Tenkai Daily — September 28, 2026

Model Releases

• TeleOCR — Multimodal OCR built on Qwen2_5_VL for conversational text extraction and grounding across Chinese and English documents. Backed by arXiv:2608.12898 on document-level multimodal reasoning. 📄

Open Source Releases

• sst/opencode v1.18.33 — Cloudflare AI Gateway timeout handling fixes, better MCP browser error reporting, credential redaction in debug output. 🛠️
• k9-aif 1.12.8 — Architecture-first framework for governed multi-agent systems with structured interfaces, lifecycle management, and governance hooks. 🛠️
• eval-judge-mcp 0.1.0 — Output judging for agents/RAG without LLM-as-judge; uses the Laya framework for relevance, hallucination, and prompt injection checks. 🛠️
• mylonite 0.10.4 — Autonomous AI agent security testing that generates regression tests from findings and gates CI pipelines on vulnerabilities. 🛠️
• ACCO 1.23.0 — Context optimizer for AI coding agents: manages retrieval, memory, routing, and output to cut token overhead via MCP. 🛠️
• shieldpi-mcp 0.5.1 — MCP server for ShieldPi Watchtower: LLM security scans, guardrail generation, incident triage, and kill-chain reconstruction from IDEs. 🛠️

Research Worth Reading

• When Is a Multi-Agent Code Judge Actually Grounded? — Two label-free diagnostics for code judge grounding, plus a judge that abstains when it lacks evidence. Tackles overconfident LLM verification. 📄
• Auditing LLM-as-Judge Failures in Text-to-SQL — Production pipeline audit finds Cohen’s kappa of 0.04 vs human gold; proposes repair strategies to reduce over-flagging. 📄
• Cartograph: Federated Tool Discovery — MCP proxy that cuts tool discovery from O(n) to O(k) via operator attestation. Scalable agent tool usage in large environments. 📄
• ScopeBench — Bench for whether agents respect engagement boundaries during autonomous pen testing under goal pressure. 📄
• TGL-NSGA-II for TinyML Architecture Search — Teacher-guided fitness ranking approximation for expensive evolutionary NAS; keeps search quality while slashing compute. 📄
• AI-Text Detection Neurons in Frozen BERT — Sparse probing and activation patching reveal specific BERT neurons driving AI-text detection across generator models. 📄

AI Dev Tools

• claude-mem — Persistent context compression across agent sessions; captures, compresses, and re-injects relevant history into Claude Code, Codex, Gemini, and Copilot. Solves the multi-session memory gap. 🧠

Today’s Synthesis

The area of the triangle with vertices at (0,0), (3,4), and (5,0) can be found using the coordinate formula or base-height method. Using the base from (0,0) to (5,0) with length 5 and height 4 (the y-coordinate of the third vertex), the area is (\frac{1}{2} \times 5 \times 4 = 10).

10