What serviceability means
Serviceability is not "is there signal." It is a four-part claim, and the calculation must produce all four:
- A plan: which product tier (FWA 25 / 50 / 100 / 300, Business) can be delivered at this address.
- A device class: which CPE architecture — indoor gateway, window-assisted, fixed-aim outdoor, self-aiming outdoor.
- A confidence: how likely the delivered experience matches the quote, stated honestly.
- A capacity check: whether the serving sector has budget left to take the subscriber at that tier.
The verdict ladder — YES / YES with external CPE / SITE SURVEY / NO — is how the four parts compress into one answer a salesperson can use.
The calculation chain
address string → geocode (lat/lon + BBL/BIN) [NYC GeoSearch; parcel identity] → raster sample (outdoor RSRP, SINR) [prediction layers, 3×3 majority vote] → environment offsets (floor height) [optional refinement] → per-CPE effective signal [device link-budget ladder] → speed model (per CPE, per scenario) [SINR → spectral efficiency → Mbps] → plan eligibility [tier thresholds with 1.2× margin] → verdict + device recommendation [ladder rules + best-server margin] → capacity check [sector ledger] → confidence + report [uncertainty statement, serving sites]
2.1 Geocoding and parcel identity
Serviceability is quoted for a rooftop/parcel, not a lat/lon. Geocode to coordinates AND parcel identity (BBL/BIN via NYC GeoSearch), because the parcel gives building height, unit count, and ownership — all of which change the answer. If geocoding confidence is low or the point falls outside the prediction bounding box, the verdict is NO DATA, never a guess.
2.2 Sampling the prediction rasters
- 3×3 majority vote around the target pixel, not a single-pixel read. Ties resolve pessimistically.
- Palette snapping with a distance threshold: pixels farther than the threshold from any legend color are discarded from the vote.
- Band-floor semantics: a band means "at least the bottom of the band." All downstream math uses the band floor, never the midpoint.
Output is the outdoor street-level grade: Strong (RSRP ≥ −100 dBm), Serviceable (≥ −105), Marginal (≥ −110/−112), Edge/None below.
2.3 Environment offsets
Rasters predict outdoor signal at street level. Real subscribers are not. Height gain: conservative +1 dB per floor above the 2nd, capped at +6 dB, when floor is known. Rooftop qualification for MDU bulk: sample at the parcel and add the full outdoor-CPE placement gain; DOB roof height from building-footprint data supports the assumption.
2.4 The device ladder: from outdoor grade to effective signal
| CPE class | Offset vs outdoor prediction | Rationale |
|---|---|---|
| Indoor gateway, arbitrary placement | −15 dB | NYC masonry / low-E glass penetration at 2.5 GHz |
| Indoor, guided window placement | −8 to −10 dB | Penetration through one aperture, report tells the user which window |
| Indoor + window paddle antennas (FX3100 class) | −2 to −4 dB net | Aperture loss partly recovered by +6–9 dBi gain |
| Fixed-aim outdoor ODU (8–9 dBi) | +8 to +10 dB net | No penetration, antenna gain, professional aim |
| High-gain outdoor (15 dBi) / SOCA-aimed | +13 to +15 dB net | Deep-edge recovery; measured aiming |
Working rule: ~5 dB is one prediction band, so the ladder spans roughly six bands from worst indoor to best outdoor. Serviceability must always be computed per CPE class: "is this address serviceable" has no single answer, only "serviceable with what." HPUE: PC2 (+3 dB) devices effectively extend every threshold in §2.6 by ~half a band for tiers where uplink matters.
