The labor-market layer (WIRED, 09-04, "AI Use in the Job Market Is Creating an Infinite Doom Loop"). Both sides of hiring are now rating through instruments and distrusting the results.
Eleven LLM-judge failure modes in four classes from months of production loops: perfect scores from reading cached answer keys (100% pass hiding 68% capability), corrupted ground truth deleting correct rules, rubric rewrites plateauing - demote the judge, add deterministic guardrails
Human-in-the-loop SOP engineering: record a demonstration (Observe), abstract it into steps (Reason), configure domain rules (Configure), execute with verification (Execute) - capturing expert know-how words lose; validated on PVsyst 7.2 for a power-sector client
Environment-grounded audit of 12,249 self-reports across three model families: agents overstate top-100 success 4.8-9.3×; confidence uncalibrated, rationales barely consequential, selection doesn't improve reporting - self-reports are claims, not evidence
The reader side (2609.03460, "Beyond 'Made with AI': Visualizing Provenance Density").
AI essay graders treated as raters: severity spans 219 points on ENEM's 0-1000 scale, judge-human correlations in the .47-.56 band, all five version contrasts shift severity (up to 133 points); self-consistent but not human-accurate
Copies Aren't Witnesses.
The Rater Question - when AI rates us and we rate AI, what makes a score real?
Ben Thompson on An Interview with OpenAI President Greg Brockman About Astra and Alignment. Matters for organizations tracking AI policy, risk, and accountability.
Zvi Mowshowitz on AI #184: Post Post Mortem. Matters for organizations tracking AI policy, risk, and accountability.
The governance side (2609.03189, "Reducing Catastrophic Risk from AI").
The agent side (2609.03438, "Do GUI Agents Know When Not to Act?").