flowchart LR
  T[Task] --> H{Harness layer}
  H --> M1[Persistent state<br/>across runs<br/>Prime Agent]
  H --> M2[Event log +<br/>Python kernel<br/>Scroll]
  H --> M3[Harness synthesized<br/>per task<br/>JIT-Agent]
  M1 --> MOD[Model<br/>weights unchanged]
  M2 --> MOD
  M3 --> MOD
  MOD --> OUT[Task outcome]
  OUT --> MEAS[Paired-run<br/>measurement<br/>Skill Lift]
  MEAS -->|rho 0.14 vs<br/>structural scan| GATE[Old review gate<br/>predicts little]
  OUT --> J[Judge lifecycle<br/>Netflix]
  J -->|meta-judge over<br/>judge reasoning| J
  OLD[Flat transcript<br/>re-read every turn] -.->|token cost| MOD
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class T input
  class H decision
  class M1,M2,M3,OUT output
  class GATE,OLD warn
  class MOD,MEAS,J aux

Why this cluster matters to you specifically. Your saved reading has been dominated by loop and harness engineering for a month, and this is the week the theme produced runnable artifacts rather than arguments. Two things to take from it. First, Scroll and Prime Agent independently reached for the same move, replacing transcript re-reading with code over typed state, which is a token-cost reduction before it is a capability gain and is directly comparable to the ALTK-Evolve result showing DeepSeek-V3.2 going from 80.4% to 89.3% task completion at 41% of the token cost. Second, Skill Lift's rho of 0.14 is the most immediately actionable number in the batch, because it invalidates a process that is currently running inside a lot of companies, and it costs nothing to replace with paired runs.

Routing, KV cache, compression, GPU

The efficiency thread arrived through papers, not through social

There is no social layer to synthesize on routing, KV cache, compression or GPU work today, so this section is short by design rather than padded. Two items are worth carrying here because they change what an efficiency-minded reader should do next week, and both are treated fully in the 08-31 digest.

• Compression now has a third cost line nobody prices. A COLM 2026 paper finds that standard weight pruning degrades the faithfulness of sparse autoencoders fit on the model, so the compressed model you actually serve is the one your interpretability tooling can least describe. Cost angle inverted: this is a cost of compression, not a saving from it. → summary
• Runtime enforcement got cheap enough to argue about. LMSM puts a pluggable interpretability-backed monitor inside the vLLM generation loop and retains 98.14% of unmonitored throughput while cutting attack success from 39.20% to 3.32%. At two percent, monitoring stops competing with capacity for GPUs. → summary
• Kernel agents got memory. A Kurate leaderboard entry gives a kernel-optimization agent an experience graph over past attempts and measured outcomes, arguing that accumulated structure beats more rollouts. Influence angle: if experience ports across hardware targets, the CUDA moat becomes a decaying cost rather than a fixed one. → summary

Industry and business

• SemiAnalysis published a negative result against its own expectation, finding no CVE-rate change in the Nvidia driver, CUDA, PyTorch, Kubernetes and Docker despite the industry's AI-cyber narrative. Influence angle: a checkable claim beats a press campaign, and they pre-committed to a series anyone can re-check.
• They also shipped free tooling rather than only an argument. pip install clustermax then cmax audit security auto-detects Slurm and Kubernetes clusters, VMs, bare metal and containers and checks versions against known vulnerabilities.
• Outcome-based pricing became a two-vendor pattern this week. OpenAI now lets some major customers pay only when the AI completes a task, days after Salesforce began negotiating Agentforce contracts priced on revenue closed or service cost automated. Cost angle: the vendor now eats every failed task, which pushes serving-cost optimization down the stack.
• Employee sentiment is moving the other way from the product roadmaps. Glassdoor positive AI mentions fell from 81% to 43% since 2019, split by role rather than by tool, with forced adoption and surveillance ranking alongside job-loss fear.
• Anthropic's legal exposure stacked again, with Sony, Warner and other publishers suing over musical compositions months after a $1.5 billion settlement with book authors.