Figure: the KV-state control plane drawn as a closed cycle. Model and agent state, carrying tool calls and deadlines, feeds a dashed band labelled KV-STATE CONTROL PLANE containing three stages in series: SELECT, PLACE and MOVE. The chain continues to ATTEND and then OBSERVE, and a dashed feedback arc runs from OBSERVE back to model and agent state, closing the loop. SELECT asks which candidate regions or blocks are worth considering at all; status design, unvalidated. PLACE asks which physical tier a block should occupy and whether its policy state survives the crossing; this is the only stage with evidence, status simulated and provisional: persistent identity-keyed policy state is worth plus 34.15 points to LFU at a 32 gibibyte HBM budget, 15.56 per cent rising to 49.71 per cent, and it is a protocol result testable with no CXL hardware. MOVE asks when and how the bytes cross and whether the cost can be hidden behind compute; status design, unvalidated. ATTEND is the model reading whatever is resident; status design, unvalidated. OBSERVE records what was actually used and closes the loop; status design, unvalidated, conceptual only. The standing caveat applies: a single data-structure defect in Revision 1 of the simulator inverted four of five findings, so Revision 2 is provisional, not settled.
Figure 3.2 · Illustrative — proposed architecture · one stage Simulated (provisional)
The KV-state control plane: SELECT → PLACE → MOVE → ATTEND → OBSERVE
The v6.0 headline structure. Five stages and a feedback loop, each annotated with the question it answers and the evidence that stands behind it. Exactly one stage — PLACE — holds a mechanism with any empirical result, and that result is simulated, not measured.
The cycle
What each stage decides, and what stands behind it
- SELECTDesign — unvalidated
Question it answers: Which candidate regions or blocks are worth considering at all?
Relevance and reuse-prediction signals narrow the information hierarchy before any tier decision is made. Past-only signals (recency, frequency, EMA/LRFU) and predictive signals (agent/tool-call state, tool-return timing, prefix identity, reuse history, attention signals) both belong here; only past-only signals have been evaluated by this package.
- PLACESimulated (provisional)
Question it answers: Which physical tier should it occupy — and does its policy state survive the crossing?
Holds the one mechanism in this architecture with an empirical result behind it: persistent identity-keyed policy state. Letting policy state survive HBM eviction, rather than re-admitting the block as if newly created, is worth +34.15 pts to LFU at a 32 GiB HBM budget (15.56% → 49.71%). A protocol result — testable with no CXL hardware.
- MOVEDesign — unvalidated
Question it answers: When and how do the bytes cross, and can the cost be hidden behind compute?
Transfer scheduling, batching, chunk sizing and prefetch depth. No CXL device has been measured by this package; the only movement numbers anywhere in it are analytical.
- ATTENDDesign — unvalidated
Question it answers: What does the model actually read this step?
The model consumes the placed and selected KV state. This stage does not query the control plane; it acts on the control plane’s prior decisions.
- OBSERVEDesign — unvalidated
Question it answers: What was actually used, and what was reused?
Records the outcome and closes the loop, updating the signals that feed the next SELECT. Conceptual only — nothing here is implemented or measured.
One stage of five carries evidence, and only one sub-question inside it.Architecture is a design artifact: measurement validates performance claims, not whether something qualifies as an architecture. The +34.15-point persistence result inside PLACE is simulated and provisional. It is not evidence that SELECT works, that MOVE hides transfers, or that the five stages compose into a working system — none of which has been built.
Illustrative (proposed architecture) · PLACE result Simulated (provisional) — ARCHITECTURE-v6.md §2 and §4; CANONICAL-NUMBERS.md §3.3