Uncropped Film Scans and Analysis Data Addressing PubPeer Query on JBC M112.361083

1) Methods / Code / Results (lane‑agnostic + low‑res replication) 1A) Lane‑agnostic similarity search on the legacy film scans (full resolution) Data. /mnt/data/1c_akt_jbc2012.jpeg and /mnt/data/1e_akt_jbc2012.jpeg. Approach. No lane/band detection. The script samples windows across the entire image (multiple sizes), and for eachwindow slides it over the other image using normalized cross‑correlation (OpenCV TM_CCOEFF_NORMED). For the best match it computes: ZNCC_{\text{peak}} (peak normalized cross‑correlation): \rho=\dfrac{\sum\limits_{i}(T_i-\bar T)(I_i-\bar I)}{\sqrt{\sum_i(T_i-\bar T)^2\sum_i(I_i-\bar I)^2}} SSIM (structural similarity; luminance+contrast+structure). HPcorr (high‑pass residual correlation; a proxy for micro‑texture / grain agreement). Duplication‑consistent thresholds (pre‑registered, strict). We’d expect all three to be high simultaneously for reused patches: ZNCC_{\text{peak}} ≳ 0.95, SSIM ≳ 0.92, HPcorr ≳ 0.90, with overlays that visually collapse (no residuals). Result (full‑res, lane‑agnostic; light sweep). Windows kept after texture filtering: n = 44 Best scores among all cross‑image windows: ZNCC_{\max} = 0.971, SSIM_{\max} = 0.517, HPcorr_{\max} = 0.674 → These do not converge on duplication. Correlation alone is not enough; structure and grain disagree. Overlays of top pairs show clear red/green residuals—not copy‑paste behavior. Reproducible outputs (produced now): ZIP (deposit‑ready): zenodo_bundle_lane_agnostic_v2.zip Script: analysis/lane_agnostic_similarity.py (in the ZIP) Full‑res results: CSV: lane_agnostic_matches.csv Summary: summary.json Still‑frame “blend” triptychs (A | blend | B) for the top pair (to mimic the video’s “morph” effect without animation): frame0, frame1, frame2 These stills show that even for the best pair at full resolution, the band interiors and micro‑texture don’t lock. 1B) Low‑resolution replication of the Illex overlay idea (downsample + JPEG) To test the method, not the films, I down‑sampled each film scan to 400 px width and re‑encoded with heavy JPEG compression (Q≈40) to emulate a figure/screenshot pipeline. Then re‑ran the same lane‑agnostic search. Windows kept: n = 58 Best scores (low‑res): ZNCC_{\max} = 0.942, SSIM_{\max} = 0.713, HPcorr_{\max} = 0.483 → SSIM inflates at low‑res (visual smoothing makes shapes look more alike), but micro‑texture collapses(HPcorr drops). This is exactly the pattern expected when low‑res overlays look convincing while the underlying grain does not match. Low‑res outputs: CSV: lane_agnostic_matches.csv Summary: summary.json Low‑res triptychs (A | blend | B) of the top pair: frame0, frame1, frame2 Take‑home: The overlay video (Illex_Pubpeer) is a low‑res demonstration tool. At that resolution, different bands can be warped/blended to appear similar (higher ZNCC/SSIM), but grain/halation/speckle—the physics of film + scanner—do not match (HPcorr low). On the actual films, all three measures fail to reach duplication‑consistent levels. 1C) Why low‑res morphs can mislead (derivation‑level clarity) NCC is variance‑normalized covariance. Down‑sampling + lossy compression reduce intra‑patch variance and depress high‑frequency content, inflating the numerator/denominator ratio for similarly shaped intensity profiles. SSIM increases when local means/contrasts align—even when fine structure diverges. Low‑res blots are dominated by smooth “blob” envelopes; distinct blobs can score moderate SSIM. HPcorr isolates micro‑texture (grain, dust, halation). True duplication preserves that micro‑texture; overlayscreated from different bands, even if shape‑aligned, cannot reproduce it. Put simply: shape can be faked by scale/warp; grain cannot.