e25fcda67d
Enrolling a packshot from an external set is normal - it is how F-12 was closed - and it silently turns one of those rows into training data. The external sets were scored by an ad-hoc script that never checked, so sku_l06_1.jpg was being reported as a held-out correct accept while being byte-identical to the enrolled view. code/eval_external.py now scores those sets with the same leaked() check the shelf evaluator runs, and refuses a row that has since been enrolled. That check needed fixing first. Its 64x64 signature fires on two different photographs of the same product: the l09 rear label from a retailer page sits 19.5 from the enrolled front packshot, under the 20.0 line, while being a different photograph (resized MAD 26.3). A signature hit is now confirmed by correlation - a genuine re-cut, re-encode or rescale correlates at 0.99+, that pair at 0.28 - so the two signals together are decisive where either alone is not. Verified against both directions: identical, q55 re-encode and 700px rescale are caught; the different-photo pair is not. Also adds the van backend (always-argmax, no gate) so the stream comparison in DECISION.md is reproducible rather than asserted, with two tests pinning that it answers where the shipped gate abstains and still refuses an empty score set. External result on the corrected basis: 19 scored, 12 correct, 7 abstain, 0 false accepts, 1 leak excluded. Suite 307 -> 309. |
||
|---|---|---|
| .. | ||
| kewpie | ||
| l12-reconciliation | ||
| other-mayo | ||
| tier3-sources | ||
| README.md | ||
External evidence (2026-09-29)
Independent-source media used to stress the identity pipeline. None of this is
Tier 2. DECISION.md's Tier 2 is a second labelled store visit; these are
external_proxy artifacts, and every manifest here says so.
| Directory | What it holds |
|---|---|
other-mayo/ |
8 accepted images for sku_l04, sku_l07, sku_l09 from 6 source domains, plus the scored result |
l12-reconciliation/ |
What sku_l12 actually is, from the shelf tag and package pixels |
kewpie/ |
Kewpie 450 g research (the near-unknown case) |
tier3-sources/ |
Two reusable pan clips and 11 rejected candidates for a future Tier-3 replay |
Leak test
Every accepted image was template-matched against add/image6.jpg. Highest
score across all 8 is 0.675 (sku_l07_1), and the 0.95+ band that identifies a
crop of the evaluation photo is empty. The images are genuinely external.
Result: 8 images, 0 false accepts
Scored through the shipped LocalLaneAdapter + DecisionPolicy against
stock-lopia-clean (other-mayo/eval_result.json):
| Verdict | Count | Files |
|---|---|---|
| Correct accept | 4 | sku_l07_1, sku_l07_2, sku_l09_1, sku_l09_2 |
| Abstain | 4 | all sku_l04_* |
| False accept | 0 | — |
Two things this establishes that the image6 shelf cannot:
sku_l07is answered with no enrolled image. Both l07 images accept through the text lane alone:sku_l07_1reads ラッキー フヨネーズ, andsku_l07_2reads the rear label's 製造者：丸和油脂株式会社 with 内容量：200g. The catalog row is 丸和油脂 ラッキーマヨネーズ 200g, so the manufacturer and the size together name it. This is the text lane's coverage claim - that a product with no reference image can still be named - holding on data that was not used to build it.- The l04 abstentions are correct behaviour, not a miss. l04 has no
enrolled image either, and its reads are weak:
sku_l04_2reads only コと旨 タ 味 (all rows score 0.0), andsku_l04_4reads the rear label with やまや garbled to まや, scoring 0.379 - below the accept line. Abstaining on a weak read is the designed outcome.
sku_l07_2 accepts with a negative local score (-0.55) on text evidence alone.
That is the lane doing its job, but it is worth naming: the accept rests on two
matched tokens (manufacturer + size), not on geometry.
What is still missing
No external source gives per-SKU identity for the whole catalog, and no reusable clip combines the target category with shelf-distance text resolution. The Tier-2 and Tier-3 gates in DECISION.md still need the operator's own recorded walks.