Morning Brief: Monday, November 16

Seventy-one feeds reporting. Two weeks. 5,850 items reduced to what follows. (what we track, how we crawl, subscribe)

Monday answers the week's largest open question, and the answer is much smaller than anyone wanted: one router, in rural Arkansas. The research batch spent the same morning moving agent governance out of policy prose and into runtime enforcement.

Saturday's brief ran on the fact that two federal officials had given two incompatible non-AI explanations for a national airspace stop. The cause now named is neither of them. A single FAA router in middle-of-nowhere Arkansas is being identified as the origin of nationwide flight chaos, which means the agency's denial about its new AI advisory system was probably accurate and the underlying condition is worse than an AI failure would have been. An AI system can be rolled back. A load-bearing router that nobody could name until it failed is a design, and it was not an isolated case this morning: ULA's Vulcan destroyed forty Amazon Leo satellites on ascent, and pranksters took over Bitly's link resolution inside X posts for several major brands. Three unrelated systems, three single points, one morning.

The arXiv crop is working the opposite problem and working it well. Soteria pairs formally verified planning with runtime enforcement; Securing Multi-Agent Systems by Design and Only as Safe as the Weakest Agent argue the containment has to be structural; TradeGauntlet and the Zafin AIOS case study test compliance controls on agents that move money; Co-Sight Harness externalizes control assets so enforcement is deterministic rather than prompted. The shift is from "the model should refuse" to "the harness will not permit," which is the only version of that sentence an auditor can check. The adversarial counterpart arrived the same day: weight-space programming got both its first benchmark and its first backdoor paper, several hours apart.

Top (5-7 min)

FAA Router In Middle-of-nowhere Arkansas Identified as Cause of Nationwide Flight Chaos
Slashdot, 2026-11-16. The resolution to the week's biggest unexplained outage, and it exonerates the AI advisory system everyone suspected. One router, national scope, no announced redundancy.
ULA Vulcan Rocket Failure Destroys 40 Amazon Leo Satellites
Slashdot, 2026-11-16. Forty satellites on one vehicle. Constellation economics assume launch is routine; this is the price when a single launch is not.
Soteria: Formally Verified Planning with Runtime Enforcement for Safe LLM Agents
arXiv cs.AI, 2026-11-16. Verification at plan time plus enforcement at run time, which is the combination most published agent safety work still skips. The enforcement half is what survives a jailbreak.
Securing Multi-Agent Systems by Design
arXiv cs.AI, 2026-11-16. The architectural argument: multi-agent security cannot be retrofitted onto a topology chosen for throughput.
Only as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-Agent
arXiv cs.AI, 2026-11-16. A failure mode with no single-agent analogue. Per-agent evaluation cannot see it, which is most of why it is still shipping.
Can LLMs Write Backdoors? An Empirical Security Study of Weight-Space Programming
arXiv cs.AI, 2026-11-16. Published the same day as the benchmark below. A subfield acquired an evaluation suite and an attack paper before it acquired a defense.
WeightBench: Evaluating Language Models as Weight-Space Programmers
arXiv cs.AI, 2026-11-16. Models editing model weights directly, measured. Read it alongside the backdoor paper rather than on its own.
Between Obligations and Evidence: A Participatory Study of Translating the EU AI Act into Technical Requirements
arXiv cs.AI, 2026-11-16. The gap between a statutory obligation and a testable requirement, studied with the people who have to bridge it. The honest version of compliance engineering.
How AI agents are changing the rules of software governance
InfoWorld, 2026-11-16. The trade-press framing of the same shift, useful for the version you can hand to a governance committee.
Should Jensen Huang Return as Nvidia's CEO? Supporters and Critics Disagree
Slashdot, 2026-11-16. Day six of an unresolved governance fight at the company the entire buildout depends on. No new facts, which is itself the state.
Is bad news for Jensen Huang also bad news for the AI industry?
TechCrunch, 2026-11-15. The question worth asking rather than the petition coverage. How much of the sector's capital plan was one person's relationships.

Themes

The governance layer moves into the runtime

Today's batch stops arguing about whether agents should be constrained and starts shipping the mechanisms. The common move is to pull enforcement out of the prompt and into something external, deterministic and auditable.

Co-Sight Harness: Externalized Control Assets and Deterministic Enforcement for Long-Horizon Business Agents
arXiv cs.AI, 2026-11-16. Control assets as artifacts outside the model, enforced deterministically. The design that makes an audit possible.
TradeGauntlet: Evaluating Compliance Controls for Autonomous Trading Agents
arXiv cs.AI, 2026-11-16. Compliance controls tested as controls rather than described as policy. Trading is the right domain because the rules predate the agents.
Can You Trust Your Finance Agent? Drift, Diagnosis and Auditable Intervention on One Shipped Financial Research Agent
arXiv cs.AI, 2026-11-16. One production agent, instrumented. Drift and auditable intervention measured on a thing that actually runs.
Operationalizing Governed Agentic Workflows in Regulated Finance: A Customer Zero Case Study using Zafin AIOS
arXiv cs.AI, 2026-11-16. The customer-zero writeup. Useful for the operational detail vendors usually omit.
G-CAST: Governed Continual Self-Training for E-Commerce Agents under Evolving Demands
arXiv cs.AI, 2026-11-16. Governance applied to the training loop rather than the inference call, which is where the drift originates.
Learn then Bet: Learning Interpretable Rules for Reliable Agent Intervention
arXiv cs.AI, 2026-11-16. Interpretable intervention rules, learned. A middle path between a hand-written policy and an opaque guard model.
Smart Home Agents Make Capable but Unpredictable Grid Operators
arXiv cs.AI, 2026-11-16. Capable and unpredictable in the same sentence, about physical infrastructure. The clearest case for enforcement outside the agent.
From AI-Enabled Remote Patient Monitoring Toward Agentic Robotics: Embodiment, Governance, and Care Integration
arXiv cs.AI, 2026-11-16. Governance treated as a design input to embodiment rather than a review stage after it.
AI-Voyager: A Public-Interest System for Human-AI Civic Reasoning to Contest and Shape the Rules That Govern AI
arXiv cs.AI, 2026-11-16. The participation side. Contestability as a system feature, not a consultation period.
When evaluation becomes intervention in clinical AI
arXiv cs.AI, 2026-11-16. Evaluating a deployed clinical system changes it. The measurement problem that most governance frameworks assume away.
Aligning AI Teams
arXiv cs.AI, 2026-11-16. Alignment at the level of the team of agents rather than the model. Short, and worth it for the framing.
Alignment by Socialization: Learning Tradeoffs Under Selection
arXiv cs.AI, 2026-11-16. Alignment as a selection pressure rather than a training objective. The tradeoffs are the contribution.

