RAI Daily · Preview · full text pending fact-check
A safety rule can survive context compaction in words and die in behaviour
A single context-compaction cycle can leave a safety rule's wording intact while stripping its force, on behavioural replay the degraded residue lets the model perform the prohibited action far more often than an intact rule does (+34 and +57 point gaps), and an audit that checks the text sees nothing wrong.
Focus areas
Why you can’t read the full briefing yet
This briefing is still being fact-checked.
Every daily briefing is researched on the day, then independently checked against its sources before the full text goes public. This one has not cleared that check yet, so for now you see the headline, the summary, and its focus areas. The full edition will appear at this same address once verification completes.
In the meantime
- Read the latest verified briefing: Monitor ensembles: skill governs, decorrelation does not
- Listen to this edition’s podcast. The recorded episode has already passed its own checks.
- Browse verified coverage of Agentic systems
- Subscribe to the RSS feed and the full edition reaches you when it clears.