← course  ·  MODULE 9 · THEORY + KPI OPTIMIZATION SCENARIOS

Owning the Node

The capstone in book form: the cold audit's two products, the ranking logic that compresses the whole course, the ninety-day program in three waves, and the trial calendar that sequences ten features without collisions — closing with the program-level optimization scenarios.

No fake things. The audit below is the real one: the NYC tri-band node's kget (25,224 MOs) against the operator golden file (6,027 rows, 533 activations). The dormant inventory was verified attribute by attribute in Modules 2–7. For the full parameter-by-parameter tuning designs, see the companion Optimization Design chapter.

CH 1The audit's two products

The machine pass (Module 8's parse → join → except → rank) returns a headline known since Module 1: this network's drift is near zero — the golden residue is short. The findings that matter are the second category: staged capabilities. A dozen features sit licensed, configured, tuned, and sleeping. This network's optimization frontier is not parameters to fix; it is decisions to revisit — and recognizing a clean network is itself a professional verdict. Mature optimization is rarely repair: it is the disciplined conversion of dormant capability into measured improvement.

CH 2The dormant inventory

ModuleCapabilityStaged state (verified live)
M2VoLTE barring skip · DUAC chainparked by documented policy; golden agrees
M3Prescheduling 86 B / 5 ms / 200 ms · frequency hopping (restricted) · zero-interference · single codewordprofiles tuned, switches off
M4TTI bundling armed after HO/re-est, thresholds tuned — DEACTIVATED vs golden ACTIVATED · PDCCH boost operable at max 0the crown finding — a governance question
M6Voice-quality mobility detectors UL/DLDISABLED, thresholds staged 5/6 · 4/10
M7Rate recommendationbitRateRecommendationEnabled=false, license-gated

CH 3The ranking logic

Evidence first: a feature answering a measured pain (the KPI evidence attached per candidate) outranks one answering a theory. Blast radius and reversibility second: single-boolean, cell-scoped trials outrank architecture-touching ones. Interaction debt third: conflicting features sequence, never parallel. This is the course compressed into three sorting keys.

CH 4Three waves of thirty

Wave 1 — cheap, evidence-backed reversals: the TTI bundling governance question resolved first (oldest open finding: classify, history, then trial on edge-concentrated cells), and the PDCCH boost magazine loaded where traces show control-channel misses. Wave 2 — the uplink latency family: prescheduling per its trial criteria where SR volumes run high, against matched controls. Wave 3 — the quality detectors on complaint clusters whose drop autopsies show strong-signal failures, and the rate-recommendation question on edge-integrity evidence. Every wave ends adjudicated: accepted features graduate into the golden file through process; rejected ones are documented into institutional memory. The must-nots are deliverables too: restricted features get vendor conversations, corridors get capacity reviews, and staged-and-parked is a valid verdict — the program revisits decisions, it does not presume them wrong.

CH 5The trial calendar

The interaction matrix is the load-bearing artifact: the uplink family collides internally (serialize on shared populations); bundling × service-specific HARQ × robust LA share cells never; the mobility optimizer learns from statistics trials perturb (freeze it on trial relations); rate recommendation changes what loss means (conflicts with anything measured on integrity). Then the licensing pass (keys present, states writable, capacity counted — before any window opens), the standing guard-rail architecture (protected KPIs declared once; trial A's cells never in trial B's controls; populations allocated like spectrum), and rollback choreography written in peacetime (dependents unwind first, prerequisites last). The quarter: weeks 1–2 baselines + license pass + audit refresh; weeks 3–10 wave 1 parallel on disjoint cells, wave 2 serial; weeks 11–12 wave 3 entries + adjudication week. The measure of a quarter is decided questions: three accepts + four rejects + three extensions is excellent; ten half-observed enthusiasms is a wasted one.

CH 6Program-level scenarios

CASE 1

Two uplink trials, one population — the collision avoided

Situation
Prescheduling and an SR-period change both queued for the same latency-evidence cells.
Matrix verdict
Both move SR volume, grant latency, and PUSCH cost — the same three metrics. Parallel runs would be unattributable.
Action
Serialize: three weeks each, shared baseline, second trial inherits the first's accepted state.
Outcome
Clean attribution; the second trial's hypothesis sharpened by the first's data.
CASE 2

The license that would have wasted a baseline

Situation
Rate recommendation scheduled for wave 3; the licensing pass finds the FAJ 121 5014 key absent.
Action
Procurement starts in week 1, not week 11 — the pass exists to prevent a trial window opening on a feature that cannot legally start, with the baseline week already spent.
Outcome
The window opens on time, key installed and verified writable.
CASE 3

One degradation, three open windows

Situation
Mid-quarter, paging discards step on a guard-rail review with three trials live.
Action
The pre-commitment acts without debate: implicated windows pause; the cell-allocation map identifies which trial's population overlaps the degradation; the others resume.
Outcome
Root cause was external (an MME parameter change) — windows resume with clocks restarted, and the program's weekly one-pager records the pause honestly.

The closing word is ownership. The mark of it is not activity — it is the quality of your questions: evidence before action, surface before cause, decision before enthusiasm, and the network you leave behind cleaner, better documented, and more measured than you found it. Own the node.