Weight-space programming, benchmarked and attacked

A subfield arrived complete this morning: a benchmark, an interpretability result, a weight-optimization evaluation, and a security study of the attack surface, all dated today.

Interpretable weight-space programming builds transparent transformers and brain encoding models
arXiv cs.AI, 2026-11-16. The constructive case: programming weights directly can produce models that are legible by construction.
Can Large Language Models Optimize Pretrained Neural Network Weights? A Multi-Phase Empirical Evaluation of Gradient-Free Editing in YOLO26 Object Detection
arXiv cs.AI, 2026-11-16. Gradient-free weight editing, measured on a real detector. The negative results matter as much as the positive.
Binarization as an Attack Surface in Weightless Neural Networks
arXiv cs.AI, 2026-11-16. Adjacent surface, same lesson: the representation choice is a security decision.
A Spectral Framework for Universal Proxy Regularization in Model Merging
arXiv cs.AI, 2026-11-16. Merging as weight-space surgery, with a regularizer that does not depend on the task.
LoRAWeaver: Directional Adapter Routing in LoRA Representation Space
arXiv cs.AI, 2026-11-16. Routing in adapter space. The practical version of the same idea.
BoostLoRA: Growing Effective Rank by Boosting Adapters
arXiv cs.AI, 2026-11-16. Effective rank as the thing being grown, which is a cleaner target than parameter count.
Hidden not Deleted: How Networks Suppress Entangled Features
arXiv cs.AI, 2026-11-16. Suppression is not removal. Relevant to anyone claiming an unlearning result.
Is Abliteration Getting Harder? Not Uniformly: Refusal Removal by Direction-Based Ablation and ARA in Heretic on the Qwen3.8, Gemma 4, Olmo 3.1 and Mistral Small 3.2 LLMs Versus Their Predecessors
arXiv cs.AI, 2026-11-16. Refusal removal across four model families and their predecessors. "Not uniformly" is the finding worth carrying.

One router, forty satellites, and a link shortener

Three independent systems failed at a single point this morning, in three different industries, with no common cause beyond the shape.

Pranksters Redirect Bitly's Links in X Posts for Trump, McDonald's and Burger King
Slashdot, 2026-11-15. Link resolution is infrastructure, and it was reassigned under major brands by people doing it for fun. The benign version of this attack.
Understanding the limits of snapshot-based fault diagnosis in distribution networks
arXiv cs.AI, 2026-11-16. Why diagnosing a grid fault from one instant's measurements fails, separated by error source. The methodological companion to the day's news.
FAA Lifts Nationwide Ground Stop for All Aircraft (UPDATE)
Slashdot, 2026-11-14. Saturday's lift, now with a named cause. Read in sequence with the denial and the router.
America Cancels 4,000 More Flights. FAA Denies Impact of New AI System.
Slashdot, 2026-11-14. The denial we flagged as unsupported on Saturday. On today's evidence it appears to have been correct, issued without the basis.
Transportation secretary blames 'routine' equipment upgrade for grounding flights nationwide
TechCrunch, 2026-11-14. The other official explanation. A router in Arkansas is closer to this account than to the AI one.
PUMA-Former: Uncertainty-Aware Agentic Computation Offloading in MEO-LEO Satellite Networks With Incomplete Information
arXiv cs.AI, 2026-11-16. Offloading decisions in satellite constellations under partial information, published the morning forty of them were destroyed.

Scan

Models and model economics

Claude Fable 5.2 vs. Opus 5.5: Same scores at almost 4x the cost
The New Stack, 2026-11-15. Same benchmark result, four times the bill. The comparison that should drive model selection and usually does not.
Qwen4-Max now available on AI Gateway
Vercel, 2026-11-15. Another frontier Chinese model behind a Western gateway. The routing layer keeps absorbing the geopolitics.
LVMH's Media Division announces strategic partnership with OpenAI
OpenAI, 2026-11-16. Luxury media as a named vertical. Worth noting for who is buying rather than what was built.
More Agents, More Samples, or More Thinking? Untangling Coordination, Ensembling, and Test-Time Compute in Multi-Agent LLMs
arXiv cs.AI, 2026-11-16. Three things usually conflated in multi-agent results, separated. Read before quoting a multi-agent improvement.
Measured from the feed: the unsustainable growth of AI video generation
arXiv cs.AI, 2026-11-16. Growth measured from the consumption side rather than vendor disclosures. "Unsustainable" is defended, not asserted.
Robustness Evaluation of GPU Cloud Pricing Benchmarks under Incomplete Observations: A Controlled Missingness Study
arXiv cs.AI, 2026-11-16. GPU price benchmarks are built on patchy observation. How much the gaps move the ranking.
Superfluid: Local AI Server for Distributed LLM Inference
arXiv cs.AI, 2026-11-16. Local-first inference across a small cluster. Fits the week's other local-stack items.
VelServe: Taming Decode Jitters in LLM Inference with Tail-Calibrated Scheduling
arXiv cs.AI, 2026-11-16. Tail latency in decode, scheduled for explicitly. The number users actually feel.
Do Mixture-of-Experts Routers Benefit from Ensembling?
arXiv cs.AI, 2026-11-16. A narrow, answerable question about MoE routing. Useful negative space.
Advancing Quantization for R&R Models with Microscaling Formats
arXiv cs.AI, 2026-11-16. Microscaling formats applied to retrieval-and-reasoning models.
Evaluating Large Language Models for Scientific Paper Search: Retrieval Success and Inference Latency on Mac Studio and NVIDIA A100
arXiv cs.AI, 2026-11-16. Desktop silicon against a datacenter part on one real task. The latency gap is smaller than the price gap.
ROCKET: Rapid Optimization via Calibration-guided Knapsack Enhanced Truncation for Efficient Model Compression
arXiv cs.AI, 2026-11-16.
Efficiently Aligning Draft Models via Parameter- and Data-Efficient Adaptation
arXiv cs.AI, 2026-11-16.
LACCO: Local Accumulation with Computation–Communication Overlap
arXiv cs.AI, 2026-11-16.

