Feed glint-validated candidates back as training labels (active learning)

OPEN, never implemented (as of 2026-08-11)

This remains a proposal. The hard-negative half of the idea shipped by a different route: hard_negatives.py mines confirmed negatives bi-temporally rather than by glint. The glint route is blocked in practice by the unresolved scene-coverage bug in Tile-batched glint coverage.

Labels: enhancement, model, glint, active-learning

Problem

The model's training positives come from OSM mapping, which lags reality badly in the target regions (that gap is the project's reason to exist). The human validation loop (MapRoulette) closes it slowly and at human cost. Meanwhile the glint check produces a stream of physically corroborated candidates -- ≥2 mutually-consistent spike dates recovering a single panel orientation -- whose false-validation floor is measured: 8.7% on 69 no-PV control buildings (results/glint_validation_pakistan/REPORT.md + the corroboration experiment's controls). That is high-precision evidence sitting unused as supervision.

Proposal

Treat glint-validated candidates as auto-confirmed positives for the next training round, without waiting for human validation:

1. After a postprocess --check-glint run, select candidates with glint_consistent >= 2 and a roof-plausible fitted orientation (glint_fit_tilt/glint_fit_az -- e.g. tilt < ~35° for rooftop placement, any tilt for ground candidates).
2. Restrict to area buckets where the calibrated likelihood ratio is meaningful (≥ 500 m², matching postprocess._GLINT_BUCKET_EDGES_M2 / min_lr -- below that the 8.7% floor eats the signal).
3. Burn these polygons into the chip set as positives via the existing chips machinery (same path as OSM labels), tagged with a source=glint column so they can be weighted, audited, or ablated separately.
4. Retrain; compare against the OSM-only baseline on the held-out val split.

The inverse direction is also open, through the existing hard-negatives review path (never auto-demotion, matching the reward-only contract): candidates showing persistent bright spikes with geometrically inconsistent dates over many scenes look like specular noise sources (water, greenhouses, metal roofs) and are good seeds for the hard-negatives stage's human review queue.

Design constraints

• Label noise budget: the 8.7% false floor means ~1 in 12 auto-positives is wrong. That's comparable to OSM noise in fast-growing regions, but it should be (a) measured per training run (fraction of glint-sourced chips), and (b) ablatable -- keep source=glint all the way into the chip index.
• No circularity: glint-sourced positives must be excluded from any val/test split used to evaluate the retrained model (they were selected by the pipeline being evaluated). Spatial holdout as usual (val_tiles).
• Ordering: most valuable after tile-batched coverage (docs/issues/glint-tile-batched-coverage.md) multiplies how many validated candidates exist per run; a few hundred queried candidates yield too few confirmations (validation rates are 16–31%) to move an 8k-chip training set.
• The training positive threshold (MIN_PV_AREA in chips.py) still applies -- glint positives below it would be burned as ignore, which is fine.