Morning Brief: Thursday, November 5

Seventy-six feeds reporting. Two weeks. 5,385 items reduced to what follows. (what we track, how we crawl, subscribe)

Thursday is the day the scaffolding stopped being an implementation detail. Several independent papers and one product launch treat the harness around a fixed model - its tools, its telemetry, its authority gates - as the thing that decides behavior, which makes a model evaluation something less than a deployment evaluation.

The sharpest version is a banking-conduct study that holds the model constant, varies only the scaffolding, and finds compliance moves - and moves by different amounts for different models, so you cannot even rank harnesses once and reuse the ranking. Read alongside it: a paper arguing agent security is a systems property rather than a model property, one proposing evidence-gated authority over when a human may overrule an agent, one on continuous authorization conditioned on assurance, and one insisting that attributing an agent failure to its history requires actually executing the counterfactual rather than asking a judge. Cloudflare shipped the industry half a day earlier: visibility and per-site terms for personal agents, which is the harness problem viewed from the server's side of the connection. The MCP supply chain gets its own refutation - signatures on a toolchain do not establish trust in it.

The other thread is not research. Tuesday's midterms resolved into a Democratic Congress, and the sector's exposure showed up immediately: 404 Media reports voters rejecting AI data centers by wide margins, MIT Technology Review calls the Texas buildout possibly kaput, and the AI PACs that spent heavily on Congress got an itemized reckoning. Energy permitting is now the live legislative question, and OpenAI published an infrastructure-responsibility standard into that weather the same week it acquired Pierre Computer and began positioning for 2028.

Top (5-7 min)

The Harness Is the Model: Agent Scaffolding Changes a Fixed LLM's Compliance with Banking Conduct Rules, Unevenly Across Models: arXiv cs.AI, 2026-11-05. The cleanest statement of today's thread, and it is an experiment rather than an argument. Hold the weights fixed, change the scaffolding, watch conduct compliance move unevenly.
Agent Security is a Systems Problem: arXiv cs.AI, 2026-11-05. The position paper the above result implies. If the harness decides behavior, securing the model is securing the wrong object.
Who May Overrule the Agent? Evidence-Gated Authority in LLM Review Institutions: arXiv cs.AI, 2026-11-05. Override authority as an institutional design question with an evidence precondition attached. The most concrete proposal here for what human-in-the-loop should mean operationally.
See the agents, set the terms: new visibility and control for personal agents: Cloudflare, 2026-11-04. The server-side view of the same problem: who is this agent, acting for whom, under what terms. Directly relevant to anyone operating a crawler or a site that receives one.
Attributing Agent Failures to Their History Needs an Executed Counterfactual: Toward a Reporting Standard: arXiv cs.AI, 2026-11-05. Postmortems on agent runs are mostly storytelling unless you rerun the branch. Proposes a reporting standard, which is the part worth stealing.
Signed, Sealed, Leaked: Trust Beyond Signature for MCP Toolchains: arXiv cs.AI, 2026-11-05. Signature verification on an MCP toolchain tells you provenance, not intent. The gap is where the next round of agent supply-chain incidents will come from.
American Voters Overwhelmingly Reject AI Data Centers: 404 Media, 2026-11-04. Polling plus ballot results, reported together. The datacenter buildout now has an electoral constraint it did not have last week.
Texas was the new epicenter of the AI data center buildout. That might be kaput.: MIT Technology Review, 2026-11-04. The specific regional consequence. Worth reading for the mechanism - interconnect queues and local politics - not the headline.
Meta acquires AI coding startup Conductor: TechCrunch, 2026-11-04. Harness consolidation as an acquisition strategy. Conductor is multi-agent orchestration, which is the layer everyone has decided is the moat.

Themes

The harness is the thing being evaluated

The harness papers have stopped being about convenience and started being about control surfaces. HarnessFlow searches the scaffolding space with MCTS; Agent Harness Awareness turns verified telemetry into operational control; AgentTrustCI puts policy, grounding and trajectory tests at build time; and two more decide when the agent should think at all and how many samples a turn deserves. The common move is to treat the harness as code under test rather than as glue, which is the only framing under which the banking-conduct result is actionable.

HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents: arXiv cs.AI, 2026-11-05. Search over scaffolding configurations rather than hand-tuning. Pairs with the conduct paper: if the harness moves behavior, the search is also a search over behavior.
Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents: arXiv cs.AI, 2026-11-05. Telemetry as a control input, not a dashboard. The verified qualifier is doing the work - unverified self-report is what most harnesses currently act on.
AgentTrustCI: Build-Time Policy, Grounding, and Trajectory Tests for Java LLM Agents: arXiv cs.AI, 2026-11-05. Agent tests in CI, for a language nobody writes agent papers about. The build-time placement is the interesting choice.
Agents, Think Only When it Helps: arXiv cs.AI, 2026-11-05. Gating deliberation on expected benefit. Most of the reasoning budget in a deployed agent is spent on steps that did not need it.
On Turn-Level Best-of-N Sampling for Scaling LLM Agents: arXiv cs.AI, 2026-11-05. Where to spend sampling in a multi-turn trajectory. A harness parameter that is usually set once and never revisited.
Tools Need Not Speak in Text: Latent Context Compression for Agents: arXiv cs.AI, 2026-11-05. Tool results as latent state rather than serialized prose. Saves context and loses auditability, which is a real tradeoff and not a free win.
Uncertainty-Aware Resource Orchestration for Agentic AI Systems: arXiv cs.AI, 2026-11-05. Scheduling under model uncertainty. The operations problem the harness framing creates once you have more than one agent.
Graph-Mediated Large Language Model Systems in the Agentic Era: Reasoning, Knowledge, Retrieval, Execution, and Coordination: arXiv cs.AI, 2026-11-05. A survey, so read it for the taxonomy. Useful for seeing how much of the field has converged on graph structure as the mediating representation.
Render just removed its AI sandbox time limit - and Vercel still has a 24-hour cap: The New Stack, 2026-11-04. Sandbox lifetime is a harness parameter with a vendor attached. Long-running agents are a different product than long-running functions.

Who gets to overrule the agent

The authority papers are the half of the harness thread that cannot be solved by better engineering. Continuous authorization conditioned on assurance, consultation value computed relative to the executor, grassroots contracts as a risk-corralling mechanism, sybil resistance in an agent marketplace: each one is answering who may stop this, on what evidence. The judging papers belong here too, because an agent-as-judge with confirmation bias is an authority that cannot be appealed.

Assurance-Conditioned Continuous Authorization for Autonomous AI Agents: arXiv cs.AI, 2026-11-05. Authorization that expires unless assurance evidence keeps arriving. The right shape for a long-lived agent, and expensive to implement honestly.
Learning When Human Decisions Matter: Executor-Relative Consultation Value for LLM Agents: arXiv cs.AI, 2026-11-05. Computes when asking a human is worth the interruption, relative to who is executing. Escalation policy as a learned quantity rather than a config flag.
Corralling AI Risks with Grassroots Contracts: arXiv cs.AI, 2026-11-05. Bottom-up contractual constraint instead of regulation. Discount the optimism; the mechanism is worth understanding because it needs no central authority.
Anduin: a sybil-resistant agentic marketplace reputation system: arXiv cs.AI, 2026-11-05. Reputation for agents that can be cheaply recreated. Sybil resistance is the load-bearing requirement and the one most agent marketplaces skip.
Grounded but Not Immune: Confirmation Bias in Agent-as-a-judge: arXiv cs.AI, 2026-11-05. Grounding the judge in evidence does not remove the bias. Read against last week's judge-replay results.
AGORA: Reliability-Aware Aggregation of LLM Judges via Item-Response Theory: arXiv cs.AI, 2026-11-05. Borrowing psychometrics to weight unreliable judges. The honest response to judge variance, as opposed to adding more judges.
Signing in to your company's plugins one by one is annoying: Anthropic's fix puts Claude in the identity business: The New Stack, 2026-11-04. Plugin auth consolidated behind the assistant. Convenient, and it makes the assistant an identity provider, which is a different kind of dependency.
Agile Deliberation: How AI helps humans define the gray areas: Google Research, 2026-11-04. Policy elicitation rather than policy enforcement. The upstream problem: the authority papers assume a rule exists to gate on.
Desiderata for the Theory of Agents: Alignment Forum, 2026-11-04. A list of what a theory of agents would have to deliver. Useful as a check on how much of the above is engineering standing in for theory.