Agent harnesses, memory, and control

ModHarness-RL: Recasting Harnesses as Learnable and Evolving Programs for Agentic RL
arXiv cs.AI, 2026-11-16. The harness as a program that gets optimized rather than a fixed scaffold. The most interesting structural claim in today's batch.
CarryOn: Learning What to Keep in Context for Long-Horizon LLM Agents
arXiv cs.AI, 2026-11-16. Context retention as a learned policy. Directly actionable if you maintain a long-running agent.
When Should LLM Agents Forget? Memory Filtering Under Query-Critical Resolver Errors
arXiv cs.AI, 2026-11-16. The complement to the above, and the harder half. Forgetting the wrong thing is a silent failure.
Can Agents Keep Up with Time? An Empirical Analysis of Temporal Adaptation After a Conversational Hiatus
arXiv cs.AI, 2026-11-16. What happens when an agent resumes after a gap. An underexamined production condition.
Vision-as-a-Tool: Visual Harness for Multimodal Agentic Coding
arXiv cs.AI, 2026-11-16. Vision as a tool call rather than an input channel. Clean separation.
Interpreting Coding Agents Through Their Behavioral Trajectories
arXiv cs.AI, 2026-11-16. Trajectories as the unit of analysis, which is where the pathologies live. Pairs with trace-based tooling.
Can Mechanistic Interpretability Help Explain LLM Agent Failures? A Case Study on Web Element Hallucination
arXiv cs.AI, 2026-11-16. One concrete failure, interrogated mechanistically. Honest about how far the method gets.
Do Verifier Judgments Survive Abstraction? Measuring Verification Transport for LLM Agent Evaluation
arXiv cs.AI, 2026-11-16. Whether a verifier's verdict holds when the task is restated. If it does not, your eval is measuring phrasing.
Generated Evaluation Suites as Diversity-Driven Search for Real Agent Regressions
arXiv cs.AI, 2026-11-16. Generating the eval suite as a search problem. The regression-hunting framing is the useful part.
WebArena-Pro: A Heterogeneous, Multimodal, Reproducible Benchmark for Web Agents
arXiv cs.AI, 2026-11-16. Reproducible is the operative word. The original's drift made cross-paper comparison unsafe.
OutfitterBench: Evaluating LLM Agents on Long-Horizon Resource Allocation and Commitment Coordination
arXiv cs.AI, 2026-11-16. Commitment coordination, which is the capability most agent demos quietly avoid.
Getting started with AG-UI — the agent-user interaction protocol
InfoWorld, 2026-11-16. Another protocol in the agent interop stack. Worth a skim for where it sits relative to MCP.
Teach Genie One your workflows: The ultimate guide for skills in Genie One
Databricks, 2026-11-16. Skills as the configuration surface, now in the data-platform vendors too.
What LLM Agents See and What Users Say
arXiv cs.AI, 2026-11-16. The gap between agent observation and user description. Where most web-agent failures start.
Time-Consistent Evaluation of Hierarchical Code Graphs for SWE Agents
arXiv cs.AI, 2026-11-16. Time consistency in code-graph evaluation, which is how leakage sneaks into SWE benchmarks.
GraphScent: Calibrating Directional Relevance Signals for Bounded-Context Code Localization
arXiv cs.AI, 2026-11-16.
ProdPilot: A Layered Model Context Protocol System for Automated Production Readiness and Deployment of Vibe-Coded Projects
arXiv cs.AI, 2026-11-16.
AutoRL: A Multi-Agent System towards Autonomous Reinforcement Learning
arXiv cs.AI, 2026-11-16.
Benchmarks Don't Talk Back: Learning Through Self-Disagreement
arXiv cs.AI, 2026-11-16.
Imitation Bootstrapped Multi-Agent Reinforcement Learning
arXiv cs.AI, 2026-11-16.
When Cooperation Must Change: Semantic Strategy Unlearning in Multi-Agent Reinforcement Learning
arXiv cs.AI, 2026-11-16.
MASTARS: Multi-Agent Sequential Trajectory Augmentation with Return-Conditioned Subgoals
arXiv cs.AI, 2026-11-16.
Who Should Be on the Panel? Task-Conditioned Persona Selection for LLM Trading Agents
arXiv cs.AI, 2026-11-16.
From Local Prediction Error to Final Outputs: A Theory of Expert Sharing in Multi-Step Reasoning
arXiv cs.AI, 2026-11-16.

