Negative curation audit

Experimental negative curation, 2026-10-20

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

A first versioned collection of eleven partially supervised negative masks has been created and added to a new training snapshot. Four other candidates remain entirely ignored and unassigned. Original datasets, the pretrained baseline checkpoint and the frozen validation/test records were preserved.

The user explicitly authorized AI-reviewed labels for experimental training after two separate visual reviews. This collection claims AI review, not human approval. The aggregate audit records fingerprints, review-file hashes, split counts and the training smoke result. Real pixels, polygons, per-image decisions, masks and checkpoints remain outside Git. See the curation workflow for reproduction.

Annotation process

Two separate AI review passes inspected all fifteen original native 320 × 320 images, checking sample identity, image hashes and unassigned source status before pixel access. Neither pass used segmentation predictions or the other's annotations. Both followed the annotation guide. Separate passes do not establish statistical independence; correlated AI mistakes remain possible.

Each pass proposed conservative background polygons or deferred the crop. Reviewer A approved eleven crops and deferred four; reviewer B approved twelve and deferred three. A mask received background 0 only where both approved polygons overlapped, and only after both approved the sample with at least 256 intersection pixels. All other pixels remained 255 ignore. No PV 1 labels were generated. Disagreement deferred the whole crop.

Result	Count
Input crops	15
AI-reviewed partial negative masks	11
Deferred, all-ignore masks	4
Supervised background pixels	217,276
Ignore pixels	1,318,724
New supervised PV pixels	0

Only 14.15% of the combined crop pixels became background supervision. The remaining 85.85% were not silently interpreted as negatives. A partial negative does not assert absence of PV in ignored areas, on the entire roof or on the whole property. The final masks and overlays were also inspected on three native-resolution contact sheets covering all fifteen crops.

The collection pvextent-negative-curation-v1 retains every sample as unassigned. Approved masks have annotation_status = "reviewed" with explicit AI origin in new source entries and provenance; deferred masks remain unreviewed. The collection passed canonical validation with no errors and 24 warnings, including inherited unknown source information and deferred masks.

Original imagery licenses and notices are preserved. Newly created masks are identified as Valeford annotations under Apache-2.0, separately from source imagery and original upstream annotations. Redistribution must include the applicable license and notices alongside source attribution.

Training snapshot and holdout protection

The curated collection was merged into the frozen geographic PV4GER source as pvextent-training-v2:

Partition	Original	New snapshot
Train	3,282	3,293
Validation	278	278
Test	421	421
Unassigned	46	42
Total	4,027	4,034

The eight PV4GER candidates replaced only previously unassigned records; seven NRW records were added. Eleven reviewed negatives passed the training gate. All originally reserved geographic-block assignments remain reserved across successive augmentations. Duplicate held-out image content, spatial groups or building identities are rejected before candidate pixel validation.

The minimum new-training-to-holdout centre distance was 3,258.64 m, exceeding the fixed 1,000 m buffer. Distances use source centres in EPSG:25832, not verified building footprints. All 699 validation/test record dictionaries are exactly unchanged. Their raster references were checked with same-file inode checks; the augmentation did not open, decode or evaluate held-out pixel files.

Inherited image/mask files use hard links, so snapshots must remain immutable. Deleting a snapshot does not delete the other link, but editing linked bytes would affect both. Future corrections must create new files in another new snapshot. New masks are copied exclusively after hash checks, and target-path collisions are rejected before linking inherited files.

Output validation checks the complete metadata/manifest and all curated pixels. It does not newly decode inherited holdout rasters or certify their annotation truth. Parent/source fingerprints and full review provenance are retained with the snapshot.

Executed training check

A separate negative-only subset containing the eleven newly eligible records was prepared for the existing tiny training smoke workflow. Eight deterministically selected examples supplied four MPS optimizer updates at batch size two.

• Fixed training-probe loss: 1.802925 → 1.756776.
• Pre-clipping gradient norm: approximately 0.69896–0.72759.
• Parameter-update L2 norm: 0.313694.
• Checkpoint reload: exactly identical probe logits.

This verifies background/ignore supervision and the training/checkpoint path. It does not fine-tune the pretrained PV baseline, measure generalization or show a lower false-positive rate. Validation and test pixels were not used. The earlier baseline score remains attached to its original dataset snapshot.

Verification and next experiment

All 235 tests passed, including exact polygon intersection and ignore handling, authorization and review validation, source immutability, forbidden held-out pixel access, geographic exclusions, preservation of block reservations through two merges, and protection against inherited hard-link path collisions. The wheel and source distribution also built successfully; packaged curation, augmentation and CLI code match the source, with dataset pixels and checkpoints excluded from the packages.

The next model experiment should compare a fixed training protocol using this new snapshot against the existing baseline. Eleven partial negatives form a small, selected addition; their sampling weight and effective supervision must be recorded. Any validation comparison remains development evidence. No new model-quality or independent benchmark result is claimed by this curation step.