2.5 The speed model
Effective SINR converts to throughput with an attenuated-Shannon model, deliberately conservative:
SE = min( 0.75 · log2(1 + 10^(SINR_dB/10)) , SE_cap ) SE_cap = 7.0 bps/Hz LTE (256QAM) layers = f(MIMO config, SINR): 4×4 → 3.4 (SINR ≥ 20) / 2.6 (≥ 10) / 1.0 2×2 → 1.9 / 1.6 / 1.0 DL_Mbps = BW_MHz · dl_share · SE · layers · overhead overhead ≈ 0.90
Evaluated twice: as-built (current 2×2, 20 MHz) and upgraded (n41 with RSRP/SINR uplift). Quotes state a range, floor-to-typical, never a single number. Peak numbers do not appear on subscriber quotes.
2.6 Plan eligibility and the verdict ladder
eligible(tier) := DL_floor ≥ tier_rate · 1.2 AND UL_floor ≥ tier_ul_min
The 1.2× margin absorbs busy-hour scheduling and model error. Selling a 100 Mbps plan on a modeled 100.0 floor is how refund queues are built.
- YES — target tier eligible with indoor CPE.
- YES, EXTERNAL CPE — eligible only with outdoor CPE; quote includes the install.
- SITE SURVEY — Marginal grade, or best-server margin < 3 dB, or raster/parcel disagreement.
- NO — no CPE class reaches the lowest tier's floor, or address outside prediction bounds.
Best-server margin ≥ 4 dB → fixed aiming unambiguous. < 3 dB → triggers SOCA recommendation and lowers confidence one notch.
2.7 Capacity-aware serviceability
RF eligibility is necessary, not sufficient. Every verdict checks the serving sector's ledger — booked busy-hour demand vs deliverable capacity. If the target tier would breach the sector ceiling, the verdict degrades ("FWA 100 available; FWA 300 waitlisted on this sector") rather than silently overselling.
2.8 Confidence
Every verdict carries a confidence class from: prediction age vs network state, sample agreement (unanimous 3×3 vs split vote), best-server margin, geocode quality, and local calibration residuals. Report language maps confidence to commitment:
- HIGH — quotes a rate range.
- MEDIUM — quotes a range with "confirmed at install."
- LOW — SITE SURVEY by definition.
Calibration: the loop that keeps numbers honest
- Every install reports back — post-install RSRP/SINR from CPE management is written next to the predicted values.
- Residuals drive offsets — per-neighborhood mean(measured − predicted) becomes a local bias correction.
- Drive/walk tests seed the loop before install volume exists, prioritized to SINR seams and the Marginal fringe.
- Threshold review quarterly — grade boundaries, the 1.2× margin, and CPE offsets are tuning parameters.
- Re-baseline events — any re-tilt / power change / carrier add / n41 upgrade invalidates rasters in affected sectors; confidence drops until new predictions or measurements land.
Failure modes this methodology is designed against
| Failure mode | Guard |
|---|---|
| Optimistic band reading | Band-floor semantics + pessimistic tie-break |
| Single-pixel artifacts | 3×3 majority vote + palette distance threshold |
| "Serviceable" without saying with what | Per-CPE-class computation, verdict carries device |
| Quoting peak rates | Floor-to-typical ranges, 1.2× eligibility margin |
| Selling into full sectors | Capacity ledger check in the verdict |
| Seam addresses aimed wrong | Best-server margin → SOCA rule |
| Stale predictions after network changes | Confidence class + re-baseline flags |
| Indoor quotes in masonry | −15 dB default; window guidance is an explicit upgrade |
| Silent model drift | Install-telemetry residual loop, quarterly threshold review |
Summary formula (one paragraph)
Geocode the address to a parcel; majority-sample the outdoor RSRP/SINR rasters at band-floor semantics; apply the CPE-class link-budget ladder (and floor-height credit where known) to get effective signal per device class; run the attenuated-Shannon speed model per class under the network's TDD split and the class's carrier-aggregation ability; a tier is eligible when its floor rate clears the tier by 20% on a sector with ledger headroom; compress to the verdict ladder with the best-server margin deciding fixed-aim vs self-aiming outdoor; attach a confidence class from prediction age, vote agreement, and local calibration residuals; and feed every install's measured KPIs back into the residual model so the next quote is better than this one.