Sealed predictions for AME2025: 7,780 nuclear mass excesses from 11 candidates, timestamped before publication

Pre-registered predictions of atomic mass excesses for 7,780 nuclides (N, Z ≥ 8) of the WS4 table that AME2020 does not list as measured. They were sealed and timestamped before the publication of the Atomic Mass Evaluation 2025 (AME2025), and will be scored with a rule fixed in advance on the nuclides that AME2025 lists as measured. 11 candidates: 4 reference mass models (WS4, DZ31, HFB-27, FRDM2012), sealed by the SHA-256 of the tables used and not redistributed; 6 corrections found by symbolic regression in earlier campaigns, none of which beat WS4 on a chronological test; 1 correction to WS4 proposed by an LLM (Claude Opus 5.5) from binned residual statistics only. Each correction comes with per-nuclide values and pre-registered uncertainties. The SHA-256 of the manifest (seal root d7d3c4dc1b7b4209234afb3934c8f6282addba6018b5ee9009f825878e443834) is anchored with OpenTimestamps in Bitcoin block 974431, mined on 2026-11-03 at 19:49:12 UTC. Main rule: for each of the 7 corrections, rms below WS4 and a one-sided sign test against WS4 at p ≤ 0.05/7 on the newly measured nuclides. Expectation written before sealing: "no gain" is the most likely outcome. Failures will be published too. Part of Projet Samaritain, an independent effort to test AI-assisted theory generators against strong canonical models. Developed with AI coding assistants (Claude, by Anthropic). See README.md in the archive; run verify_seal.py to check integrity.