Security, injection, and disclosure

Universal XSS in uBlock Origin from website-driven element picker
Lobsters, 2026-11-15. Universal XSS in the blocker itself, triggered by the page it is blocking. The trust inversion is the lesson.
Preventing Cross-Service UDP Loops in QUIC
Lobsters, 2026-11-15. Amplification loops between services that both speak QUIC. Google's writeup, with the mitigation.
Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners
arXiv cs.AI, 2026-11-16. The symbolic interface is the attack surface, not the natural language. A useful corrective.
Safety Side Effects of Preventative Persona Steering: An Empirical Audit
arXiv cs.AI, 2026-11-16. Steering installed for safety reasons with measured safety costs elsewhere. Audit rather than advocacy.
A Push Down the Mountain: Framing Effects and Internal State Interventions on Collusive Behavior in a Long-Horizon LLM Agent
arXiv cs.AI, 2026-11-16. Collusion induced by framing in a long-horizon agent. Relevant to anyone running multiple agents against one market.
Do Machines Learn Deviance From Their Peers? A Randomized Test of Differential Association With AI Agents
arXiv cs.AI, 2026-11-16. Criminology's differential association, tested on agents, randomized. Unusual method and a real result.
Adversarial Tracking: Attacks on Image Sensors for Object Tracking in Autonomous Vehicles
arXiv cs.AI, 2026-11-16. Sensor-level attacks rather than model-level. Harder to patch in software.
PrivacyPR: A Dataset of Real-World Privacy Fixes for Benchmarking Privacy-Aware Code Review
arXiv cs.AI, 2026-11-16. Privacy fixes mined from real pull requests. Ground truth rather than synthetic violations.
From Bytes to Behavior: Request-Level Attribution from Live GPU Hidden States
arXiv cs.AI, 2026-11-16. Attribution read off live GPU state. Both an observability tool and a side channel, depending on who holds it.
Causal AI-Driven Secure mmWave Beam Alignment with Defensive Variational Autoencoders
arXiv cs.AI, 2026-11-16.
Condensing Law of Security Information
arXiv cs.AI, 2026-11-16.
Windows 0-Day Dropped By Disgruntled Researcher Hands Hackers System Rights
Slashdot, 2026-11-14. Disclosure as retaliation. The vulnerability is ordinary; the release path is the story.

Platforms and developer infrastructure

Kernel prepatch 7.4-rc3
LWN, 2026-11-16.
Migrating from GitHub to Codeberg
Lobsters, 2026-11-15. A concrete migration with the friction documented. More useful than the arguments about whether to.
AWS gives developers a free local stand-in for its cloud, built from scratch by AI agents
Pinboard, 2026-11-15. The provenance is the claim: AWS says agents wrote the emulator. A large production artifact to point at.
AWS distills its cloud into a local emulator that runs on your PC
Pinboard, 2026-11-15. The skeptical read of the same launch, and the scope limits.
How AI is warping both developers and developer tools
InfoWorld, 2026-11-16. Both directions of the deformation. Light on evidence, right about the shape.
Rust Leaves Top 10 in TIOBE Index, Making Room for Go
Slashdot, 2026-11-15. TIOBE measures search volume, not use. Included because people will cite it.
A datetime library for Rust
Lobsters, 2026-11-15. Jiff, written up by its author. Datetime libraries are where correctness arguments go to live.
Designing nextest's install scripts
Lobsters, 2026-11-15. Install scripts treated as a design problem with a threat model.
clice v1.0 Released: Welcome to Modern C++
Lobsters, 2026-11-15.
Nix: Using the eval-profiler
Lobsters, 2026-11-15.
beer.drv.drv.drv
Lobsters, 2026-11-15.
This Month in Redox - October 2026
Lobsters, 2026-11-15.
ggml-org/ggml
GitHub, 2026-11-16.
NVIDIA/SkillEvaluator
GitHub, 2026-11-16. Nvidia publishing a skill evaluator the week its governance is in question. Worth a look for the rubric.
elementalsouls/Claude-BugHunter
GitHub, 2026-11-16.
Logback 1.6.7
logback releases, 2026-11-14.

Data platforms and analytics

JSONBench is back, and ClickHouse still leads
ClickHouse, 2026-11-16. Vendor benchmark, vendor wins. The methodology section is the part worth reading.
How Infinitas Learning uses ClickHouse Managed Postgres to process billions of events for 1.6M unique learners
ClickHouse, 2026-11-16.
DuckDB has an answer to "bloated" database messages
The New Stack, 2026-11-15. A wire protocol designed for columnar transfer rather than inherited from row-at-a-time clients.
Please Implement A Quorum Database
Lobsters, 2026-11-15. An argued gap in the storage landscape, from someone who would know.
datasette 1.0a43
Simon Willison, 2026-11-15.
datasette 0.65.6
Simon Willison, 2026-11-15.
datasette-agent 0.6a0
Simon Willison, 2026-11-15. An agent plugin for Datasette. The read-only-SQL-over-a-known-schema shape is the one agents handle well.
datasette-apps 0.4a0
Simon Willison, 2026-11-15.
datasette-maplibre 0.1a0
Simon Willison, 2026-11-15.
Experts Across Schemas: Agent-Guided Mixture-of-Relational-Experts for Cross-Database Generalization
arXiv cs.AI, 2026-11-16.
LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM Agent
arXiv cs.AI, 2026-11-16. A fast path in front of a slow agent. The two-tier design is more interesting than the SQL numbers.
ScrapeGraphAI-100k: Dataset for Schema-Constrained LLM Generation
arXiv cs.AI, 2026-11-16.

Infrastructure, energy, and the buildout

Huawei Set To Challenge Tesla Powerwall with New, 20 kW Home Battery
Slashdot, 2026-11-15.
Bill Gates Has Earned Billions Investing in Climate-Polluting Landfills, Reports ProPublica
Slashdot, 2026-11-15. ProPublica's reporting, with the holdings. The climate-philanthropy framing does not survive it intact.
Helsing backer Plural raises $1.1B to fund Europe's next moonshots
TechCrunch, 2026-11-16. European deep-tech capital at scale, from the fund behind Helsing.
Indian chip startup Bharat Semi lands $24M seed round as it eyes global AI sensor market
TechCrunch, 2026-11-15.
TechCrunch Mobility: The Nvidia backup plan for China's self-driving industry
TechCrunch, 2026-11-15. What the Chinese AV industry does if Nvidia access ends. The contingency planning is already concrete.
Large Language Model-Empowered Intelligent Agent for Adaptive Fast-Charging Control of Lithium-Ion Batteries
arXiv cs.AI, 2026-11-16.
DySENSE: Dynamic Spatiotemporal ENergy Synthesis and Emulation with Satellite Imagery
arXiv cs.AI, 2026-11-16.
AscendGYM: A Cost-Aware Kernel Optimization Benchmark for Real-World Workloads on Ascend NPUs
arXiv cs.AI, 2026-11-16. Kernel optimization on Huawei's accelerator, cost-aware. The non-CUDA ecosystem is accumulating its own benchmarks.
RL\(^{\mathrm{RT}}\): On-Device Reinforcement Learning Runtime
arXiv cs.AI, 2026-11-16.
A\(^3\): Adaptive Audio Agent Framework for Collaborative Edge Computing
arXiv cs.AI, 2026-11-16.

