Training extension audit

Six-epoch continuation audit

Canonical source: versioned documentation. This wiki copy is maintained manually.

Local date: 2026-10-21, Europe/Berlin. UTC logs use 2026-10-20. The aggregate JSON records measured metrics, checkpoint identities, input hashes and the separate training-review pools. The baseline guide gives the exact continuation command.

Training result

The augmented candidate was continued from epoch three through epoch six, adding three epochs and 4,941 optimizer updates. The run uses the unchanged 3,293-image training split and 278-image internal validation split. Its final global step is 9,882. It adds no newly collected NRW examples or labels.

The best checkpoint is epoch 5, with internal pixel IoU 0.92183825 and Dice 0.95932969. IoU improves by 0.280326 percentage points over the parent best. The fixed last epoch has lower IoU (0.91783037), so it does not replace the best. The gain is small and does not demonstrate better external transfer.

Internal validation measure	Parent best, epoch 3	Continued best, epoch 5	Fixed final epoch 6
Pixel IoU	0.91903499	0.92183825	0.91783037
Dice	0.95780952	0.95932969	0.95715490
Precision	0.96084849	0.96085002	0.95256894
Recall	0.95478971	0.95781417	0.96178524
TP pixels	5,823,842	5,842,290	5,866,512
FP pixels	237,303	238,045	292,110
FN pixels	275,765	257,317	233,095
TN pixels	22,130,290	22,129,548	22,075,483

The best checkpoint is selected solely by internal validation pixel IoU. The fixed final epoch is reported even when it is not the best. These are global masked pixel metrics against upstream labels, not array/roof scores or an independent external benchmark. The earlier positive pilot motivated concern about transfer; it was not rerun to choose checkpoints, thresholds or settings for this control. No new external performance claim follows from this run.

Additional epoch	Validation pixel IoU	Dice	FP pixels	FN pixels
4	0.91074304	0.95328678	223,900	340,517
5	0.92183825	0.95932969	238,045	257,317
6	0.91783037	0.95715490	292,110	233,095

Registered configuration and immutable inputs

The duration-control protocol and six-epoch TOML were committed as 6a32373 before training. Only epochs changes, from three to six. AdamW, learning rate 0.0001, weight decay 0.0001, batch size two, seed 20261020, horizontal/vertical flips, native RGB normalization and threshold 0.5 remain unchanged. Regular resume enforces the original dataset/split fingerprints, encoder, preprocessing, device, package versions and optimizer/RNG state.

• Dataset fingerprint: 1e78b1ce00d826e22d60c50f4c2f7c1b337bc3eeb9f36303b51e53daf0e54f14.
• Parent last.pt SHA-256: cdb958053609ee1b3cb4781f59518da9eb921ab28c5f8d2d379a6fcbd728e720.
• New best checkpoint SHA-256: 2e04c0e009f3b76a1a0c233aa9efea064251f6aed572d8db9272c2455c83edc6.
• New last checkpoint SHA-256: d7ed0d58e3230c55edac19aaca23d593bf16e1b8695f399a948fd9a70a298747.

The continuation is published under runs/baseline-resnet18-continued-001. The original parent and reference checkpoints remain unchanged. Start-validation confusion counts exactly reproduce the parent's final counts. Reloading the new best checkpoint also exactly reproduces its recorded validation counts. Model parameters changed during optimization; the run records a finite, nonzero parameter-update norm. All 26 retained source/model artifacts and all 699 original validation/test record dictionaries remain unchanged. The original 421 test rasters and the external twenty-case pilot rasters were not opened for this work.

Training ran on Apple M2 / MPS with the same package versions as its parent. Recorded wall time is 1376.68 seconds (about 22.94 minutes). Acquisition and CPU tests ran concurrently, so elapsed time is not a dedicated throughput benchmark. Training implementation bytes match the registered commit; the acquisition refactor does not change model or trainer code.

Separate new NRW training-review inputs

A new training acquisition API was implemented and registered as d769751. The selection plan and initial acquisition protocol precede RGB inspection. ID, image-hash, building and spatial-group exclusions cover all 4,054 records in the full training snapshot and twenty-case pilot. Geographic distance protects 761 val/test/unassigned centres, including every pilot case. New training imagery may be near existing train-only examples; this does not relax evaluation acquisition's all-split rule.

Two combined-pool attempts encountered a WMS timeout and connection reset; neither published a partial dataset. The region-batching record, committed as 939a3c9, retains the same regions, 24 slots, distance rules and quotas. Each region then publishes independently. There is no replacement sampling or prediction-based choice. Actual successful acquisitions are:

Region	Native crops	Capture date	Minimum actual affine-centre distance to protected reference
Schleiden	8	2025-03-07	11,399.725 m
Dülmen	8	2026-04-07	10,296.754 m
Marsberg area	6	2025-06-30	5,635.661 m

The 22 native RGB crops are all 320 × 320 at verified 0.1 m/pixel, retain exact source response bytes, and have unique IDs, buildings and image hashes. An independent metadata audit checked full-crop GFI support, reference hashes and actual affine crop centres. Global candidate spacing is at least 207.294 m. Two Marsberg slots cross a returned metadata-tile boundary; the tile reports 0.10 m but does not cover the full crop, so full-crop native resolution/date are unconfirmed and those slots are excluded. This is not proof that their source imagery has 20 cm resolution.

Every canonical mask remains entirely 255, status unreviewed, split unassigned, and training_use_allowed: false. There are zero new canonical supervised pixels, and no new crop was added to the continuation. The source images retain DL-DE-Zero-2.0 and their acquisition notices. Region fingerprints, exact request counts and minimum distances are in the aggregate JSON; building responses are bounded feature prefixes, not complete regional surveys.

Two new independent blind AI workflows each inspected all 22 originals and approved the same seven positive crops, deferring fifteen. Their exact review hashes are recorded in the aggregate JSON. Proposal schema, simple polygon geometry, image hashes and within-review PV/background non-overlap passed validation; native proposal overlays were inspected. The intersections supply 8,938 proposed PV pixels and 20,899 proposed background pixels across seven cases at the preliminary 100-PV/256-background gate. These are partial experimental AI proposals, not human-confirmed truth or imported canonical labels. The canonical datasets remain unchanged and wholly ignored. No model prediction was used for these proposals.

Validation and next work

The complete suite passed 334 tests, including unchanged negative-acquisition contracts, near-train eligibility, metadata-only protection of all twenty pilot cases, strict training split rules, native metadata gates and atomic publication. Wheel and source-distribution builds passed. Changed module bytes match the working source, package metadata contains the final README, and real imagery, masks, reviews, runs and checkpoints are excluded. Real imagery, labels, review decisions, checkpoints and raw run outputs remain outside Git. Original model/data artifacts are retained.

Longer optimization is a development control. It does not establish a remedy for the observed external failure. The new raw images and any review proposals still require positive-mask curation, an immutable training snapshot and a separate registered NRW-enriched experiment. A handful of positive examples can be overwhelmed by 3,293 uniformly sampled records; record sample and supervised pixel shares and design sampling explicitly. Broader independent positive evaluation is needed before relying on internal validation gains.