Claude's 26.8% protein success hides 0% failure rate

---

Quick Scribbles

• Anthropic — Claude's 26.8% protein design success masks 0-80% range across targets.
• Vercel — New fx agent cuts MCP tool schemas from 37 to 15.
• Z.ai — GLM-5.3 found 2,436 vulnerabilities but only 53 disclosed publicly yet.
• ContextFusion — Open-source tool cuts LLM context tokens by 60-99% while preserving quality.

---

Stay Connected

Subscribe to BrainScriblr for the latest AI developments delivered to your inbox.

---

Good morning, AI Knowledge Worker. Claude's protein design campaign posted 26.8% success. That headline masks 0% hit rates on specific targets. TREM2 reached 80%. Maltose-binding protein delivered zero.

What does portfolio-average performance mean when your single target might land in the zero zone?

In today's BrainScriblr:

• Claude's 26.8% protein success hides massive per-target variance
• Vercel's fx agent cuts MCP schemas from 37 to 15
• Z.ai found 2,436 vulnerabilities but disclosed only 53
• ContextFusion slashes LLM context tokens by 60-99%

---

Claude's 26.8% Protein Design Success Hides 0% Failure Rate on Some Targets

The Scoop: Anthropic's August 18 protein design campaign produced 354 binders from 1,320 designs—a 26.8% success rate. But maltose-binding protein went 0-for-90 while TREM2 hit 80%.

The Technical Details:

• Claude orchestrated 24 workflows across PXDesign, RFdiffusion3, Genie 3, and six other open-source tools.
• Independent validation by Adaptyv Bio ran 1,320 anonymized designs through SPR at five concentrations.
• Published per-target rates ranged from 80% (TREM2, 72/90) to 0% (MBP, 0/90).
• Binomial test rejects single shared hit rate at p ≈ 1.8 × 10⁻⁶³.
• Beta-binomial refit puts 38% probability a new target lands below 10% industry floor.
• Folding-model confidence scores could not distinguish 80% targets from 0% targets in advance.
• Complete design sequences and measurement data published on Hugging Face for reproducibility.

Why It Matters for You: The 26.8% headline is a portfolio average across sixteen targets. Your single target is a draw from a distribution spanning 0% to 80%. A 30-design batch carries a 20% chance of returning zero binders. Resource planning against the mean leaves you unprotected against the lower tail. Traditional procurement models assume consistent per-target performance—this variance breaks those assumptions.

The Bigger Picture: Every agentic benchmark reports one pooled pass rate across wildly different tasks. SWE-bench, ARC-AGI, and internal eval suites hide the same spread. Your ticket's success probability is not the benchmark mean—it's whatever your item-level difficulty implies.

---

Vercel's Coding Agent Solves MCP's Token Bloat Problem: 15 Schemas Instead of 37

The Scoop: Vercel's fx agent cuts MCP tool schemas from 37 to 15 per request. Traditional MCP harnesses send every tool on every turn.

The Technical Details:

• Two-tier architecture: 17 built-in tools plus dynamic subagent host for MCP servers
• Standard four-server MCP setup sends 22,226 bytes of schemas per request
• fx loads only relevant MCP servers dynamically instead of exposing full catalog
• Schema filtering keeps 15 function definitions in context versus 37 in traditional approach
• Built on technology from Vercel's recent Cursor acquisition for coding workflows

Why It Matters for You: Token bloat directly increases API costs and degrades model performance. Four MCP servers can triple your per-request payload without adding value. Dynamic loading cuts those costs while reducing tool selection errors. Models perform better with focused tool sets versus entire catalogs. This architecture pattern will influence next-generation agent framework design across the industry.

The Bigger Picture: Agent frameworks face the same challenge search engines solved decades ago. Showing everything reduces relevance and increases noise for decision-making systems.

---

Z.ai Found 2,436 Vulnerabilities But Only 53 Are Public: The Disclosure Pipeline Crisis

The Scoop: Z.ai's GLM-5.3 discovered 2,436 vulnerabilities across 269 projects. Only 53 are disclosed—2,383 remain queued.

AI got unexpectedly good at finding flaws. The patch pipeline didn't scale with it.