Policy, society, and the human read

AI-Generated Books Confront Publishing Industry With Big Questions
Slashdot, 2026-11-16. Publishing's turn at the question the kernel maintainers and the math journals already faced. Same volume problem, slower institution.
How Will AI Affect Open Source Development?
Slashdot, 2026-11-15.
All AI-generated code should be Open Source
Lobsters, 2026-11-15. A position argued rather than asserted, and the interesting part is the licensing mechanics it would require.
you won't win at copyright
Lobsters, 2026-11-16. The pragmatic counter: copyright is the wrong lever for the thing people want from it.
From Apes to AI
Kenneth Payne, 2026-11-15. Payne on the evolutionary framing of machine intelligence. Strategic studies rather than ML.
This Comic Kills Technofascists
404 Media, 2026-11-15.
AI-aided Decision-Making in Education - Uncertainty, Explainability, and Human Responsibility
arXiv cs.AI, 2026-11-16.
Personalization of Basic and Moral Values Guiding LLM-Controlled Assistant Robots
arXiv cs.AI, 2026-11-16. Personalizing a robot's values is a governance question dressed as a UX feature. The paper notices.
Visa hurdles, high tuition costs creating a 'perfect storm' for lower international student enrollment
WBUR, 2026-11-16. The research-pipeline story. Enrollment declines show up in lab output three years later.
Flaming Hydra's new Kickstarter campaign rethinks e-book ownership and digital lending
TechCrunch, 2026-11-15.
Apple launches new app to study how people use AI for health
TechCrunch, 2026-11-15. Apple collecting observational data on consumer health-AI use. The study design will matter more than the app.
Fake News Detectors are Biased against Texts Generated by Large Language Models
arXiv cs.AI, 2026-11-16.
LEGAL-UQA: Testing Constitutional Question-Answering Systems for Urdu Speakers with a Parallel English-Urdu Benchmark
arXiv cs.AI, 2026-11-16.

Robotics and embodiment

Envisioning Agentic Robotics: A New Era of Robotic Systems
arXiv cs.AI, 2026-11-16. The position paper for the term. Worth reading to know what people will mean by "agentic robotics" next quarter.
Vision-Language-Action Models: A Life-Cycle Survey of Design, Training and Deployment
arXiv cs.AI, 2026-11-16. A survey organized by lifecycle rather than architecture, which makes it usable as a reference.
X-Nav Technical Report: Real-World Long-Horizon Indoor-to-Outdoor Navigation on a Humanoid Robot
arXiv cs.AI, 2026-11-16. Indoor to outdoor on real hardware. The transition is the hard part and they report it.
VLAN: Vision-Language Accessible Navigation
arXiv cs.AI, 2026-11-16.
Racing the Clock: Teaching VLNs Budget-Conditioned Object Search
arXiv cs.AI, 2026-11-16.
D-VLA: A High-Concurrency Distributed Asynchronous Reinforcement Learning Framework for Vision-Language-Action Models
arXiv cs.AI, 2026-11-16.
E0: Expressive Fine-Grained Discrete Action Prediction for Vision-Language-Action Models via Tweedie Discrete Diffusion
arXiv cs.AI, 2026-11-16.
Recursive Belief Vision Language Action Models
arXiv cs.AI, 2026-11-16.
Vision-language models lag human performance on physical dynamics and intent reasoning
arXiv cs.AI, 2026-11-16. A gap worth keeping in mind before putting a VLM in a control loop.
TelePhysica: Agentic Reconstruction of Physically Interactive Scenes
arXiv cs.AI, 2026-11-16.
Imitation Policies Learn a Proprioceptive Anchor: A Predictive Law for Perception Noise and Its Closed Loop Cost
arXiv cs.AI, 2026-11-16. A predictive law rather than an observation. Rare and useful in imitation learning.
Courteous Anticipation: Improving Long-Lived Task Planning in Persistent Shared Environments
arXiv cs.AI, 2026-11-16.
Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent
arXiv cs.AI, 2026-11-16.

