Daily Executive Summary
This run adds five events and updates five. ThinkingBox is the most actionable: it checks terminal database state and side effects instead of trusting an agent report. DiarizationLM and Strata have public artifacts for one-day tests. Public Vault API docs move OpenAI Agents Vault from WATCH to TRIAL. Ataraxos and Figma hosted MCP now have primary evidence but remain WATCH.
Models
DiarizationLM (TRIAL)
Google opened a 4B speaker-label correction model with local GGUF files. Author tests improve WDER and cpWER on four corpora, but the multi-speaker meeting samples are small. Speaker drift means a long monologue is attached to the wrong identity even when the words are correct.
Claude Sonnet 5.5 (TRIAL, review)
Artificial Analysis reports about 193k output tokens and $7.60 per task at maximum effort, roughly 50% more than Sonnet 5 in its workload. The vendor's “up to 30% lower cost per task” is not a universal discount.
Agent & AI Engineering
ThinkingBox (TRIAL)
ThinkingBox exposes 507 stateful workflows through OpenEnv with 20 trials per model. Terminal-state grading checks the records and side effects left behind; an agent can exit cleanly without a tool error and still write the wrong field or do extra work.
OpenAI Agents Vault (TRIAL, rating change)
Public docs confirm that sandboxes receive placeholders and a network proxy injects secrets only for allowed hosts. The previous documentation gate is cleared.
Figma hosted MCP (WATCH)
Figma says only MCP Catalog clients can connect to its hosted server. Protocol compatibility does not guarantee hosted access.
Open Source
Strata (TRIAL)
Strata distributes Qwen3.8-Flash-Next experts across GPU, RAM, CPU and SSD and exposes local OpenAI- and Anthropic-compatible APIs. Adoption is strong, but speed and low-bit quality remain author-reported.
Research
Ataraxos (WATCH)
Ataraxos combines damped self-play and decision-time search for Stratego. The paper and MIT code are public, but no checkpoint is available, so reproduction is not a one-day trial.
Developer Tools
Claude Code 2.1.289 (ADOPT, update)
The release closes policy bypasses around nested commands, environment prefixes, bare assignments and IDE symlink reads, and adds teammate-agent lifecycle APIs.
Infrastructure
LiteLLM 1.104 (TRIAL, update)
The stable release signs Docker images, rejects weak master keys, detects breached passwords and fixes deactivated-user, MCP worker-routing and spend-batch replay failures.
Ollama 0.35.1 (TRIAL, update)
The release preserves Modelfile CAPABILITY declarations so clients recognize decision-only models and avoid offering chat, tools or thinking.
Business & Policy
No new policy event changed model access, API costs, chip supply or the open-source landscape.
Trend Signals
ThinkingBox, Claude Code and LiteLLM support trend #7 from consequence verification and permission enforcement. Figma's allowlist and LiteLLM governance support trend #2. No evidence-strong new trend emerged; all statuses and confidence levels remain unchanged.
Tech Radar
- New ThinkingBox (research / TRIAL): the public harness repeatedly checks terminal state and extra side effects.
- New DiarizationLM (foundation-model / TRIAL): local GGUF files enable speaker-correction tests, but meeting evidence is small.
- New Strata (infrastructure / TRIAL): adoption is strong; low-bit quality and stability need testing.
- New Ataraxos (research / WATCH): paper and code are verified, but checkpoints and independent reproduction are missing.
- New Figma hosted MCP (ai-agent / WATCH): hosted access uses a client allowlist.
- OpenAI Python SDK (developer-tools / TRIAL): Vault API docs clear the waiting gate, raising it from WATCH.
- Claude Code 2.1.289 (developer-tools / ADOPT): closes several managed-policy bypasses.
- LiteLLM 1.104 (ai-agent / TRIAL): strengthens identity, MCP worker routing and accounting.
- Ollama 0.35.1 (infrastructure / TRIAL): clients can recognize decision-only capability.
- Claude Sonnet 5.5 (foundation-model / TRIAL): independent testing shows maximum effort can cost more.
Current 30-day radar: ADOPT 9 / TRIAL 117 / WATCH 77 / IGNORE 0. See the site insights page for the full current radar.
Worth Trying
Prioritize two one-day tests: run ten ThinkingBox workflows five times each and inspect terminal state and side effects; run DiarizationLM GGUF on real calls and meetings for WDER, cpWER, latency and identity drift. Teams with 12–24 GB GPUs can also evaluate Strata.