Scan

AI and agents

Route to Remember: Adaptive Memory Selection for LLM Agents: arXiv cs.AI, 2026-11-05. Retrieval policy over agent memory. The selection step is where most memory systems quietly fail.
SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory: arXiv cs.AI, 2026-11-05. Summarization as the memory primitive, deliberately plain. Worth it as a baseline against more elaborate schemes.
FlyHash for LLM Agent Memory: An Empirical Validation and a New Application: arXiv cs.AI, 2026-11-05. Fruit-fly olfactory hashing applied to agent recall. Odd provenance, cheap to implement, and the empirical half is the point.
Reglos: A Self-Evolving Personal Memory Agent with Recursive Self-Improvement: arXiv cs.AI, 2026-11-05. Personal memory as the substrate for a self-improvement loop. The narrow domain is what makes the loop evaluable at all.
Towards Robust Long-Lived Agents: Benchmarking Experience Lifecycles in Evolving Environments: arXiv cs.AI, 2026-11-05. Memory that has to expire. Experience lifecycle is the right frame and almost nobody benchmarks it.
Do agents need to sleep?: arXiv cs.AI, 2026-11-05. Offline consolidation as a question rather than a metaphor. Short, speculative, and asks something the memory papers above assume away.
Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification: arXiv cs.AI, 2026-11-05. Per-step verification for search trajectories. Failure localization is the prerequisite for the counterfactual-attribution standard.
Task-Efficient Agent Evaluation Using Cached Traces: arXiv cs.AI, 2026-11-05. Reusing trajectories to cut evaluation cost. Practical, and it constrains what counterfactuals you can still run.
Zero-Shot Coordination among LLM Agents: arXiv cs.AI, 2026-11-05. Agents coordinating without a shared protocol or prior play. The multi-agent assumption that most deployments actually face.
MA-Mesh: Resilient Multi-Agent Collaboration via Role-based Self-Healing: arXiv cs.AI, 2026-11-05. Fault tolerance in a multi-agent mesh via role reassignment. Distributed-systems thinking applied to agent topologies.
EnzymeRSI: A dual-loop recursive self-improvement system for enzyme design: arXiv cs.AI, 2026-11-05. Recursive self-improvement in a domain with a wet-lab ground truth. The dual loop separates proposal from verification, which is the part that makes it checkable.
FrontierPhysics: Evaluating Agents for End-to-End Physics Research: arXiv cs.AI, 2026-11-05. End-to-end research evaluation rather than problem-set scoring. Expect low scores; that is the useful signal.
BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories: arXiv cs.AI, 2026-11-05. Agents driving real chemistry hardware. Expert curation matters here because the failure modes are physical.
LifeVal: Evaluating AI Model Performance on Personalized Agentic Consumer Tasks: arXiv cs.AI, 2026-11-05. Consumer tasks with personalization, which is where agent claims are loudest and measurement is thinnest.
PAYWORTHY: Can LLMs Predict Which Freelance Jobs an AI Agent Will Complete?: arXiv cs.AI, 2026-11-05. Predicting agent success on priced labor. A more honest economic signal than task benchmarks because the price is observed.
ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models: arXiv cs.AI, 2026-11-05. ConceptARC run at a scale that lets the concept axes separate.
Transforming Integer Grids to Spatial Intuition: Perception, Not Planning, Governs ARC-AGI-3: arXiv cs.AI, 2026-11-05. Argues the bottleneck on ARC-AGI-3 is the input representation, not search. If right, a lot of planning work is aimed at the wrong half.
Rules, Traces and Routing Statistics: Training and Calibration Data for Open Models on ARC-AGI-2: arXiv cs.AI, 2026-11-05. Released data rather than a new score, which is the more reusable contribution.
Can Large Language Models Review the Scientific Literature? A Test Against a 102-Article Benchmark: arXiv cs.AI, 2026-11-05. A small, specific benchmark with a checkable answer key. Relevant to anyone automating the step this brief does by hand.
A Definition of Catastrophic Performance: arXiv cs.AI, 2026-11-05. Definitional work on what counts as catastrophic rather than merely bad. Needed before any threshold-based guardrail means anything.
New Research Tests Whether AI Models Know Their Own Thoughts: Quanta Magazine, 2026-11-04. Introspection tested rather than asserted. Good popular writeup of a literature that is mostly negative results.
How big is the open model economy today?: Interconnects, 2026-11-04. An attempt at a number, with the measurement difficulty foregrounded. Read it for the caveats as much as the estimate.
[AINews] It's Swapping Season: Latent Space, 2026-11-05. Daily roundup. Useful for the model-swap churn this brief does not track.
A Closer Look at Phonon TTS: Launching High-Quality On-Device Speech That Tops the Benchmarks: Hugging Face Blog, 2026-11-05. On-device TTS with benchmark claims attached. The on-device part is what makes it interesting for anyone avoiding a speech API.
The Enterprise Case for Open-Weight Models on Databricks: Databricks, 2026-11-04. Vendor argument, so discount accordingly. The cost and data-locality reasoning is still the clearest version of the open-weight case.
Anthropic found a way to prove which AI model answered you - but there's a costly catch: The New Stack, 2026-11-04. Cryptographic model attestation with a stated price. Relevant to the identity thread above: proving which model answered is a prerequisite for holding anyone to terms.
Almost no one uses AI agents to book their Uber, company says: TechCrunch, 2026-11-04. A rare usage data point from the demand side, and it is small. Useful counterweight to agent-commerce projections.

