2026-10-19

LEASH — supervise the stronger model after handing off. A weak agent's latent uncertainty controls escalation and continued oversight. On OSWorld, the authors report 73.66% success at $0.4117 per task, versus always-on Claude Opus 5 at 72.90%, with 19.2% lower inference cost. Read it for the matched-rate random-escalation comparison; reference-trajectory acquisition costs remain unclear in the abstract. → LEASH (source)

Mosaic — train across session boundaries. Hierarchical credit assignment incorporates session usage and marginal contribution. With Qwen3.5-4B on BrowseComp, it beats advantage broadcasting while using fewer sessions in compaction and subagent-forking scenarios. Read it for learning session allocation; the abstract supplies no numerical gain or total-compute comparison. → Mosaic (source)

• Anthropic proposes committing 5% of its shares to a wholly-owned public-benefit LLC, subject to stockholder approval. The board may add 2% for each 50% valuation increase above a baseline. Governance details and milestones are still forthcoming; this is not a cash distribution. → Anthropic (source)

• Grok Voice 3.0, LEASH and Mosaic — one model record and two research notes. (voice) (routing) (training)

• Grok Voice 3.0 records EVA-Bench accuracy 74% separately from its other voice metrics; this is the vendor's attribution to ServiceNow, not a fresh independent verification. (source)