AI News HubLIVE
站內改寫1 分鐘閱讀

待翻譯:Claude Opus 4.6 returned nothing 900/900 times. Should agents retry?

AI 服務暫時不可用,以下為來源摘要,待恢復後補全翻譯:Published July 29, 2026 | Version v1 Preprint Open Cross-Vendor Semantic Void Matrix Authors/Creators Pal, Rayan (Researcher) Description This preprint reports a frozen cross-vendor evaluation of successful zero-visible…

來源Hacker News AI作者: aiagenttester

AI 服務暫時不可用,以下為來源正文,待恢復後補全翻譯。

Published July 29, 2026 | Version v1 Preprint Open Cross-Vendor Semantic Void Matrix Authors/Creators Pal, Rayan (Researcher) Description This preprint reports a frozen cross-vendor evaluation of successful zero-visible-byte language-model executions across 11 exact model identifiers from OpenAI, Anthropic, Google, and Moonshot. It defines a Void as a model execution returning a successful provider response with exactly zero visible UTF-8 output bytes; provider termination metadata determines its subtype, while explicit refusals, safety blocks, tool-mediated executions, and operational failures remain distinct non-Void outcomes. Across 31,430 completed single-turn trials, the matrix observed 11,658 Voids (37.09%). In 4,290 strict matched semantic pairs, null arms produced 2,505 Voids while matched output-licensed controls produced 0. The study distinguishes four equal-status subtypes—V0, V1, V2, and VU—and separately preserves visible responses, Near-Voids, refusals, safety blocks, and infrastructure outcomes. At a 16,000-token ceiling, 313/500 trials remained Voids, all V0 or V2, showing that recognized output-budget termination does not explain all observed cases. Complete raw attempts, retries, request configurations, provider metadata, and cryptographically linked event records are publicly preserved. Verification replayed 62,968 event hashes and checked the complete 31,484-record raw-attempt set. The findings establish a reproducible, semantically structured class of successful zero-visible-byte executions without claiming an internal mechanism, intention, shared architecture, or causal explanation. Files Cross_Vendor_Semantic_Void_Matrix.pdf Files (170.1 kB) Name Size Download all Cross_Vendor_Semantic_Void_Matrix.pdf md5:3f296f97fed27e0877138e601536e65f 170.1 kB Preview Download Additional details Software Repository URL https://github.com/theonlypal/void-matrix-complete-analysis Programming language Python