P06 · A supplied real-image comparison, now exposed practice
==========================================================
Unzip this pack into a directory named heldout inside the extracted P06 starter
package. Keep the starter's subset.json: it identifies the original eight known
training/development images. Use the pinned Pillow environment for source replay.

Both supplied models were trained on the five original training images, for
100 updates at rate 0.3. Threshold 0.5 and an illustrative 0.7 per-image IoU floor
were frozen before scoring. Equal updates are not equal compute. No candidate
was selected/promoted using these new outcomes. comparison-plan.json identifies
the exact protocol digests and archive-order selection rule. Digests bind files;
they do not authenticate chronology, authorship or a learner's work.

The cohort is the first three complete official-test archive entries after
excluding all eight original IDs. Selection used no model outcomes. The three
source photographs are distinct IDs/bytes; original photographer/animal
relations and near-duplicate independence are not established. This convenience
cohort does not support population conclusions or breed-generalization claims.

RESULTS ARE NOW PUBLIC. Every subsequent use of this cohort is exposed practice.
The archived local history records pixel first and convolution repeat because
both use the same images; both candidate protocols were already frozen. Those
labels record exposure, not candidate validity or independent selection evidence.
Preserve the included exposure-history directory when replaying; do not create
an empty ledger to make the recorded cohort look new. File deletion/editing can
defeat local history. There is no authenticated cross-device exposure record.

From the starter directory:
  python perception-prepare.py verify --manifest heldout/sources.json --batch heldout/prepared.json --out heldout/source-check-new.json
  python perception-evaluation.py evaluate --protocol heldout/pixel-frozen.json --batch heldout/prepared.json --source-manifest heldout/sources.json --ledger heldout/exposure-history --out heldout/pixel-replay-new.json
  python perception-evaluation.py evaluate --protocol heldout/convolution-frozen.json --batch heldout/prepared.json --source-manifest heldout/sources.json --ledger heldout/exposure-history --out heldout/convolution-replay-new.json

Use new output names every time. Original files are read only. Replayed masks
and metrics should match the archived results; exposure says repeat practice.
reference.json independently recomputes all predictions using NumPy matrix
operations and separate Boolean metric arithmetic, without importing the
teaching forward/score functions. This verifies arithmetic, not learner skill.

Inspect every image and both constant baselines in reference.json. The distant
dog has very few classified pet pixels: predicting all background earns 96.3%
accuracy but pet IoU is zero. Convolution improves the white dog's overlap but
still fails the other images. Neither model meets the frozen requirement on
every source. Do not replace the declared gate with a favorable aggregate.

Study tasks: predict the failure from the image before opening scores; compare
overlap versus accuracy; locate foreground misses and false positives in the
saved row-major masks; propose one intervention and explain how you would test
it on separately sourced data. A result on this exposed pack is development.

ATTRIBUTION.md and source/README.txt preserve dataset authors, current
CC-BY-SA-4.0 statement and the legacy archive notice. Additional images are
samoyed_36, german_shorthaired_3, British_Shorthair_64. Their unmodified JPEGs
and trimaps are retained; prepared.json contains the aspect-preserving BOX RGB
and NEAREST trimap adaptations. The general attribution's original eight-image
selection describes the starter; this pack's selection is in comparison-plan.
Independent source/annotation review, learner implementation review and pilots
remain pending. This pack does not award project completion.
