Mobility with a call in hand: the quality-triggered detectors, the measurement alphabet tuned for voice, the inter-frequency staircase, service-triggered policy, the SRVCC retreat, the self-optimization layer, and voice under EN-DC — with six KPI-optimization case studies.
No fake things. Every FAJ/CXC, MO attribute, threshold, and feature state below is from the Ericsson CPI (LTE RAN 25.Q4.4), the live NYC node kget (25.Q3), the operator golden file, and the LTE PM counter inventory. 3GPP citations name exact specifications.
CH 1Quality-triggered mobility — the staged detectors
Classic mobility triggers on signal (RSRP/RSRQ thresholds). But voice can die at strong signal — uplink interference, one direction failing while the other thrives (the glass-tower asymmetry). VoLTE-Triggered Mobility exists for exactly this: the uplink detector (FAJ 121 5433) and downlink detector (FAJ 121 5686) watch the voice bearer's own delivery quality and trigger mobility when quality — not signal — fails. On this node both sit DISABLED with thresholds staged: trigger ratios 5/6 (UL) and 4/10 (DL) tuned and waiting, the recurring posture. The reopening evidence is drop autopsies showing strong-signal failures (Module 8's traces).
CH 2The measurement alphabet, voice-tuned
The ReportConfig machinery (TS 36.331 §5.5.4) is the same alphabet for every service — A1 (serving good again), A2 (serving below threshold), A3 (neighbor better by offset), A5 (serving below X and neighbor above Y), B2 (serving below X and inter-RAT neighbor above Y) — but voice rides it with its own settings: tighter hysteresis and shorter time-to-trigger on QCI-1-bearing connections, because a voice call cannot ride out a deep fade waiting for a leisurely TTT, yet must not ping-pong (each handover is a small drop risk and an audio gap of tens of ms). The balance is the craft: hysteresis vs TTT vs A3 offset form a triangle — tighten any corner, the others move.
CH 3The layer cascade
A tri-band node is a staircase: capacity layer at the top, coverage layer as the refuge. The descent is built from A2 (leave-trigger arming inter-frequency measurement) → A5 (the conditioned step down) per layer pair, with each stair's thresholds forming a strict ordering — a UE must always have a next stair before the current one ends. The audit method (Lesson 6-3): walk the thresholds layer by layer and verify the chain has no gap; a missing stair shows up as drops clustered at one transition, often introduced by a refarm template (the acceptance-test case of Module 8 caught exactly this class).
CH 4Service-triggered mobility
Policy by bearer: a data-only UE and a voice-carrying UE should not make the same mobility decisions. Service-triggered mobility keys thresholds and targets to the presence of QCI-1 — voice avoids layers whose coverage is fragile, prefers the refuge band earlier, and is steered away from cells where voice KPIs run poor. The configuration surface is per-relation and per-profile; the audit point is coherence with the cascade of CH 3 — service overrides must respect the staircase ordering or they create service-specific gaps.
CH 5SRVCC — the organized retreat
Where LTE truly ends, the call continues only by becoming a circuit-switched call mid-flight. SRVCC (FAJ 121 2027 / CXC4011247, ACTIVATED toward UTRAN here) is triggered by B2 (LTE below X, UTRAN above Y), prepared by the MME/MSC ladder (TS 23.216), and executed as a handover that swaps the bearer's whole nature. Each phase has counters — preparation, execution, per-phase failures — and each phase failing means something different: preparation failures are core/neighbor configuration, execution failures are radio timing at the worst possible location. The B2 thresholds define where “the true LTE boundary” is; set them too optimistic and calls die before retreating, too pessimistic and healthy LTE calls leak to 3G.
CH 6The self-optimization layer
The node tunes its own mobility: per-relation optimization driven by observed handover outcomes, with voice given its own policy — hoOptQci1 = 5 / 50 / 200 on this node (the QCI-1 evaluation triplet: failure threshold, evaluation window, minimum samples). The automation's blind spot is the trial program's concern (Module 9): the optimizer learns from statistics that trials perturb — freeze it on trial relations or its corrections become confounds. Its everyday value is real: ping-pong damping and oscillation cleanup at a scale no human walks.
