Offline Handwritten Signature Forgery Detection: A Single-Image Classification Pilot, a Reproduction Audit, and a Preliminary Writer-Independent Verification Study on BHSig260

Offline handwritten signature verification is a writer-independent comparison problem: a questioned signature is scored against reference signatures of the claimed writer, systems are tested on writers absent from training, and performance is reported as the equal error rate (EER) with false acceptance and false rejection rates. This paper reports two linked pieces of work in that lineage, and states plainly what each one does and does not establish. The first is a final-year undergraduate project: a frozen MobileNetV2 network with a compact classification head trained to distinguish genuine from forged signature images on a small public pilot set, deployed as an authenticated Flask web prototype. The author reported 90% validation accuracy from a single unlogged run of that pilot. An independent reproduction conducted for this paper does not reproduce that figure: re-implementing the preserved code on the downloaded data measures 0.673 validation accuracy on the original folder split — a split in which 78.0% of validation images are byte-copies of training images — and 0.722 ± 0.008 (mean ± SD, three seeds) on a clean deduplicated split. The 90% figure is therefore reported here only as an author-reported, unreproduced measurement, and the reproduction gap is disclosed as a finding, not hidden. The second piece of work reframes the task as writer-independent pair verification under a frozen protocol and reports the first measured results on the Bangla subset of BHSig260: a CPU pilot on a nested 6/3/3-writer subset, and one full GPU run (seed 42, full-backbone fine-tuning) reaching test EERs of 0.157, 0.136, and 0.127 at one, three, and five references. Those benchmark figures are preliminary — a single run, not a mean over the five protocol seeds — and are labelled as such throughout. The paper's contribution is a fully documented pilot, an honest reproduction audit, a pre-registered verification protocol with its first measured run and a complete deviation log, and a prioritised experiment plan that would bring the work to the benchmark standard.