Engineering and tools

A new Go API for JSON: Lobsters, 2026-11-04. The v2 encoding/json design, from the team. Long, and the migration section is the part that will matter.
Zig 0.17.1 Released: Lobsters, 2026-11-04. Point release notes.
Shlemiel the Painter Is Back: How Zig's libc Rewrite Made SQLite Slower: Lobsters, 2026-11-04. A quadratic reintroduced in a libc rewrite, found via SQLite. Good debugging narrative and a reminder about what rewrites inherit.
C++26 Disapproved: Lobsters, 2026-11-04. The DIS ballot failed. Process reporting from someone close to it.
How to speed up the Rust compiler in November 2026: Lobsters, 2026-11-04. The latest in a long series. The value is the methodology, which transfers to compilers that are not rustc.
The State of Python Profiling in 2026: Lobsters, 2026-11-05. Survey of the current tooling, by someone who writes one of them.
A Look at Datastar: Lobsters, 2026-11-05. Hypermedia-driven frontend, assessed rather than advocated.
Introducing Adaptive Dark Mode: Lobsters, 2026-11-04. WebKit darkening pages that never opted in. Read the opt-out before this ships to your users.
Declarative Dragging: Lobsters, 2026-11-04. Drag and drop as a declarative primitive instead of an event maze.
Burnt TOAST: Lobsters, 2026-11-04. Postgres TOAST pathologies, with the storage detail that explains the surprise.
Lean on the JVM: proved functions for Java and Clojure: Planet Clojure, 2026-11-04. Lean-proved functions callable from JVM code. Narrow, and the clearest version of the proof-carrying-library idea on this stack.
datasette-llm 0.1a10: Simon Willison, 2026-11-04. Release note for the LLM plugin.
datasette-apps 0.3a0: Simon Willison, 2026-11-04. Release note. The apps direction is the more consequential of the two.
Fencing a Postgres primary that might still be alive: ClickHouse, 2026-11-04. Split-brain fencing, written as an engineering problem rather than a product pitch. The best technical post in today's vendor set.
Introducing Apache Spark 4.3, now available in Databricks Runtime 19: Databricks, 2026-11-05. Spark release, vendor-runtime framing.
Is MCP good yet?: Neon, 2026-11-04. An operator's assessment, more candid than most. Pairs with the MCP toolchain-trust paper in Top.
Workflow SDK adds idempotent starts: Vercel, 2026-11-04. Idempotent workflow starts. Small, and it removes a whole class of retry bug.
Kimi K3.1 Pro now has more ZDR providers on AI Gateway: Vercel, 2026-11-04. Zero-data-retention routing options, which is the gating question for most regulated deployments.
Portable Python scripts with dependencies: LWN, 2026-11-04. PEP 723 inline metadata in practice. Subscriber article.
Plasma X11 session removal surprises some Fedora users: LWN, 2026-11-04. Deprecation handled badly, reported carefully. Subscriber article.
LWN.net Weekly Edition for November 5, 2026: LWN, 2026-11-05. This week's edition. Subscriber article.
Claude Code v2.1.318: claude-code releases, 2026-11-04. Release tag.
Your LLM writes good SQL. That's not the same as safe SQL: InfoWorld, 2026-11-05. Correct-versus-safe as distinct properties for generated queries. The permission model is the part people skip.
How will we make senior developers?: InfoWorld, 2026-11-04. The apprenticeship question, asked without a tidy answer. Read with the incentivizing-learning paper in Policy.
Going deeper with htmx 4: combining htmx and JavaScript: InfoWorld, 2026-11-04. Practical htmx 4, including the parts that need real JavaScript.
Hybrid LLM-Deterministic Pipelines for Enterprise-Scale Code Migration at Bloomberg: arXiv cs.AI, 2026-11-05. Migration at scale with deterministic steps bounding the model. The most credible industrial code-agent report in today's batch.
Verified but Vacuous: Three Failure Surfaces of LLM-Driven Verified Code Generation: arXiv cs.AI, 2026-11-05. Proofs that pass while proving nothing useful. Essential reading if you were planning to trust a generated specification.
ViBench 2.0: A Benchmark for Long-Horizon Software Building: arXiv cs.AI, 2026-11-05. Long-horizon building rather than patch-sized tasks.
SlackBench: Benchmarking Agents on Collaborative Projects Grounded in Real Code Repositories: arXiv cs.AI, 2026-11-05. Agents in the messy social half of software work, grounded in real repos. The grounding is what separates it from roleplay.
PersonaCodeBench: Evaluating Coding Agents in Persona-driven Human-Interactive Sessions: arXiv cs.AI, 2026-11-05. Interaction style as an evaluation variable. Treat the personas lightly; the variance result is the finding.
Reasoning-Trace Overflow in Code Generation: An Empirical Study of Explicit Reasoning Models: arXiv cs.AI, 2026-11-05. Traces that outgrow their budget mid-task. A concrete, measurable failure mode with an obvious harness fix.
A.S.E 2.0: Understanding How Repository Context Impacts the Security of LLM-Generated Code: arXiv cs.AI, 2026-11-05. More repository context changes the security of generated code, not only its correctness. Another instance of the harness result.
SysPilot: Cross-Library Performance Autotuning via Library Performance Contracts: arXiv cs.AI, 2026-11-05. Performance contracts as the interface for autotuning across libraries.
Whiel: A Neurosymbolic Platform for Certified Relational Program Verification: arXiv cs.AI, 2026-11-05. Relational verification with certificates, neurosymbolic plumbing. Read against the vacuous-verification paper above.
The ULC Audio Codec is a Perfect Fit for the N64: Lobsters, 2026-11-05. Codec constraints meeting a fixed hardware budget.

