NVIDIA and Elith released Nemotron-Personas-Finance-Japan, a synthetic Japanese persona dataset. It has 1M records with 50 fields each, adds household finance and financial literacy data, and uses the CC BY 4.0 license.
How we checked
Checked the dataset page: exists, CC BY 4.0, 1M rows, 50 fields
Why it matters
Teams building financial agents for Japan get commercially usable test users without touching real customer data.
Try this
Load nvidia/Nemotron-Personas-Finance-Japan with the Hugging Face datasets library and look at a few records before using them as simulation seeds.
In t3code PR #22557, merged October 21, 2026, about 11,000 of 25,928 tests judged low-signal were deleted to make CI faster and less flaky. Review restored 22 auth-boundary tests.
How we checked · Why it matters · Try this
How we checked
The GitHub PR page confirms the title, test counts, merge date and restored tests
Why it matters
It is a real, inspectable example of cutting agent-era test bloat. The video's claim that only 4 tests came back is not confirmed by the PR page.
Try this
Read PR #22557 in pingdotgg/t3code to see which kinds of tests were judged low-signal before you prune your own suite.
Claude Code v2.1.321 fixes MCP OAuth requests that sent secrets such as refresh tokens to another host after a server redirect. It also fixes a read-only Bash command running outside the sandbox without a permission prompt.
How we checked · Why it matters · Try this
How we checked
Read the official GitHub release notes for v2.1.321; no sandbox test today
Why it matters
Older versions could leak MCP sign-in tokens or skip sandbox and permission checks. Anyone who uses MCP servers with OAuth should update.
Try this
Run claude update, then check that claude --version reports 2.1.321 or later.
Anthropic and Diode report that Fable 5.2 designed a speakerphone, a voice-assistant hub and a handheld pet, covering PCB layout, enclosure and firmware. They say the boards worked without circuit fixes, and Fable 5.2 scored 69% on Diode's PCB Gym versus 25% for Fable 5.
How we checked · Why it matters
How we checked
Read Anthropic's research post; the results and benchmark scores are reported by Anthropic and its partner only
Why it matters
This is a concrete test of agents doing hardware design. Anthropic has a commercial relationship with Diode, and the devices are unvalidated prototypes.
Anthropic's docs (quoted by Simon Willison) say Fable 5.2 merged repeated code in about half of small coding tasks when told the codebase is its own, and in under a tenth without that sentence.
How we checked · Why it matters · Try this
How we checked
Found the quoted text and numbers in Anthropic's prompting guide; these are Anthropic's own test results
Why it matters
One sentence in the system prompt changes how much the model refactors. The same docs warn it also adds unrequested changes and about 1.5x the output tokens.
Try this
Add the ownership sentence only in repos where you want refactoring, and review the full diff, not only the model's summary.