Radio-coverage — proposed completion milestones
Acceptance gates for the existing tool, followed by separate DroneRadar validation.
Goal and decision status
Proposed end state: reproducible propagation predictions for an explicitly supported domain, with trustworthy evaluation and usable outputs. These milestones are recommendations; only the present assessment is authorized. No dates, labor estimates or training budget are invented. Main/worktree evidence explains the gaps each gate addresses.12
Complete the existing radio-coverage tool
| Gate | Work and dependency | Proposed acceptance criteria |
|---|---|---|
| P0 — Definitions and artifact inventory | Define path gain/loss sign, floor, frequency, antenna/power convention, CRS, heights, outdoor mask, extent and latency goal. Inventory branch/code/config/result/checkpoint/data provenance. | Every selected baseline/run resolves to code commit, configuration, split, data version, units and available inputs. Missing/untracked assets labeled; no recipe called a completed run. Immutable evaluation fixtures chosen before model selection. |
| P1 — Reference credibility | Depends P0. Run the separately approved reference-sensitivity experiment; distinguish IRT2/IRT4/DPM, tracer settings and grid artifacts. | Scene-held-out metrics and confidence intervals reported for multiple thresholds/rings, including positive recall/precision and covered-region RMSE. Calibration influence disclosed; no field-truth label for simulator references. Any threshold transfer claim within an agreed tolerance or explicitly rejected. |
| P2 — Baseline and learned-method selection | Depends P1. Compare available baseline, geometry, hybrid and optional gate methods on the same cases. | Report accuracy, selection regret, seed variation and wall-clock/memory including preprocessing. Candidate must meet a pre-agreed error/utility target and improve over the simplest baseline on its stated task; otherwise retain baseline. Proposed research target: one-site mean regret ≤1 pp with scene-bootstrap upper bound ≤2 pp on supported simulator task, to be negotiated before execution. |
| P3 — Domain and extrapolation | Depends P2. Test continuous real extents, held-out cities, intended frequencies/heights, and scene/edge conditions. | Publish a support matrix of tested/unsupported settings. Larger/altitude scenarios have their own held-out evidence. No 3.5→5.9 GHz free-space shift, 256 m tile score or stitched mosaic used alone to certify new domains. Unsupported requests return warnings/nulls. |
| P4 — Reproducible tool and application integration | Depends P2/P3 for chosen domain. Package deterministic inference/export and link application through a defined adapter. | Fresh documented environment regenerates accepted fixture predictions within declared numerical tolerance; model/config/data hashes and latency/memory captured; unit/coordinate/floor/NoData handling checked. Endpoint result agrees with standalone prediction on fixtures. Fallback and failures distinguish baseline/learned methods. |
P2's numeric regret targets are proposed engineering acceptance thresholds, not observed universal performance or user-approved requirements. P0 must settle the actual target before P2 execution. Simulator qualification completes a simulator-qualified tool only; measured propagation claims require external measurements.
Optional research branches
Learned-ray work requires fixed-fan teacher comparison, map deposition at 100/7,200 rays, endpoint plus map metrics and CPU/GPU timing with identical scene/power settings. Audit hard-coded 256 m raster/strip constants before scaling, then add diffraction if required. A path exact-match score alone cannot pass P2/P3.2
The raster proposal requires its own feasibility gate (empty-space/single-wall energy/falloff, resolution sensitivity, cost) before any integration promise. SpectrumNet requires held-out subgroup and domain-transfer checks; coarse source data cannot gain real geometric detail from interpolation alone. Neither branch is prerequisite for P4 if the current tool meets the target.2
Later DroneRadar application gates
| Gate | Dependency and evidence | Acceptance |
|---|---|---|
| D0 — Observation definition | Partner signal classes, export sample, antenna/receiver information, data rights and task definition | Contract maps actual fields/units/time semantics; unknowns explicit. Define whether success is emitted-signal reception, valid packet, classified alert or localized track within a stated time window. |
| D1 — Controlled calibration | D0 plus independently known target-presence/exposure and receiver uptime | Record misses and negatives as well as detections; evaluate by site/flight/day held out; disclose sensitivity to height, interference, dwell and placement. No numeric probability without calibrated evidence. |
| D2 — Placement validation | D1 plus proposed-versus-baseline site comparison and field feasibility | Compare coverage/recall gain, false alarms and uptime under a fixed site budget; uncertainty and correlated receiver failures included. Set pilot targets before collecting evaluation outcomes. |
| D3 — Operational integration | Separate approved implementation/security/data-rights scope | Validated adapter, monitoring, retention/access controls and operational acceptance. Assessment completion does not authorize this phase. |
Next action
Authorize the bounded P1 experiment as a new phase if the proposed propagation-first path is accepted. Resolve available assets and encoding first; expand domain only after the reference check. Partner/interface discovery remains coordinated with IDEA-001.