a903
- Mission hash
- sha256:48c62bdaf2298fdb3e737599a0dc6ea2bd72f75dc8b70072dcf70cacf24699c1
- Planner-note hash
- sha256:68f2a28199d9c52c37fd0b9dab9a52562838a9b86b9d2b801d151934f58f6cc2
- QSR
- QSR-0.4.1
- QSR effective hash
- sha256:c306963db7e1ec109adc76dd655f234edacaf4c1a03d27a9ea13806699395fa7
- Tools
- TOOLS-0.4
- Tool-catalog hash
- sha256:4ff41b1fb43670530915c2b849d4f2e5c1a451873c70394069ce26fb0be4475b
- Model configuration
- google:generateContent sha256:bdadff507a7f715df9ad999d41eada133ffc0dce7c7060273dced69ab522e4bd
- Build
- 093d195bbb159084a629310d920875e4c513a6de
The public guide does not duplicate fully rendered dynamic prompts, because they can contain value-bearing enterprise evidence. The run record already preserves the governed request and evidence context separately.
Chapter takeaway
- Business time and request time are different clocks.
- A result set preserves membership, not values.
- A run is identified by its whole configuration, not by a model name.