Radio-coverage worktrees — implementation and recorded results
Versioned worktree evidence; map validation and transfer remain unresolved.
Inspection and evidence boundary
Read-only code/result inspection by a bounded Sol researcher, followed by coordinator synthesis. No training, inference or source tests were executed. Four checkout HEADs were rechecked after manager authorization and match the pre-gate captures. Worktree originals, Git status and hashes are in the shared archive.1
| Worktree | HEAD | Recorded state |
|---|---|---|
| serene-agnesi-18d752 | 528dadf77d81dc34f7cc3e01589c9b509385fb2b |
Learned-ray code/results tracked; three training logs untracked |
| ray-beam-energy-sim-f84c77 | 0490fb3bd1c002ef0d6e9d6c535711aa240330e2 |
Raster proposal untracked; proposed beam module absent in inspected tree |
| vigorous-burnell-99a393 | 56d7b5ff3ff6a81a8f9e0ae4493d9bb47c334584 |
SpectrumNet code/results tracked; working tree clean |
Learned rays
The teacher recorder, path deposition and step/patch models exist. Saved stage-one results report 79.6% exact rollout at 1 m on 12,672 held-out rays with at most two reflections, falling to 63.4% at four reflections. Exact rollout requires interaction types and all endpoints to agree; it is not coverage-map accuracy. One seed and shared-GPU timings limit the comparison.23
Batching and shared occupancy already exist. Fixed fans, deposition of predicted fans into maps, comparative throughput and diffraction remain proposed validation/work. Tests checking teacher recorder/deposition against the tracer exist, but were not run here; agreement with a teacher is not independent field accuracy.456
Additional code limit: raynet.py fixes RES=0.5, NOCC=512 and NB=736 with 0.5 m forward bins, covering the 256 m tile diagonal. The report's architectural suggestion that map size does not matter is not a demonstrated km-scale implementation. Larger extents require auditing strip length, coordinate normalization and exit behavior before claiming transfer.4
Raster beam proposal
The document proposes H/R/S images, GPU raster marching, scattering, diffraction and optional learned materials. Its target of roughly one-second maps and a half-day speed test are proposal estimates, not achieved results. The inspected worktree has an untracked plan and no proposed rcm_ml/beam implementation. It should remain an optional future branch, not a dependency of finishing the existing tool.71
SpectrumNet
The adapter, gated training and saved metrics exist. The local report infers pixel-to-dB encoding and transmitter positions, and identifies multi-transmitter maps and coarse 10 m pixels. Treat encoding, inferred height, missing weak-signal tails and diffraction interpretations as internal data-analysis findings; the underlying dataset/paper/license was not independently re-audited in this assessment.89
Arithmetic from preserved result JSON supports the following, without reproducing predictions:11
| Check | Recorded/derived result | Implication |
|---|---|---|
| Gated distance-input seeds, outdoor RMSE | 9.852 / 9.859 / 9.884 dB | Similar in-domain aggregate scores |
| RMSE where reference is covered | 13.207 / 13.916 / 13.701 dB | Large error on positive regions remains |
| Maximum aggregate ring coverage error | 3.9 / 2.7 / 3.7 percentage points | “Within four points” holds for aggregate rings; seed-0 single-Tx subgroup reaches 6.7 pp |
| Mosaic 100–200 m coverage | 51.4 / 13.0 / 26.3%; 38.4 pp spread | In-domain similarity conceals transfer instability |
| 400–800 m agreement | Reported 99.6–99.8%; all-dead reference baseline 99.8% | Accuracy is dominated by negative cells; does not demonstrate useful positive coverage prediction |
| Best validation steps | 12,000 / 21,000 / 12,000 of 32,064 | Longer training is not justified by these logs alone; this is not proof additional training can never help |
The mosaic JSON contains tracer coverage 53.6% at 100–200 m while the prose table says 53.7%. Seeds one/two record 38.1 minutes versus prose 32–34. Report these as small source-document inconsistencies; use JSON for arithmetic. The 3.5→5.9 GHz free-space correction is 4.535679 dB, not a validated cross-frequency propagation model.10811
Next actions suggested in the session (extra frequencies, ensembling, finer-grid work) remain proposals. None is required to finish this assessment, and none was executed here.
-
Source record plus capture-inventory.json preserve absolute paths, HEAD, dirty context and individual SHA-256 values. ↩↩
-
Exact learned-ray research note, sections 1.3–1.5. ↩
-
Exact saved step+patch test JSON; no prediction rerun. ↩
-
Exact raynet.py constants and batching/rollout code, inspected statically. ↩↩
-
Exact teacher trajectory/deposition implementation. ↩
-
Existing test definitions; not test execution evidence. ↩
-
Exact untracked raster plan; targets and phases are proposals. ↩
-
Exact SpectrumNet research note; source claims/interpretations retained with attribution. ↩↩
-
Exact SpectrumNet adapter/evaluation code. ↩
-
Exact mosaic result JSON and copied reference-method metrics. ↩
-
Derived saved-JSON arithmetic inventory; originals are enumerated in source.md. ↩↩