Evaluation, uncertainty, and interpretability

Does the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device LLMs
arXiv cs.AI, 2026-11-16. Detector rankings established on frontier models, re-run on quantized on-device ones. They do not hold.
Screen Before You Interpret: A Portable Validity Screen for Benchmark-Based LLM Confidence Signals
arXiv cs.AI, 2026-11-16. A validity screen you can apply before trusting a confidence number. Cheap and immediately usable.
Black-Box Uncertainty Quantification for Large Language Models via Ensemble-of-Ensembles
arXiv cs.AI, 2026-11-16.
Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume
arXiv cs.AI, 2026-11-16.
DECODE: Diagnosing Interpretability Failures in Automated Interpretation Process
arXiv cs.AI, 2026-11-16. Interpretability pipelines fail silently. A diagnosis procedure for the procedure.
Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering
arXiv cs.AI, 2026-11-16.
Logically Consistent Accurate Forecasts
arXiv cs.AI, 2026-11-16.
What Balanced Accuracy Cannot Identify in Neural-Signal Decoding under Test-Time Adaptation
arXiv cs.AI, 2026-11-16.
Evaluating Chain-of-Thought Reasoning through Reusability and Verifiability
arXiv cs.AI, 2026-11-16.
JFinBench: A Continuously Refreshable Japanese Financial Benchmark from the Latest Certification Exams
arXiv cs.AI, 2026-11-16. Continuously refreshable is the design feature. Contamination has a shelf life this way.
GDP.xlsx: Benchmarking Frontier Models on Professional Workbooks
arXiv cs.AI, 2026-11-16. Spreadsheets as the eval surface. Closer to most white-collar work than any coding benchmark.
BigCodeArena: Unveiling More Reliable Human Preferences in Code Generation via Execution
arXiv cs.AI, 2026-11-16.
BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions
arXiv cs.AI, 2026-11-16.
HARDTESTGEN: A High-Quality RL Verifier Generation Pipeline for LLM Algorithmic Coding
arXiv cs.AI, 2026-11-16.
LECAT: Measuring LLM Category Learning from Pretraining Data
arXiv cs.AI, 2026-11-16.
Knowledge After Collapse: Context and Wording Shape Factual Accessibility in Language Models
arXiv cs.AI, 2026-11-16.

Clinical and life sciences

When to Trust Simulated Patient Responses: Support-Gated World Models for Vasopressor Decision Support
arXiv cs.AI, 2026-11-16. Gating on support rather than confidence, in a setting where the difference kills people.
Confident and Wrong: Structure Confidence Is a Poor Selection Filter for Protein Function
arXiv cs.AI, 2026-11-16. Structure confidence used as a function filter, shown not to work. A widely used shortcut, measured.
Causally Constrained and Uncertainty-Aware Cardiopulmonary Risk Prediction
arXiv cs.AI, 2026-11-16.
TotalSegmentator Aorta Report: Automated Aortic Diameter Measurement at 11 Standardized Landmarks on CT Images
arXiv cs.AI, 2026-11-16.
Interactive deep-learning based interface for brain tumour segmentation
arXiv cs.AI, 2026-11-16.
AG-JEPA: Anatomically Guided Joint Embedding Predictive Architecture for Neurodegenerative Representation Learning
arXiv cs.AI, 2026-11-16.
Towards Explainable Sleep Stage Classification using Single-Channel Electroencephalogram
arXiv cs.AI, 2026-11-16.
Cross-Linguistic Acoustic Biomarkers of Cognitive Impairment Across Interactive and Single-Speaker Speech Tasks
arXiv cs.AI, 2026-11-16.
Generative machine learning for protein design
arXiv cs.AI, 2026-11-16.
ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
arXiv cs.AI, 2026-11-16.
AbFlowNet: Optimizing Antibody-Antigen Binding Energy via Diffusion-GFlowNet Fusion
arXiv cs.AI, 2026-11-16.
More Genes, Faster Inference: Efficient Large-Panel Spatial Transcriptomics Prediction from Histology
arXiv cs.AI, 2026-11-16.
ISLA: Ensemble learning and domain adaptation for generalizable acute ischemic stroke lesion segmentation in diffusion MRI with clinical validation
arXiv cs.AI, 2026-11-16.
AI-Based Multi-View Echocardiographic Image Synthesis using Dynamic Shape Control
arXiv cs.AI, 2026-11-16.
Tabular Population Priors as Structured Retrieval for EHR Medication Set Prediction
arXiv cs.AI, 2026-11-16.
PhysDox: Benchmarking LLMs on Physical Feasibility Auditing of Physiological Sensing Protocols
arXiv cs.AI, 2026-11-16. Auditing whether a sensing protocol is physically possible. A good example of an LLM task with a checkable answer.

Engineering and the craft

Memory management in 40 major C projects, a survey
Lobsters, 2026-11-15. Forty real codebases, surveyed. The taxonomy of what people actually do beats any single project's convention.
Some problems of recursive descent parsers (2009)
Lobsters, 2026-11-15. Seventeen years old and still the clearest statement of the failure modes.
Paint in the memory of other programs
Lobsters, 2026-11-15.
A NES emulator written in Odin
Lobsters, 2026-11-15.
Announcing Hegel-clj
Aphyr, 2026-11-11. Kingsbury on a new Clojure library. Worth the click for the writeup regardless of whether you need the library.
pumice 0.2.0
Planet Clojure, 2026-11-16.
2026 Board Election Results and Annual Meeting
Planet Clojure, 2026-11-15. Clojurists Together's governance, which funds a meaningful share of the ecosystem's maintenance.
Jonquil: a pragmatic editor for the Lisps you use
Pinboard, 2026-11-16.
20 Years of Pyrolog
Pinboard, 2026-11-16. Twenty years of a Prolog on PyPy. A retrospective on RPython as a language-implementation substrate.
Emacs is my browser
Pinboard, 2026-11-15.
curry: a modern Emacs mode for Haskell
OCaml.org, 2026-11-15.
waforth: Forth interpreter/compiler for and in WebAssembly
Pinboard, 2026-11-15.
waforth on npm
Pinboard, 2026-11-15.
WAForth VS Code extension
Pinboard, 2026-11-15.
All Code Is Slop (Except Mine) — Lauren Tan
Pinboard, 2026-11-16.
ARCHER: From Rationale Imitation to Executable Reasoning in Small Language Models
arXiv cs.AI, 2026-11-16.
Interactive Program Induction with Language Model Priors
arXiv cs.AI, 2026-11-16.
Hybrid Mechanical and Agentic Specification Inference
arXiv cs.AI, 2026-11-16. Mechanical inference where it works, agents where it does not. The division of labor is the contribution.
Active Learning of Temporal Specifications through Distinguishing Queries
arXiv cs.AI, 2026-11-16.