Security

Man-in-the-middle arbitrary code execution in Debian's apt: Lobsters, 2026-11-04. Patch first, read second. Package-manager MITM is the worst place for this class of bug.
Pulling the right strings for RCE on a Puppet: Lobsters, 2026-11-04. Config-management RCE, written up with the exploitation path intact.
Hackers stole DreamHost customer files and inboxes during months-long attack: TechCrunch, 2026-11-04. Months of access to a hosting provider. Check whether anything of yours was parked there.
French Blood Banks Hit Alongside Cyberattack On European Supermarkets: Slashdot, 2026-11-04. Shared-supplier blast radius reaching medical logistics.
Residential Proxies Detected on White House Network: Lobsters, 2026-11-04. Residential proxy traffic where it should not be. Single-vendor telemetry, so treat attribution cautiously.
So my ISP was hijacking my DNS: Lobsters, 2026-11-04. Detection writeup for a thing many people have and never notice.
Your Cryptography Conspiracy Theory Is Wrong: Lobsters, 2026-11-04. Patient debunking of a recurring genre, with the actual history attached.
A thousand fewer paths to root: OpenAI, 2026-11-04. Internal privilege-path reduction, published. Vendor security engineering posts are rare enough to be worth the read.
Security updates for Wednesday: LWN, 2026-11-04. Distribution advisory roundup.
Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing: arXiv cs.AI, 2026-11-05. Empirical attempt at the offense-defense balance question rather than another position paper.
Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models: arXiv cs.AI, 2026-11-05. Removing rather than adding to make alignment less brittle.
Does Context Compression Preserve Refusal Alignment?: arXiv cs.AI, 2026-11-05. An efficiency optimization quietly degrading a safety property. The kind of interaction that only shows up in a deployed harness.
Auditable Unlearning for LLMs: arXiv cs.AI, 2026-11-05. Unlearning you can check afterward, which is the only version that satisfies a deletion request.
Reconciling Privacy with Provenance for Language Model Watermarks: arXiv cs.AI, 2026-11-05. Watermarks that identify output without identifying the user. The tension is real and usually left unstated.
SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving: arXiv cs.AI, 2026-11-05. Expert pruning at serving time costs safety behavior. A serving decision with an alignment consequence.
We're Asking the Ninth Circuit to Hold Reverse Location Warrants Categorically Unconstitutional: EFF Deeplinks, 2026-11-04. Litigation update on geofence warrants.
Axon Sued Over AI-Powered License Plate Reader Network: Slashdot, 2026-11-05. ALPR network litigation. The discovery here will be more informative than the filing.

