+ We take real, auditable incidents and freeze them at the moment before the answer was revealed. + Models must redo the work with only what was knowable then, against a rubric frozen before any output existed.{' '} + The cases are never published. The scores always are. +
++ Same band (high), same 23 cases, same public hash index. Five rank tiers; ties are labeled as ties. + Glossary: case = one task (from a real incident); paper = one sitting; band = reasoning-effort level. + These are static chart snapshots: open an image at full size, or follow its source details. +
+})
+
+ + The nine-axis completion matrix: each cell is a face's pin-level completion. + Tied totals hide completely different profiles — the matrix shows what the sum can't.{' '} + Rank by work uses eight work types and preserves each record's own week and case count. +
+})
+
+ + Every weekly issue runs like an exam: cases sealed in advance, a clean-room sitting, + per-paper wire audits, anonymized publication, and a public index anyone can re-check. +
+})
+
+ Each comes with named boundaries and counterexamples in the original writeups.
+})
+
+ })
+
+ })
+
+ One repo per lane (model × endpoint × band), one matrix per week, wire audits and hash cross-checks included.
++ The published glm-5.3-flash 17/23 and qwen3.8-27b 16/23 totals are W37 composites: + W36 results of 15/21 and 14/21 plus two new ops cases each. The axis picker retains their W36 records.{' '} + Read the method and score-period notes. +
++ Read the method: AMBER Core Specification ·{' '} + public hash index ·{' '} + roadmap +
+AMBER / PUBLIC BENCHMARK
+ +真实公开成绩 · 2026-W38
+考于 W37 · W38 未重考
+榜首成绩 · 23 案题库
+先补齐证据，再回到首屏。
+跨家复测还未齐，暂不声称能力回退。
+便宜候选需要成体系的价格 / 成本源。
+次级区只记缺口，不替数据提前下结论。
++ 琥珀评测把真实、可审计的历史事件封存到答案揭晓前，让模型只用当时可得的信息重做一次。 + 评分规则先冻结，题库不公开，分数和校验材料公开。 +
+公开材料可以复核成绩，但不会泄露私有题库。
+每个仓对应一条模型结果记录；名称、分数、期号和链接沿用公开材料。
++ Ollama 的 glm-5.3-flash 与 CrofAI 的 qwen3.8-27b 使用 W37 补考合成成绩： + W36 的 15/21、14/21 各加两案，成为 17/23、16/23；d41f 的 17/23 为 W37 全库首考。 + 按轴榜保留这两条 W36 的 21 案记录，不是这份合成总分。 + 原始说明：Ollama W37 ·{' '} + CrofAI W37。 +
+这些链接是方法和公开记录的原始入口。
+ ++ 选择你要做的工作类型，下表先比较所选类型里的最低通过数，再比较合计。 + 这让 12 条公开车道的差异更容易核对。 +
+ +每个 chip 显示该类型里满分车道的数量。默认不选任何类型。
+规则：先比最低通过数，再比所选类型的通过总数；并列会保留。
+| 选择一种工作类型后显示车道。 |
数据来自公开成绩仓。期号和案数随原始记录保留；方法与公开哈希索引见 方法。
+