GuojieZBtrainer / ZBLL + ZBLS — 2026-10-28 An independent trainer derived from GuojieEOtrainer. The original EO website has not been changed. Application source is served as plain JavaScript and distributed under the Mozilla Public License 2.0 (LICENSE.txt). Cube facelet layout, move permutations and EO convention: https://github.com/crystalcuber/crystalcube Upstream commit b621dff9426ea8e954ca4362d29813510acfc864 (MPL-2.0). The cube module and ergonomic heuristic are reused from GuojieEOtrainer. New code adds notation normalization, verified case lookup, ZBLS last-slot matching, specialist practice, favorites, 3D playback and offline caching. 3D model: huazhechen/cuber, MIT, Copyright (c) 2018 huazhechen. https://github.com/huazhechen/cuber Version 4e8e5b4aa6611b5bc8d208efc037923b61264155. Three.js 0.130.1 is bundled locally under MIT. See licenses/huazhechen-MIT.txt and licenses/three-MIT.txt. ZBLL case identifiers and public algorithm sequences: https://github.com/CYRickyArAr/zbll_site_static https://github.com/CYRickyArAr/zbll_site_static/blob/main/data.js Retrieved 2026-10-28. Dataset generatedAt: 2026-10-25T13:32:33+08:00. 472 non-PLL cases, 875 public sequences. Mark labels are source annotations of use in reconstructions, not claims about a solver's current preferred alg. 21 conventional PLL algorithms are included separately, making 493 cases. ZBLS case identifiers and public algorithm sequences: Morgan Yeh. https://github.com/morganyeh06/zbls-trainer Classes and Data Files/algorithms.js; index.html. Retrieved 2026-10-28. 302 cases, 623 usable public sequences. Two candidate sequences were rejected by stage-state validation: 10-3: U F' U' F' U' R U R' 13-2: U R U' R' U R U' R' d' R' U R Both cases retain valid alternatives. No source website code, artwork, player portraits or externally hosted cube images are redistributed. All accepted sequences are checked against their stage prerequisites. Every displayed answer is additionally replayed in the facelet simulator and accepted only if its goal is met. ZBLL lookup covers all 7,776 legal EO last-layer states, including the four solved/AUF states. ZBLS ignores unrelated last-layer corner/edge permutation but checks the last pair, Cross, all four F2L pairs, and full EO at the goal. Already-filled last-slot inputs can receive an EO-only finish. Solutions may temporarily disturb completed pieces; their final state must satisfy the stage goal. Practice uses case selection plus optional U adjustments, with varied last-layer backgrounds for ZBLS. It is not a WCA uniform random-state scrambler. User inputs must satisfy the selected stage's prerequisites. The interface accepts the selected stage's prerequisites. It does not offer a general full-cube solve workflow or guarantee optimal algorithms. 2026-10-28 embedded search and ergonomic scoring: - Search starts only when the visitor clicks "寻找更多解法" after viewing a library answer. Initial recommendations do not start background search. - ZBLS uses GuojieEOtrainer's pattern-database iterative-deepening search with a four-F2L-pair plus full-EO goal. It may temporarily disturb pieces. - ZBLL uses cubejs's Kociemba two-phase solver, searching different cube orientations and first moves. Solver scripts are bundled locally unchanged. https://github.com/ldez/cubejs (lib/cube.js, lib/solve.js), retrieved 2026-10-28. MIT: Copyright 2013-2017 Petri Lehtinen; Copyright 2018 Ludovic Fernandez. The full license is provided in licenses/cubejs-MIT.txt. - New results are independently checked, deduplicated, and ranked together with library algorithms. Search updates preserve the selected playback. - Each round runs approximately eight seconds excluding initial table setup; a single ZBLL query may finish after the round budget. Stop and resume are supported between search steps. Finite search is not an optimality claim. - Numeric ergonomic costs, estimated regrips, awkward transitions and move-by-move grip instructions are shown. Lower cost ranks first. Plain outer turns retain the EO scorer. Wide/slice turns and rotations have explicit experience costs; grip across mixed segments is approximate. Scores are not measured execution times or guarantees of personal comfort. Favorites, settings and daily viewed counts are local to the browser. Viewing the last playback step counts as viewed, not independently solved. Favorites can be exported and imported as JSON. No account data is collected.