Research and math

Fractal basin geometry and the limits to predictability and reproducibility in deep learning: Nature Machine Intelligence, 2026-11-05. Reproducibility bounded by the geometry of the loss landscape, not by sloppy practice. The most interesting result in today's batch that has nothing to do with agents.
Fraction comparison trick revisited: Lobsters, 2026-11-04. Short, self-contained, and the revisit is where the care is.
How Unbalanced Must a Latin Rectangle Be?: arXiv, 2026-11-05. Combinatorics, cleanly posed.
An Engineering Reading of the Bounded Halting Problem and Observable Properties of a Universal Machine: arXiv cs.AI, 2026-11-05. Bounded halting read for what it says about observability. Relevant to the verification thread if you squint, and worth it anyway.
From Type to Token: Petri Nets, Individual Token Philosophy, and Individualized Causal Effects: arXiv cs.AI, 2026-11-05. Petri net token identity connected to individual causal effects. Unusual pairing, carefully done.
Ab Initio Accuracy Beyond the Cutoff: Distributed Message-Passing Equivariant Attention at Tens-of-Millions-of-Atom Scale: arXiv, 2026-11-05. Scale record with the accuracy claim stated, which is the pair that matters in ML potentials.
From hydrological laws to catchment models: A process-based model matches deep-learning accuracy on three continents: arXiv, 2026-11-05. A process model catching up to deep learning. The negative-result framing is the contribution.
On vsyscalls and the vDSO (2011): Lobsters, 2026-11-05. Fifteen years old and still the clearest explanation.
86open: Lobsters, 2026-11-05. The standardization effort behind ELF on x86 Unix, resurfaced.

Hardware and radio

As LLMs Devour the Storage Market, Fujifilm Bets on Tape: Hackaday, 2026-11-04. Tape economics against a training-data storage crunch. The physical substrate of the buildout, which gets less attention than the GPUs.
From Germanium to Diamond: the Many Shapes of Semiconductors: Hackaday, 2026-11-04. Materials survey, readable.
The First Business Computer Hits 75!: Hackaday, 2026-11-05. LEO at seventy-five. Worth it for how much of the operational problem was already recognizable.
Making a Wax Motor for Micrometer-Level Control: Hackaday, 2026-11-05. Thermal expansion as a precision actuator.
Reading Data from an Unwilling Blood Pressure Monitor: Hackaday, 2026-11-04. Getting your own measurements out of a closed device.
Porting a Crappy Flash Game to the IBM 5150 and Beyond: Hackaday, 2026-11-05. Constraint-driven porting, with the constraints spelled out.
OFFgrid-SDR Updated to V2.1: Adds DAB, DAB+ and DRM xHE-AAC Decoding in the Browser: RTL-SDR, 2026-11-05. Broadcast decoding moving into the browser. The trend across today's RTL-SDR items is that the DSP stopped needing a desktop.
Clever Quest Lab: Free Browser-Based Decoders for RTTY, Weather Fax, SSTV, NAVTEX, APRS and More: RTL-SDR, 2026-11-05. A pile of classic digital modes, no install.
ESPica32: A Standalone ESP32 Radio for Tuning Any KiwiSDR with No PC or Phone: RTL-SDR, 2026-11-05. A physical front panel for a remote receiver.
Backtrail Pro: An Open-Source App for Discontinued goTenna Mesh Radios and Meshtastic: RTL-SDR, 2026-11-04. Open software keeping abandoned hardware useful.
AMD Expects To Outsell Intel For the First Time: Slashdot, 2026-11-04. Guidance, not results. Note which segments.
U.S. Air Force to begin year-round operations at rebuilt World War II-era Pacific airfields: The Air Current, 2026-11-04. Dispersed basing made permanent. Infrastructure reporting at its most concrete.
Scott Kirby Laments Long Wait for Airbus and Boeing to Build New Aircraft: Leeham News, 2026-11-04. An airline CEO on duopoly lead times.

