Work

Pre-registration: equal-budget benchmark of group-blind target-image acquisition for PAD-UFES-20 to DDI transfer in skin-cancer classification (protocol v1.1)

Year: 2026
Type: article
Abstract: Pre-registration (protocol v1.1, frozen 15 November 2026) of an equal-budget benchmark of group-blind target-image acquisition for skin-cancer classification transferred from PAD-UFES-20 to the Diverse Dermatology Images (DDI) dataset, with the exact code, analysis plan, sandbox feasibility study and review history. No DDI or MRA-MIDAS data are included. Protocol v1.1 (`plan/PROTOCOL_v1.1.md`), frozen 15 November 2026. Authors: Sakchai Tangprasert (KMUTNB, corresponding author) and Suwit Chanthasen (KMUTNB). The DOI of this deposit is the only value absent from the frozen text. Review status of this registration. This protocol and its release package were reviewed three times in full by an independent AI reviewer (ChatGPT web, temporary chats, 15 November 2026); each full review returned FAIL, and every blocker raised was addressed in a subsequent revision. A fourth, scoped closure check of the last ten findings returned RESOLVED for seven and NOT RESOLVED for three (B15 nested public schema, B17 evidence-derived cohort eligibility, B6 selection-time observability of the exploratory per-tone comparator). Those three were fixed afterwards and checked by the project's Design Owner (Claude, an AI model) by reading the code and recomputing the deposited sandbox decision; they were not re-checked by any external reviewer, and the Design Owner did not execute the test suite. No human peer review has taken place. All raw review answers, the authors' responses and the Design Owner's decision records are included unedited under `reviews/` and `plan/`. > What was and was not verified. The sandbox decision (synthetic shift on public PAD-UFES-20 features) was reproduced byte-for-byte by the external reviewer from the deposited raw rows. Hashes of restricted inputs (DDI and MRA-MIDAS features, manifests, split components, partitions) are commitments that no external party has checked; see protocol §15b for how a data holder can verify them. > State of the data at registration. At the time of deposit no acquisition policy had been run on DDI and no metric using DDI labels had been computed; DDI labels had been read only to assert the ingestion counts stated in the protocol. DDI and MRA-MIDAS had been analysed in the authors' earlier work, so the DDI analyses registered here are pre-registered analyses of previously analysed data, not first-use confirmatory tests. > Authorship of the protocol text. The protocol was drafted by AI models (Claude) under the direction of the authors, who approved it and are responsible for it.
Language: English
Subfield: Oncology
Field: Medicine
Open Access status: green
OpenAlex ID: W7233580904
FWCI:
Cites:
Cited by:
Related to: