self.md radar — 2026-10-29
the dispatch log disagrees
a silo-services client says the business operates only in Alabama. its dispatch log records crews in five states. Surge AI has put that contradiction inside a tax assignment for an agent to discover.
1. DAYJOB: Tax’s best reported pass rate is 10.8%
Claude Sonnet 5.5 leads Surge’s new tax benchmark at a 10.8% pass rate. a pass requires satisfying at least 95% of an assignment’s rubric criteria, including the analysis and finished deliverable. that is a task-completion threshold, not a measure of how many individual facts a model gets right.
there are 50 expert-created assignments and 1,789 workspace files: returns, payroll records, contracts, correspondence, even a photographed handwritten receipt. the agent has to work out which evidence changes the client’s request before producing the return, workpapers or advice.
the public dataset contains two sample tasks, not all 50. the documented environment blocks web browsing and package downloads during a run.
reading: Surge AI’s DAYJOB: Tax release
2. MicroZig’s AI ban reaches the spelling fix
MicroZig’s new participation policy, effective October 28, bars LLM-generated code and prose. it also excludes paraphrasing, grammar editing, translation, shared brainstorming and finding bugs with an LLM. a human-written issue comment based on chatbot advice is unwelcome too.
the translation rule has a specific alternative: English is encouraged but not required. contributors may post in their native language and leave readers to choose translation tools themselves.
the policy covers the Zig Embedded Group’s GitHub organization and Discord server. it is not a rule for every Zig project.
reading: MicroZig’s published Code of Conduct
3. the fingerprinting script has instructions for its reader
an anti-bot script examined by veritas contains misleading biological-research names and hidden Unicode text aimed at LLMs. the distraction is intended to interfere with model-assisted deobfuscation.
in the author’s attempts, Astra refused to analyze it; enabling Daybreak Blue did not change that. GLM-5.3 complied. these are reported attempts, not a controlled comparison between models.
underneath, the author found familiar browser fingerprinting signals: browser automation markers, language settings, hardware concurrency and WebGL renderer information. a refusal did not answer what the script was collecting.