Research and math

From Pen-and-Paper to Machine-Checked Proofs: Lean for System Identification
arXiv cs.AI, 2026-11-16. Lean applied to control theory's proofs. The domain is unusual and the mechanization is real.
Certified Improvements to Best-Known Thomson-Problem Configurations with a Search Pipeline Built with Large Language Model Agents
arXiv cs.AI, 2026-11-16. New best-known configurations, certified. An agent result with a verifiable artifact at the end, which is the only kind that counts.
Verifiable pXRD: Hill-climbing pXRD structure determination with LLM agents
arXiv cs.AI, 2026-11-16. Same pattern in crystallography: the agent searches, the physics verifies.
Agent-written heuristics versus commercial solvers on unsolved job-shop benchmark instances
arXiv cs.AI, 2026-11-16. Head to head against Gurobi-class solvers on open instances. The honest comparison.
Agentic AI for Discovering Fault-Tolerant Quantum Circuits
arXiv cs.AI, 2026-11-16.
AI for Parameter Selection in Sample-based Quantum Diagonalization
arXiv cs.AI, 2026-11-16.
A step toward defining the Standard Model on a lattice
OpenAI, 2026-11-15. Lattice field theory, from the lab's new science track. Judge it on the physics, not the press release.
Our new blog for science
OpenAI, 2026-11-15. A separate publication channel for scientific results. Note what moving it off the main blog signals.
Introducing MathHelm: A Place to discuss AI-generated papers
Terence Tao, 2026-11-14. Still the week's most constructive institutional response. A venue instead of a ban.
The Lean 4 Theorem Prover and Programming Language
Pinboard, 2026-11-16.
First Proof Project
Pinboard, 2026-11-16.
Two-Clustering Regime of Token Dynamics in Causal Attention
arXiv cs.AI, 2026-11-16.
Probabilities of Causation in Nondeterministic Causal Models
arXiv cs.AI, 2026-11-16.
Tracing Actual Causes with Counterfactual Witness Maps
arXiv cs.AI, 2026-11-16.
A Synthetic Benchmark for Target-Centric Causal Discovery in Industrial Processes
arXiv cs.AI, 2026-11-16.
Time series causal discovery with variable lags
arXiv cs.AI, 2026-11-16.
Astrocyte Biocomputing: Astrocytic Contributions to Neural Network Computation
arXiv cs.AI, 2026-11-16.
An In-Silico Audit of Topographic Organization in a Multimodal Brain-Response Model
arXiv cs.AI, 2026-11-16.
A Variational Latent Equilibrium for Learning in Neuronal Circuits
arXiv cs.AI, 2026-11-16.
The Transformer Paradigm: Dynamic Processing and a New Principle of Inference
arXiv cs.AI, 2026-11-16.
An AI-enabled high-resolution chlorophyll-a forecasting model that outperforms the operational global biogeochemical system
arXiv cs.AI, 2026-11-16.
Tracing Uncertainties of Extreme El Ni\~{n}o Prediction to Mesoscale Ocean Currents
arXiv cs.AI, 2026-11-16.
Elliptic Sampling: Optimal Interpolant Schedules for Gaussians and Beyond
arXiv cs.AI, 2026-11-16.
Benign Overfitting for Regression with Trained Two-Layer ReLU Networks
arXiv cs.AI, 2026-11-16.

Hardware and radio

Purism Announces 'Librem 5 v2' Linux Phone with RISC-V SoC
Slashdot, 2026-11-16. A RISC-V phone SoC in a shipping product. The interesting number will be how much of the stack is actually open.
Siri Glitches Ruin Reviewer's Test of Apple's New 'HomeView' Smart Speaker
Slashdot, 2026-11-15.
VapourDeck is the 486 Handheld You Didn't Know You Needed
Hackaday, 2026-11-16.
It's A NAS, For Sony Mavica Disks
Hackaday, 2026-11-16.
Follow The Grid With An Arduino Uno
Hackaday, 2026-11-16.
Loop Antennas for Transmitting
Hackaday, 2026-11-15.
Printing Smooth Upper Surfaces with a Non-Planar Slicer
Hackaday, 2026-11-15.
An Orbital Sanders Crash Course and the Threat of HAVS
Hackaday, 2026-11-15.
Making CO₂ Sensors Fail Visibly, Not Silently
Hackaday, 2026-11-14. Fail-visible as a design requirement for a sensor. The right instinct, and the FAA story's lesson in miniature.
Toward Interference-Agnostic RF Source Separation
arXiv cs.AI, 2026-11-16.
Typing Matters: Muscle, Language and Typing Behaviour in Live Keystroke Decoding from sEMG
arXiv cs.AI, 2026-11-16.
Optimizing Drilling Trajectories in Multi-Material Structures Using Quantum Annealers
arXiv cs.AI, 2026-11-16.

The crawler zoo

Fifteen new user agents overnight, and the first weekly report with a request count attached: 61,090 in week 46. Note Grokcelot-recon, two directory-check bots, and the EFF showing up as a crawler.

Zoo Report, Week 46, 2026: 61,090 bot requests
The Crawler Zoo, 2026-11-16. The first number big enough to reason about. Worth comparing against your own logs.
Grokcelot-recon
The Crawler Zoo, 2026-11-16.
EFF
The Crawler Zoo, 2026-11-16.
EternalAI-DirectoryCheck
The Crawler Zoo, 2026-11-16.
LocalWebReference
The Crawler Zoo, 2026-11-16.
AdmaticMarketCrawler
The Crawler Zoo, 2026-11-16.
NavCrawl
The Crawler Zoo, 2026-11-16.
LargerioBot
The Crawler Zoo, 2026-11-16.
OusanBot
The Crawler Zoo, 2026-11-16.
key-master
The Crawler Zoo, 2026-11-16.
cis5550-crawler
The Crawler Zoo, 2026-11-16.
WhyNoIPv6Bot
The Crawler Zoo, 2026-11-16.
Kaupr
The Crawler Zoo, 2026-11-16.
MglennP-RadialScout
The Crawler Zoo, 2026-11-16.
my-agent
The Crawler Zoo, 2026-11-16.
remotejobsiq-bot
The Crawler Zoo, 2026-11-15.
MadeOnSol-Metadata
The Crawler Zoo, 2026-11-15.