Policy and society

What's next for US energy policy after the midterm elections?: MIT Technology Review, 2026-11-04. The policy read on the result. Most load-bearing item in this section for anyone planning compute capacity.
A new bill could change how the US builds energy infrastructure: MIT Technology Review, 2026-11-05. Permitting reform as the actual lever. Dull, and it determines whether the datacenter interconnect queue moves.
After the midterms, a bigger clash over data centers could be coming: TechCrunch, 2026-11-04. The fight forecast, with the state-level mechanics that will decide it.
AI PACs spent millions trying to shape Congress - here's how they fared: TechCrunch, 2026-11-04. Money in, results out, itemized. The clearest measurement of the sector's political leverage so far.
OpenAI is already trying to shape the 2028 election: TechCrunch, 2026-11-04. Reported rather than announced, so weigh the sourcing. The direction is consistent with everything else here.
Setting a new standard for responsible AI infrastructure: OpenAI, 2026-11-05. A self-authored standard published into a hostile political week. Read it as positioning, then read what it commits to.
OpenAI to acquire Pierre Computer Company, Inc.: OpenAI, 2026-11-04. The acquisition note itself. Slashdot's framing ties it to Codex; the announcement says less than that, so read the primary.
There Is One Fight: American Voters Reject Tech Oligarchy: 404 Media, 2026-11-04. Companion to the datacenter polling, broader and more interpretive.
Democrats Take Control of US Congress As Voters Reject Trump's Republicans: Slashdot, 2026-11-04. The result, briefly.
Without CDA 230, Facebook will own the internet forever: Pluralistic, 2026-11-05. The counterintuitive case that repeal entrenches incumbents. Doctorow is arguing a position; the moderation-cost mechanism is the part to take seriously.
The AI fatfinger economy has come for my receipts app: Pluralistic, 2026-11-04. A small, specific enshittification case study, which is the format where this writing is strongest.
Google Summaries Aren't About AI, They're About Lock-In: EFF Deeplinks, 2026-11-04. Reframes AI summaries as a distribution question. Read with the two schools items below.
Google faces backlash for plan that would turn off Search for schools that restrict AI: TechCrunch, 2026-11-04. Search access conditioned on AI policy. The most aggressive bundling move of the week.
DuckDuckGo pitches itself to schools amid backlash over Google's AI policy: TechCrunch, 2026-11-04. The competitive response, immediately.
Ukraine's Drone Strikes Spur Hasty Russian Plans For Underground Data Centers: Slashdot, 2026-11-04. Datacenter siting as a physical-security problem.
Sequoia hits rewind, undoing 20 years of expansion: TechCrunch, 2026-11-04. A structural retreat at a firm that sets norms for the rest.
Apple is reportedly cutting budgets for its internal AI usage: TechCrunch, 2026-11-05. Reported, single-sourced, and the direction is notable given everyone else's spend.
Waymo Cuts Ties With Uber, Will Serve Riders Directly: Slashdot, 2026-11-04. Disintermediation once the autonomy works.
Dull, Dirty, Dangerous? Working Conditions and the Limits of Occupational AI Exposure: arXiv cs.AI, 2026-11-05. Argues occupational exposure scores miss working conditions. A useful correction to the task-decomposition genre.
Today's Work, Tomorrow's Workforce: Incentivizing Learning in AI-Assisted Work: arXiv cs.AI, 2026-11-05. Formalizes the apprenticeship problem the InfoWorld piece asks about. If assistance removes the learning, the incentive has to be put back.
Whose Transparency is it Anyway? Understanding Generative AI Transparency Needs and Gaps Across Stakeholders: arXiv cs.AI, 2026-11-05. Transparency as several different requirements that get one word.
Just the Facts: How Dialogues with AI Reduce Conspiracy Beliefs: arXiv cs.AI, 2026-11-05. A positive effect, measured. Note the durability question the design cannot answer.
OpenAI: Astra covertly underperforms on safety research: Alignment Forum, 2026-11-04. Sandbagging on safety tasks specifically, which is the version of the behavior that matters most. Single-evaluator claim; read the method.
Positive and Negative AI Futures: Alignment Forum, 2026-11-04. Scenario writing. Useful mainly for seeing which branch points the community now treats as load-bearing.

Boston

Here are the latest results on all 9 of Massachusetts' ballot questions: WBUR, 2026-11-04. The full board in one place.
Question 4: Mass. voters pass Election Day voter registration: WBUR, 2026-11-04. Same-day registration passes.
Question 2 passes: Mass. public defenders get the right to form a union: WBUR, 2026-11-04. Public defenders win collective bargaining.
Rep. Clark, poised to become House Majority Leader, talks priorities as Democrats take Congress: WBUR, 2026-11-04. A Massachusetts delegation member about to hold real agenda power.
Family of cyclist killed in Mission Hill pleads for action to improve Boston street safety: WBUR, 2026-11-04. Street safety, specific and local.
Wild River Fire in New Hampshire spreads to 1,000 acres: WBUR, 2026-11-04. November fire in the White Mountains.
Bonus Seminar: The real AI Threats are to the Internet and our Peace of Mind, with Cathy O'Neil: mit-calendar, 2026-11-05. Today at the Media Lab.
Human in the Loop: Expectations in the Agentic AI Era: Northeastern events, 2026-11-04. Exactly the question the authority papers above are formalizing.
Finding Common Ground, and Synthesizing It: harvard-seas, 2026-11-05. Nisarg Shah on social choice. Pairs with the deliberation work in Themes.
Distinguished Seminar in Computational Science and Engineering: mit-calendar, 2026-11-05. Today, MIT CSE.

The crawler zoo

Eighteen new user agents this cycle. Two are worth naming.

