AI signal coverage — readiness and recommended completion path
A useful research asset; propagation validation and application integration remain incomplete.
Assessment
Proceed with a baseline-first propagation tool and an explicit evaluation contract. Existing code and simulator experiments justify preserving and completing this asset. They do not yet justify presenting its output as measured passive detector coverage. Prioritize reference calibration and domain validation before another training campaign or tracer architecture. This is an agent recommendation within the approved assessment, not a user implementation decision.12
The bounded assessment answers scope v002. Recommendations about the overall DroneRadar boundary remain for manager reconciliation; shared project scope has not been edited.
Readiness
| Capability | Evidence state | Readiness conclusion |
|---|---|---|
| Baseline application | Terrain/building propagation, link-margin service and FastAPI path implemented | Starting platform; runtime and endpoint behavior not exercised in this assessment |
| Learned map/site ranking | Reported low DPM error and small site-selection regret on small simulated street-level tiles | Promising narrow benchmark; model loading absent from inspected application path |
| Long-range prediction | Calibration-dependent artificial-mosaic convergence; IRT4 disagrees with IRT2 nearby | Needs reference sensitivity and real-extent validation; far-range physical accuracy unestablished |
| Learned ray paths | Stage-one code/results; ~80% exact paths at 1 m for ≤2 reflections | Research option; map fidelity, diffraction and comparative speed incomplete |
| Raster beam method | Untracked proposal and performance targets | No implementation established |
| SpectrumNet transfer | Gated models reduce spurious far coverage; 38.4 pp near-range seed spread | Diagnostic evidence, insufficient stable transfer; do not promote to production coverage mask |
| DroneRadar detector coverage | No calibrated detector/export evidence in this assessment | Concept/contract stage; requires signal/receiver/field-validation layer |
Readiness derives from bounded code/result inspection, not model execution. Main PoC JSONs include untracked outputs; all preserved evidence records identify checkout/commit context.12
What is already valuable
The project has three reusable pieces: a geometric baseline application, learned-map evaluation/placement machinery, and documented failure investigations. Recorded floor-clipped all-pixel DPM RMSE improves from geometry's 2.353 to 1.994 dB in one 128 px run; recorded one-site hybrid regret is 0.14 percentage points on 99×80 simulator cases. Raw all-pixel RMSE for that pair is 10.0865 versus 9.9993 dB; the much smaller clipped errors should not be interpreted as uncensored channel-gain accuracy. These support a controlled research continuation, not performance promises across sites or sensors.1
The later convergence work is especially useful because it challenges the earlier short-range interpretation: IRT4 differs from IRT2 by roughly 16–19 percentage points across selected 100–260 m rings. A stable numerical calculation can still be calibrated to the wrong reference. Likewise, a 99.8% far-field agreement can be matched by predicting every pixel dead in the sparse-positive reference. Acceptance must measure positive-region utility and reference sensitivity.12
Recommended path and alternatives
- Freeze definitions and baseline fixtures, reconcile run provenance, then audit reference sensitivity. Prefer path gain/loss as the core output, with optional received-power/link-margin calculations only when antenna, power and receiver assumptions are supplied.
- Evaluate existing geometry/physics/hybrid methods on scene-held-out cases, multiple thresholds and positive-region metrics. Select the simplest useful method. A learned model is optional if it fails to improve the intended task.
- Establish a supported geometry/height/frequency domain and reproducible inference package. Only then propose application integration.
- Treat DroneRadar receiver placement as a later application gate, using receiver-specific evidence and field observations; keep uncertainty visible.
Alternatives: continue learned-ray fan/deposit validation if faster map generation is demonstrated to be the bottleneck; use SpectrumNet as an explicitly different domain diagnostic; build a raster tracer only if existing methods fail an agreed speed/fidelity requirement. No current evidence requires pursuing all branches. Full-resolution/multi-seed completion of the historical ML protocol remains an artifact gap to resolve before authorizing costly reproduction.123
Propagation versus detector coverage
Analysis: a propagation map supplies a channel prediction under assumed geometry/materials/frequency/heights. A passive detector also requires a compatible emitted signal, receiving hardware/antenna, scanning/dwell and decoding or classification, interference conditions, operational uptime and a defined observation window. Different signal classes require different evidence; Remote ID observation, RF-link classification and active radar are different measurements in the existing landscape.4
Do not convert a controller-to-drone link margin directly to detector probability. The receiver location and performance, antenna patterns, transmitter duty cycle and target output differ. The existing DJI aggregate note reports an imbalanced controller-link dataset and assumed controller position; it does not calibrate third-party detector performance. No raw flight logs were accessed.1
See proposed interface for explicit units/null behavior and completion plan for gates. The tool's first defensible DroneRadar role is a candidate measurement/placement prior to test, not a certification that unobserved airspace is covered.
Evidence gaps and decisions
- Completion target: path-gain prediction plus benchmark is recommended; required geography, frequency set, height layers, extent, latency and accuracy tolerance remain user/product choices.
- Reference truth: simulator disagreement, coarse grids, clipped floors and inferred SpectrumNet encoding; no measured propagation validation or detector probability calibration established.
- Reproducibility: checkpoint availability and data rights/encoding need a run inventory; small selected documents/results are archived, weights/large arrays omitted.
- Partner interface: BlueBird and Chuika are user-identified co-founders, but legal/product mapping, sensor outputs, sensitivity, confidence semantics and data rights remain unresolved. Reuse IDEA-001's ongoing partner research; no capability is inferred from co-founder status.5
- Field design: negative exposure, true target presence, receiver uptime and interference need observation independent of detection events. Missing detection records are not automatically false negatives.
These unknowns do not prevent planning. They do prevent performance claims or a detector-integration commitment. First experiment proposal addresses reference choice before new training.
Checks, capture limits and review
The bounded researchers inspected selected code and saved JSON; coordinator rechecked unchanged HEADs, examined learned-ray fixed-grid constants and synthesized evidence. No source tests, runtime, inference or training were run. Hash integrity/knowledge validation are separate from scientific performance. Four Claude conversations were selectively sampled; missing messages/images/tools are disclosed in the inventory.
Pre-gate online discovery found ITU P.1411 and an ASSURE report candidate. Archive attempts failed (network resolution / reader size); they are listed in the manifest but are not used to support technical conclusions here. No fresh downloads were attempted under the start exclusions. Dataset licensing and external paper contents remain unaudited. The independent Sol review checked representative exact captures and recorded two medium findings: clipped/all-pixel RMSE labeling and rounded-floor threshold contamination. Both were corrected and rechecked, with no remaining acceptance-blocking defect in that bounded review. Reviewer checked 66 retained-evidence integrity entries; coordinator checked all 67 after adding the source-byte recheck record. Knowledge validation passed after finalization. No human or field-performance verification is claimed.6
-
Main findings link exact code, run JSON and study-note captures, including DPM/site ranking, convergence and aggregate flight-note limitations. ↩↩↩↩↩↩
-
Worktree findings link exact ray code/results, SpectrumNet JSON and arithmetic checks, and the raster proposal. ↩↩↩↩
-
Curated session inventory identifies pending proposals and original intent with selective-transcript limitations. ↩
-
Existing landscape synthesis and its canonical/technical source chain; reused read-only, not repeated. ↩
-
Exact user clarification; does not establish legal/product/capability mapping. ↩
-
Independent reviewer record with scope, corrected findings, recheck and limitations. ↩