CH 7Voice under EN-DC
EN-DC bolts an NR leg onto the LTE anchor. For voice the question is which leg carries QCI-1's uplink — and this node answers conservatively: endcSplitAllowedMoVoice=false — voice does not ride the split bearer. The reasoning: split-bearer scheduling adds a coordination layer (PDCP reordering across legs) whose jitter voice pays for; the LTE anchor's known latency beats the split's theoretical capacity, which voice doesn't need. The headroom interaction (Lesson 6-7): NR leg power demand can starve the LTE uplink's power headroom — a population signature (EN-DC-attached UEs degrading alone) rather than a cell signature.
CH 8Six KPI-optimization case studies
CASE 1
Strong-signal drops — waking the staged detectors
KPI symptom
Drop cluster on two waterfront cells; RSRP at failure healthy in every autopsy.
Evidence
Traces: uplink SINR collapsing under marina radar interference while downlink thrives; classic one-direction death no signal trigger can see.
Root cause
Quality failure at strong signal — the exact prey of FAJ 121 5433.
Action
Module-8 trial: enable the UL detector on the two cells — thresholds already staged at 5/6, so the trial is a FeatureState flip with tuned machinery behind it.
Verification
Quality-triggered HOs appear in per-cause counters; cluster drop rate falls; no ping-pong inflation on the target relations.
Rollback
FeatureState; thresholds stay staged.
CASE 2
The riverfront flutter — relation-scoped hysteresis
KPI symptom
Evening drop spike, no cell pattern; Module 8's trace method found 31 of 40 drops sharing one shape.
Evidence
A3 ping-pong between two cells across a promenade, then RLF mid-handover; crowds flutter the balance only at evening densities.
Root cause
Hysteresis adequate for the empty street, inadequate for the crowded one.
Action
Relation-specific hysteresis increase (one relation, not the cell-wide ReportConfig — blast-radius discipline).
Verification
Evening drops fall by the cohort size; daytime handover latency unchanged.
Rollback
One relation attribute.
CASE 3
The missing stair after a refarm
KPI symptom
Drops cluster at one specific inter-frequency transition after a band refarm; everywhere else clean.
Evidence
Threshold walk (CH 3's method): the A5 pair for one layer transition missing from the refarm template — UEs descend onto a stair that isn't there.
Root cause
Template gap; archaeology traced it to a festival-era config (Module 8's audit case).
Action
Restore the A5 thresholds; register the template in the golden audit so the class is caught monthly.
Border-area complaints split two ways: dead calls at the LTE edge (no retreat) and “my call went 3G” deep inside LTE coverage (premature retreat) — different months, same parameter.
Evidence
Per-phase SRVCC counters: execution failures at edge in month one (B2 too pessimistic — triggered too late, by then the LTE leg was unworkable); successful but unnecessary retreats in month two after an overcorrection.
Root cause
B2 threshold pair oscillating between camps without a measured definition of “the true boundary”.
Action
Drive-data the boundary (Module 8's repeatable route), set B2 so preparation begins where LTE still has ~5 s of life, and freeze it with the rationale documented.
Verification
Both complaint classes fall; retreat rate matches the geography, not the calendar.
Rollback
Previous threshold pair, documented.
CASE 5
The optimizer as confound
KPI symptom
A hysteresis trial's control cells drift mid-trial; results uninterpretable.
Evidence
hoOpt machinery adjusted control-relation parameters in response to the trial's traffic shift — automation learning from perturbed statistics.
Root cause
Optimizer not frozen on trial relations (the Module 9 matrix rule, learned the hard way).
Action
Freeze hoOptQci1-driven adjustment on trial and control relations for the window; restart the trial clock.
Verification
Control baselines hold; the rule enters the program's standing architecture.
Rollback
Unfreeze post-trial.
CASE 6
The EN-DC population signature
KPI symptom
Voice MOS sags for one device cohort only; cells, bands, hours all flat.
Evidence
Population split (the Module 8 playbook habit): EN-DC-attached UEs alone degrade; PHR traces show NR leg consuming uplink power headroom during calls.
Root cause
NR leg starving the LTE voice uplink — configuration interaction, not illness.
Action
Verify endcSplitAllowedMoVoice=false intact (it is, ×9); tune EN-DC power-sharing/SCG config for voice-active UEs rather than touching the voice bearer.
Verification
Cohort MOS converges to population; non-voice EN-DC throughput cost measured.
Rollback
SCG config attribute set.
References
Ericsson CPI, LTE RAN 25.Q4.4 — FAJ 121 5433 / 5686 (VoLTE-triggered mobility UL/DL), 2027 (SRVCC)