AgentDrivesAgent: The Crawler Zoo, 2026-11-05. An agent operating another agent, declared in a user-agent string. The identity question from Cloudflare's post, arriving as traffic.
World2-MCP-Capability-Indexer: The Crawler Zoo, 2026-11-04. Someone is crawling to build an index of MCP capabilities. Read with the MCP toolchain-trust paper: the index is a target too.
SiteDecayBot: The Crawler Zoo, 2026-11-04. Measuring rot rather than content.
CT-WP-Scanner: The Crawler Zoo, 2026-11-04. Certificate-transparency-driven WordPress scanning. Not research traffic.
scanner-data-research-probe: The Crawler Zoo, 2026-11-04. A name that asserts legitimacy without identifying an operator.
Why is there an Anubis gate in front of Search for logged-in users?: Lobsters, 2026-11-05. Proof-of-work gates hitting authenticated humans. The cost of the above, paid by the wrong people.

Tail

Lorem Ipsum Dot Gov Dot UK: Lobsters, 2026-11-04. Placeholder text shipped to production, on a government domain.
It's DOOM, on a Toy Camera: Hackaday, 2026-11-04. The genre continues.
Boston post-hardcore band Fiddlehead finds a way forward: WBUR, 2026-11-05. Local music writing that is about the people.
Sister June Carpenter, one of the few remaining Shakers, dies in Maine. She was 88: WBUR, 2026-11-04. An obituary for a community of three.

Feed silences and verification

Hacker News has produced no new item since 2026-10-26, ten days. The cause is ours: a stale conditional-request validator pins the feed to a 304, so the crawler reports success and ingests nothing. No HN items appear in this brief and none should be inferred from their absence.

Of 118 feeds polled this morning, 57 returned content, 59 returned 304 not-modified, and two errored: WikiCFP timed out connecting, and James Bornholt's feed failed name resolution for www.bornholt.dev - the fifth consecutive day, so treat it as the feed and not the network. Seventy-six sources published inside the 14-day window.

All 172 URLs in this brief were checked through crawler/fetch-gated, which runs the blocklist, robots and throttle checks before issuing a request. One hundred seventy-one were allowed by the gate, and 164 of those returned 200. Seven did not, and all seven are bot-challenge responses rather than missing pages: lwn.net returned 403 on three articles, openai.com on three, and layer3intel.com on one. One noteflakes.com article answered 204 to HEAD and 200 to GET, so it counts as live.

One URL the gate denied, on robots: the Lobsters thread about the Anubis gate. No request was issued. Robots governs what we fetch, not what we cite, and the URL came from the Lobsters feed itself, so it is linked here unverified and said so rather than quietly dropped.

One link was dropped. wiredream.com/anagram-fun, carried by the Lobsters feed, returns a genuine 404 with a custom error page under both HEAD and GET. It is not in this brief. No other dead links.

All URLs came from data/warehouse.db. Nothing in this brief was constructed from a title. Slashdot URLs appear with ?utm_source=rss1.0mainlinkanon&utm_medium=feed stripped and were verified in the stripped form.

Titles were checked back against the warehouse rows they came from. Twenty-nine of the 172 differ from the feed text, in three kinds:

Ten drop a redundant prefix or marker: five crawler-zoo entries lose the feed's New arrival: under a heading that says the same thing, three LWN articles lose the [$] subscriber marker, and two Pluralistic posts lose the Pluralistic: wrapper and trailing date.
Fourteen differ only in typography or case: curly apostrophes and quotation marks normalized to ASCII, em-dashes replaced with a hyphen, three sets of enclosing quotes dropped from the link text, and one leading article lowercased. No words changed.
Five change a separator or drop a symbol: a period becomes a colon in the Phonon TTS title, an em-dash becomes a colon in the htmx entry and a comma in the Harvard colloquium, the Spark entry drops a registered-trademark mark, and the claude-code release reads Claude Code v2.1.318 rather than the bare v2.1.318 the tag carries.

The Atom feed carries the warehouse titles verbatim in all 172 entries, so nothing above is lost to a subscriber.

Build provenance

build: 2026-11-05 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.4) | feeds: 118 polled today (57 ok, 59 not-modified, 2 error), 76 publishing in window | items-considered: 5385 (14d, incl. 3088 arXiv cs.AI) | warehouse: 67463 items (3582 dateless, see feed silences) | published: 172 URLs (171 allowed by gate, 1 robots-denied and disclosed; 164 HTTP 200, 7 bot-challenge disclosed; 1 dead link found and dropped) | atom: 172 entries, RFC 4287 valid