Grail Computer · research record
Dated experiment record. Statements describe this run, not current submission readiness. Historical H. pylori category names do not establish infection; copy screens are bounded tests. Claims that a source does not exist mean none was identified in that recorded search, not proof of absence. Licensing interpretations in the notes remain unconfirmed. Local access details have been omitted. Current limitations and remaining work.

Copy screens on the ten final files

Both screens were re-run on the ten files in images/, that is on the post-overlay bytes for the five images that carry organisms. Nothing is flagged.

1. phikon-v2 embedding screen

research/sgh-synthetic-histopathology/scripts/pixcell_embed_screen.py, 1024-px tiles, model owkin/phikon-v2, device mps, fail line 0.95, --records screens/records-for-screen.jsonl.

file category arm provenance tiles max cos median tile cos margin own-source tiles prov. fields
SGH_synthetic_01.png gastritis f3 B_si3_donor procedural-layout 8 0.8968 0.8662 0.0532 8/8 (1.000) 1
SGH_synthetic_02.png gastritis f3 B_si3_donor procedural-layout 8 0.8995 0.8924 0.0505 8/8 (1.000) 1
SGH_synthetic_03.png gastritis a5 match-si3 mosaic 8 0.8044 0.7528 0.1456 3/8 (0.375) 10
SGH_synthetic_04.png IM a5 match-si3 mosaic 8 0.8282 0.7576 0.1218 3/8 (0.375) 9
SGH_synthetic_05.png IM a5 match-si3 mosaic 8 0.8134 0.7720 0.1366 4/8 (0.500) 12
SGH_synthetic_06.png mixed a5 match-si3 mosaic 8 0.8462 0.7605 0.1038 3/8 (0.375) 12
SGH_synthetic_07.png mixed a5 match-si3 mosaic 8 0.8401 0.7561 0.1099 2/8 (0.250) 11
SGH_synthetic_08.png normal v1 a1 pass2-si3 single-donor 8 0.8721 0.8092 0.0779 5/8 (0.625) 1
SGH_synthetic_09.png normal f3 B_si3_donor procedural-layout 8 0.8651 0.8318 0.0849 7/8 (0.875) 1
SGH_synthetic_10.png normal f3 B_si3_donor procedural-layout 8 0.8623 0.8349 0.0877 6/8 (0.750) 1

own-source counts a candidate tile whose nearest reference among all 118 is any of the fields that actually supplied that canvas's tokens. The screen script itself can carry only one conditioner per output in its --records file, so it is given the primary donor and the full provenance set is applied afterwards to its per-tile records. For the two calibration points recorded by C2: a recoloured copy scores 0.987 on this screen and two unrelated real fields score about 0.66.

Read honestly: the tightest margins in the package belong to the four F3 images (0.0505, 0.0532, 0.0849, 0.0877) and to the one single-donor image (0.0779), and those are the five images conditioned on a single real donor field, at own-source fraction 0.625-1.000. That is donor dependence, not copying - the pixel screen below puts the same images at NCC 0.457-0.567 against those donors, far below the 0.92 flag and below the 0.85 that a generated field scores against its own reference in C2's calibration. It is nonetheless the axis on which these images are least defensible, and the five mosaic images (margins 0.104-0.146, own-source 0.250-0.500 over 9-12 real fields) are clearly the more conservative provenance.

scripts/pixcell_patch_search.py, 64-px windows, stride 256, threshold 0.92, min_flagged 2, in the harness's targeted mode: each image against every real field its own tokens came from, plus every reference field that is the nearest neighbour of any of its tiles in the embedding screen.

file pairs max NCC worst reference
SGH_synthetic_01.png 1 0.4565 hiesd-28f9b423__hpylori_gastritis__34688_2112__mpp0.25.png (own token source)
SGH_synthetic_02.png 1 0.5665 hiesd-9ded7e41__hpylori_gastritis__13504_5824__mpp0.25.png (own token source)
SGH_synthetic_03.png 14 0.6216 hiesd-4fdce538__intestinal_metaplasia__21376_28992__mpp0.25.png (phikon neighbour)
SGH_synthetic_04.png 13 0.6527 hiesd-324885ef__intestinal_metaplasia__87424_33792__mpp0.25.png (own token source)
SGH_synthetic_05.png 16 0.7889 hiesd-d123b43f__intestinal_metaplasia__64192_5376__mpp0.25.png (own token source)
SGH_synthetic_06.png 16 0.5655 hiesd-d123b43f__mixed__65408_6464__mpp0.25.png (own token source)
SGH_synthetic_07.png 14 0.6209 hiesd-d123b43f__mixed__53632_9728__mpp0.25.png (own token source)
SGH_synthetic_08.png 4 0.4740 hiesd-4fdce538__hpylori_gastritis__96000_4544__mpp0.25.png (phikon neighbour)
SGH_synthetic_09.png 2 0.4908 hiesd-40f3c0cf__normal__23744_3072__mpp0.25.png (own token source)
SGH_synthetic_10.png 2 0.5035 hiesd-e3367780__normal__70784_4416__mpp0.25.png (phikon neighbour)

Calibration for this screen (C2, 6 September): a recoloured copy scores 0.9995, a 64-px shifted copy 1.0, a stain-matched copy 0.9997, a generated field against its own reference 0.85, and two unrelated real fields 0.88. The 0.789 maximum here is below all of those.

Two honest limits. First, the NCC is targeted, not exhaustive: 83 pairs were searched, and an exhaustive search of 10 images against all 118 fields would be 1180 pairs, about 3.3 hours at the 10 s/pair this run averaged, which was not done. The embedding screen is the prefilter that justifies the targeting - a pixel copy always produces a very high embedding cosine. Second, neither screen tests the risk that actually applies to these images, which is inherited layout and donor dependence, not copied pixels; that is discussed in SELECTION.md.

3. Files

Download public Markdown export