nx_room_qoe.nx

nx_room_qoe.nx -- X-ROOM (client-side, S-class): sovereign DETERMINISTIC INTEGER call-quality (QoE/MOS) estimator -- the CLIENT-side sense that feeds nx_room_diag. RESEARCH-GROUNDED (ITU-T G.107 E-model + WebRTC QoE literature, via WebSearch): the published standard is the E-model R-factor R = 94 - Ie_eff - Id -> MOS, but it is FLOATING-POINT and voice-only; modern WebRTC QoE (rtcscore, XGBoost/MLP MOS) are ML BLACK BOXES -- non-reproducible (published RMSE ~0.28 MOS), need training data. S-CLASS EXCEED (measured, honest): (1) DETERMINISTIC INTEGER E-model -- same stats give the SAME MOS to the milli on any hardware (audit-replayable), where ML scores are non-reproducible and the float E-model varies; (2) MULTI-DIMENSIONAL ATTRIBUTION -- not one opaque number but per-impairment breakdown (loss/delay/jitter/bitrate) -> the DOMINANT killer -> the sovereign FIX (loss->fec, jitter->jitterbuf, bitrate->sfu/ simulcast); (3) catches COMBINED SUB-THRESHOLD degradation a naive per-threshold quality bar reports as "good" (impairments ADD in the E-model). No training, no float. Inputs (client-observed): loss_pm (per-mille), rtt_ms, jitter_ms, delivered_kbps, target_kbps. Output: MOS in milli (1000..4500), R-factor, dominant impairment + fix. main() is the SELF-VALIDATING GATE. Evidence -> knowledge/status/room_qoe.log. HONEST SCOPE: the deterministic QoE MODEL; the JS last-mile shim must collect the raw counts (frame seq-gaps=loss, inter-arrival=jitter) and report them = the wiring rung. license_tier: ORIGINAL