$ cat briefs/daily/2026-11-17.md
2026-11-17
November 17, 2026 (Tue)
2 top stories · 2 paper picks · 1 research-access update
Top Stories
1. OpenAI names its embedded assessors
The 2026-11-16 announcement names Apollo Research, METR, Transluce and Goodfire. Scope and access remain bounded; publication allows confidentiality redactions and possible disclosure delays.
Why it matters: named engagements make assessment terms inspectable; they are not completed safety findings. → Embedded Evaluation (source)
2. Healthcare computer use gets a rubric-based comparison
OpenAI reports 94.3% for GPT-6 Astra and 91.8% for GPT-6 Luna on HealthAdminBench at maximum reasoning effort.
Why it matters: these measure earned rubric points, not completed workflows or proven clinical outcomes. → Agents (LLM Agents) (source)
Paper Picks
VibePrevisBench: 242 tasks test both executable Blender scenes and what their rendered videos visibly show. Across 11 agents, authors find that inspecting a render does not reliably lead to correcting it. Read it for verification beyond successful execution. → VibePrevisBench (source)
Continual-ARC: recurring tasks compare persistent harnesses with weight updates under a shared feedback budget. The current abstract reports little or no unseen-task gain from tested harness methods. Read it to distinguish memory from transferable learning. Both picks are abstract-level, author-reported evidence. → Continual-ARC (source)
Watch
New in Wiki
- The two paper pages add methods and evidence limits; no new entity, concept or person page. → VibePrevisBench · Continual-ARC (scene) (learning)
Updates
- Astra and GPT-6 Luna now link to the healthcare comparison's measurement limits. (source)