self.md radar — 2026-11-11
the measurement grid stayed put
how do you check whether a beamline obeyed an agent’s change?
in VISION-CC’s new study, the scientist inspected the measured positions after the team lead confirmed the setting
1. a scientist spots the grid the agents had missed
the team-lead agent confirmed a minimum step of 0.01 mm
the measurements still landed on a 0.05 mm grid
a scientist noticed the discrepancy in a visualization; the specialist agent responsible for measurement steering then fixed the implementation
this appears in a new preprint on VISION-CC, which uses Claude Code teams to operate synchrotron X-ray experiments
separate agents handle the instrument, analysis, measurement selection and visualization
another reconstructs decisions and corrections from the conversation logs
the published repository includes 44 scattering frames and a sanitized ledger covering 25 minutes of one experiment
the study covers three experiments at one beamline, with one scientist and one agent platform
its audit labels and links were model-generated and carry uncertainty
reading: the X-ray study, Figure 4 and limitations · VISION-CC’s released data and ledger
2. Muse Ember’s headline price stops at a 200K prompt
Meta’s new muse-ember-1.0 has a 1M-token context window
the launch page’s $0.10 input and $0.50 output prices, per million tokens, apply only to prompts up to 200K
above that prompt length, input costs $0.30 and output $1.50
cached input moves from $0.004 to $0.012 per million tokens
if the prompt exceeds 200K, budget at the higher rates for both input and output
the coding comparison needs its footnote too
Meta reports Terminal-Bench 4.0 results using mini-swe-agent for Ember, Claude Code for Haiku 5.5 and Codex for GPT-6 Luna
those scores compare model-and-harness combinations, not models under identical conditions
Ember is available through Meta’s hosted API and Muse Code
reading: Muse Ember’s launch, availability and evaluation footnotes
3. five microphones narrow the search to one pile of boxes
the culprit was an old carbon monoxide detector asking for batteries
Simon Willison found it after having Claude Code build Beep Locator
two iPads, a phone, a MacBook Pro and a Surface Ultra listened around the room; he reports a location within about 40 cm, then dug through that pile
automatic positioning of the listening devices worked poorly
he added a mode for dragging their rough locations onto the map and checked sound localization by clapping from different places
a Python server coordinates browser microphones over WebSockets
replacement JavaScript reaches listeners without reloading their pages and triggering fresh microphone-permission requests
this is one reported household trial, not a room-to-room accuracy test
reading: Willison’s trial and failed calibration · Beep Locator’s setup and controls
the repository’s optional --auto-update mode pulls branch changes and restarts the server
its README warns that enabling it lets unreviewed branch code run on the host