The loop closes on wait, not on Q

by Fudoshin Research·2026-08-20·S3·F-2026-08-20-I

A card is not a loop. The working loop measures sojourn, acts, and measures again. Official Q stays on the tick as audit. At a hot window the wait-loop sheds; the official hint cuts.

Problem under load

The operator card recommends a lever and stops. Nothing applies it. Nothing measures the next window. Official Q is the wrong plant: it is a published invariant, not a sojourn. Scoring dQ keeps saying cut the bottleneck, because Q grows as nu*^2. The job feels wait.

Primitive / definition

Plant state is (M, Pi, tau, u). Observation is end-to-end sojourn from the Lindley tandem (same Stage D simulator; no sojourn formula inside). Wait controller: if Pi >= 1 shed; if u < 0.25 repair; if measured wait exceeds the mean SLA, shed; else hold. Q is printed on the tick. It is not an input.

Method / protocol

Hot window Pi = 0.60 against a 400 ms mean SLA. Three controllers from the same start: open (observe only), wait (this loop), Q (official hint). Replay accepts a live measured_wait_s as the exogenous residual. 24 checks. Bind table, not a product integration: shed = drop low-priority retrievals or tool calls; cut = cache / faster retrieval; repair = stop writes.

Result

Open loop stays over 400 ms. Wait-loop first action is shed; a short window later measures 402 ms against a 390 ms formula and sheds again — the controller believes the measurement. Q moves 50.274 to 50.221. The Q-loop cuts twice; after the first cut the formula is already under the wall and the measurement is not; Q jumps to 76.5. Wait is not why it moved. After the mean is repaired, miss rate is still 0.328. The tail is a different loop.

Limitations and next