The Technical Details:

• GLM-5.3 scored 84.5% on CyberGym (finding flaws) but only 54.4% on ExploitBench (weaponizing).
• The 23.6-point gap means discovery capability outpaced exploitation by a full tier.
• Ledger breakdown: 107 critical, 990 high-severity findings across Linux kernel, Redis, WebKit, FreeBSD.
• Vulnerabilities averaged 26.6 years old—the oldest from 1981—always there, never affordable to find.
• Real-world measurement (Aug 20): 39.3% of top PyPI packages shipped zero releases in 90 days.
• On npm, 28.3% of core packages had no release in the same window.

Why It Matters for You: Coordinated disclosure was designed for scarcity. Discovery just became abundant.

Triage load arrives before exploit risk. Staff accordingly.

Maintainer cadence sets your ceiling. Median package ships every six weeks.

Inventory blindness turns fixes into incidents. Know dependencies before disclosures land.

The Bigger Picture: Finding vulnerabilities now costs tokens. Fixing them still costs maintainer hours.

This mirrors algorithmic discovery systems. The generator scaled. Everything downstream did not.

---

ContextFusion: Multi-Objective Knapsack Optimizer for LLM Context Windows

The Scoop: Most LLM apps waste tokens by stuffing retrieved chunks into prompts. ContextFusion cuts 60-99% of tokens while preserving answer quality.

The Technical Details:

• Multi-objective knapsack solver selects maximum-value content blocks within token budgets.
• Task-specific representations generate QA extractive, code signature, and agent condensed variants.
• Delta fusion computes incremental context changes across agent turns, preventing token churn.
• Provider adapters compile OpenAI chat.completions, Anthropic messages, and Ollama local formats.
• Cache-aware assembly segments stable system instructions from dynamic content for reuse.

Why It Matters for You: Token reduction translates directly to API cost savings. A 98.9% context reduction means 100x fewer input tokens billed. Latency drops proportionally—smaller prompts generate faster responses. Implementation requires middleware integration, not architecture rewrites. Delta fusion solves the agent memory problem without conversation history explosions.

The Bigger Picture: LLM cost optimization mirrors database query optimization from the 2000s. As usage scales, naive "send everything" approaches become unsustainable economically.

---

🌍 AI for Good

1. Humanitarian AI: Lessons Learned and Trends for 2026 The Humanitarian Leadership Academy releases comprehensive research mapping current AI practices in humanitarian work, including insights from their January 2026 pulse survey. This report provides critical guidance for aid organizations navigating AI implementation in crisis and conflict settings. — Humanitarian Leadership Academy, 2026-02-01

2. UNICEF Harnesses AI to Scale Solutions for Children's Climate, Health, and Education Needs UNICEF's Office of Innovation is building purpose-driven AI solutions with partners to create social good for children in emerging markets. The initiative backs diverse builders and scales AI applications addressing critical challenges in underserved regions globally. — UNICEF Office of Innovation, 2026-08-15

3. AI for Good Initiative Releases Impact Report on AI Serving Humanity The ITU's AI for Good platform publishes comprehensive analysis showing how AI is transforming global efforts in healthcare, education, resource management, and environmental challenges. The report guides innovators and policymakers in leveraging AI to address society's most pressing problems. — AI for Good (International Telecommunication Union), 2026-07-09

---

Partner Spotlight

Support BrainScriblr while discovering powerful AI tools (affiliate links):

• n8n — No-code automation platform for AI workflows
• Hume AI — Emotional intelligence API for human-centered AI
• Railway — Cloud platform for deploying AI applications
• Cudo Compute — Distributed cloud computing for AI workloads

---

Worth Your Inbox

Discover more quality AI and tech content:

• SemiVision — Semiconductor industry insights and AI chip developments
• Turing Post — Deep technical analysis of AI research and breakthroughs
• FinOps Weekly — Cloud cost optimization and financial operations
• CoreUpdates — Essential tech updates and startup intelligence
• The Multiverse School — Learning and development in the AI era
• Simple AWS — Practical AWS tutorials and cloud architecture
• EarthConscious — Sustainable living and environmental consciousness