Boston

Terence Tao: Six Math Essentials, Harvard Science Center
wal.sh, 2026-11-16. Tonight. Same week he opened MathHelm.
AI and Society: The Future of Math
MIT, 2026-11-16. The institutional version of the same conversation, two miles down the river.
Barbara Grosz Distinguished Lecture in AI and Society
wal.sh, 2026-11-16.
2026 Dresselhaus Lecture — Connecting nano to macro: Thermal materials research inspired by Millie Dresselhaus
MIT, 2026-11-16.
Harvard SEAS PhD Virtual Information Session
Harvard SEAS, 2026-11-16.
MS Program Webinar – Department of Chemical Engineering
Northeastern, 2026-11-16.
Fusion Book Club Meeting
Northeastern, 2026-11-16.
English Programs Lightning Book Talks
Northeastern, 2026-11-16.
Harvard Book Store author talks
wal.sh, 2026-11-16.
Right whale population dropped in 2025 after a few years of growth
WBUR, 2026-11-16. Reversal after several years of recovery. The survey methodology is in the piece.

Tail

GTA 6's Upcoming Launch Celebrated by Wendy's, YouTube, the City of Miami, and 20% of America's Workforce
Slashdot, 2026-11-15.
SNL gathers all the tech tycoons for a lesson in dating and spycraft
TechCrunch, 2026-11-15.
The tech products we enthusiastically recommend this holiday season
TechCrunch, 2026-11-15.
Hackaday Links: November 15, 2026
Hackaday, 2026-11-16.
Lions' 35-10 rout over Patriots after Maye injured in Munich
WBUR, 2026-11-15.
MeReader: An Offline, Privacy-Preserving, Progress-Aware AI Assistant for Narrative eBook Reading
arXiv cs.AI, 2026-11-16.
OmniScore-Bench: Evaluating Vision-Language Models on Musical Score Reading
arXiv cs.AI, 2026-11-16.
SceneAgent: A Multi-Agent Framework for Portable Game Scene Generation
arXiv cs.AI, 2026-11-16.
OCTAVE: A Capability-Resolved Benchmark for End-to-End HTML5 Game Generation
arXiv cs.AI, 2026-11-16.
TReC: Evaluating Move-Induced State Changes in Instruction-Conditioned Go
arXiv cs.AI, 2026-11-16.
Activation Steering the Nan-Shan in Typhoon
arXiv cs.AI, 2026-11-16.
Photogrammetric 3D Model-Based Animal Behavior Classification
arXiv cs.AI, 2026-11-16.
Creative Product Design under Brand Image Constraints
arXiv cs.AI, 2026-11-16.
TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation
arXiv cs.AI, 2026-11-16.
CoTTAGE: An Explainable Judge for Text-to-Audio Generation
arXiv cs.AI, 2026-11-16.
Structure and Interpretation of Computer Programs
Pinboard, 2026-11-16.
ElixirConf EU 2027, 22-23 April, Lisbon
Pinboard, 2026-11-16.
50th Anniversary of 'Star Wars' Celebrated with iMax 70mm Re-Release
Slashdot, 2026-11-15.
Solving the fox game
Lobsters, 2026-11-14.

Feed silences and verification

One hundred eighteen feeds polled this morning: 47 returned content, 70 answered 304, one errored.

A correction to Saturday's brief. We reported Hacker News as answering 304 on twenty consecutive polls and called it a stale conditional validator in our own cache. The logs do not support that. Hacker News has returned HTTP 419 on twenty-one consecutive polls, from 2026-10-27 through this morning, and has never once returned 304 in any log we hold. 419 is not a conditional-request response; it is a rejection. Our conditional-fetch path classifies any response without changed content as :not-modified and logs the real status parenthetically, so a bot challenge has been recorded as a cache hit for three weeks. That is our bug, not HN's behavior, and the diagnosis we published was wrong about which side the fault sits on. HN last returned content on 2026-10-26 and contributes nothing to this brief.

arXiv cs.AI returned a 206-paper batch this morning, all dated today. The weekend gap closed exactly as the announcement calendar predicts. C-004 remains unrefuted.

www.bornholt.dev is the one error. It has now failed to resolve on 174 consecutive polls since 2026-05-26 and has never once succeeded from this host. It is excluded from the publishing count rather than reported as a silence, because there is no evidence it was ever live from here.

One hundred twenty-six items in the window carry source = 'wal.sh research' in the warehouse: 121 Slashdot URLs across thirteen subdomains, one Hackaday post, one Netflix Tech Blog post, one All Things Distributed post, one blog.sesse.net post, and one Apache Slashdot post. That is not a source, it is us. Our own Atom feed is crawled back in, url is the primary key, and re-ingestion overwrites the original publisher on every item we have previously curated. Those items are attributed above by their URL domain rather than by the stored source field. The count has gone 111 (11-13) → 113 (11-14) → 125 (11-15) → 126 (11-16). The fix belongs in the store, not in the brief.

Every link above was checked through the compliance gate before publication: 285 distinct URLs, 279 answered 200 and none answered 404. Four returned 403: OpenAI three times and npmjs.com once. These are bot challenges against our identity, not broken pages. One URL, the VS Code Marketplace listing for WAForth, answered 404 to HEAD and 200 to GET with a 74 KB page; it is published on the GET. Two URLs our own robots check declined to fetch, blog.sesse.net and theregister.com, are published on feed provenance: robots governs what we retrieve, not what we cite.

Honoring arXiv's Crawl-delay: 15 across 163 abstract URLs took 41 minutes of the verification pass on its own. Noted because it is the cost of the policy, not a reason to skip it.

build: 2026-11-16 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.3) | feeds: 118 polled, 71 publishing | items-considered: 5,850 (14d, incl. 3,621 arXiv) | warehouse: 71,678 items | published: 284 (285 URLs verified, 279 × 200, 0 × 404)