Wolf Digest — Wednesday, July 15, 2026

The Space Development Agency awarded L3Harris $955 million and Sierra Space $798 million — $1.75 billion combined — to build 36 additional missile warning and tracking satellites, 18 each, under the Accelerated Missile Defense Tranche 3 program for the Tranche 3 tracking layer of the Proliferated Warfighter Space Architecture. SDA awarded 72 space vehicles in December to Lockheed Martin, Rocket Lab, Northrop Grumman, and L3Harris; AMDT3 brings Tranche 3 to 104 spacecraft total, with launch to low Earth orbit expected by end of 2028.

L3Harris's 18 derive from its Hypersonic and Ballistic Tracking Space Sensor, prototypes of which launched in 2024 for MDA, and will carry medium-field-of-view payloads producing fire-control-quality data. Sierra Space's are based on the Horizon bus developed for the Tranche 2 tracking layer. A 2025 executive order on Golden Dome called for accelerating HBTSS and PWSA deployment; in March, Golden Dome director Gen. Michael Guetlein disclosed a $10 billion budget increase also funding AMTI technology and the Space Data Network. The piece is sourced entirely to SDA and contractor statements, with no dispute raised.

The latest Accelerate the Procurement and Fielding of Innovative Technologies round exceeds $500 million, including more than $70M for software — totals that exclude classified procurements, so the true round value is undisclosed. APFIT sits under the Research and Engineering directorate and targets small, non-traditional, and venture-backed vendors; it has delivered 100-plus capabilities and now passes $2 billion cumulative.

The inaugural software-only procurements: ARCHER AI decision-advantage platform for PACOM ($10M); Cybergenome AI binary analysis for CYBERCOM ($10M); an intelligence and operations threat-planning tool for the Air Force ($15.2M); machine assisted information sharing for rapid intel disclosure for SPACECOM ($10M); operator tactical intranet systems for SOCOM ($17M); and SCEPTER AI-based course-of-action generation for SOUTHCOM ($10M). Hardware awards include the ALPV Liberty autonomous low-profile vehicle (Marine Corps, $32.58M), Drake counter-drone (Navy, $11.35M), a low-collateral effects interceptor and multi-mission UAS (Air Force, $17.5M), the Mira maneuverable spacecraft (SPACECOM, $19.75M), a persistent subsea autonomous profiler (Navy, $30.8M), and "Unruly" one-way attack drones (SOCOM, $31M). Sourced to a DoD release; no dispute raised.

DeepSeek's annualized revenue is approaching $500 million, strengthening both a new fundraise and IPO planning. It could file as soon as this year, weeks after completing a round that valued it at $52 billion — itself following a $7.4 billion raise. Chinese labs Zhipu AI and MiniMax have already listed; OpenAI and Anthropic are planning similar moves. Goldman Sachs analysts characterize a shift from the "DeepSeek moment" to the "Zhipu moment," with Zhipu prioritizing AGI over near-term application monetization.

Reflection AI signed a $1 billion compute deal with Nebius for access to Nvidia's latest chips, following a similar SpaceX compute deal weeks earlier. Reflection is valued at $8 billion, was founded in 2024 by two former Google DeepMind researchers, and has raised close to $2.6 billion from Nvidia, Sequoia, and Lightspeed. Nebius previously took a $2 billion Nvidia investment, then signed a five-year Meta infrastructure deal worth up to $27 billion and a multi-year Microsoft deal worth up to $19.4 billion. No duration, chip counts, or delivery timeline were disclosed, and neither party had responded to TechCrunch at publication. TechCrunch frames Reflection among open-weight developers drawing attention as data-retention concerns and government intervention push interest toward open models.

RLVR reliably improves language-model reasoning but is expensive to repeat on every new strong model, because the target must generate many rollouts during training. This work distills on-policy from a weaker trained model directly into a stronger one, avoiding the rollout cost on the expensive target. The weak-to-strong framing matters because it decouples the cost of the RL run from the size of the model that ultimately benefits — the reasoning gains transfer without re-paying for them at frontier scale.

DIU published two Director memos on July 14 (the memos themselves are dated 2 and 9 March). The 'Day 1' memo, carrying a handwritten Secretary of War endorsement, states DIU exists to accelerate conversion of commercially derived technology into combat power, and narrows investment priorities to three: high technologies producing overmatch domination; technologies substituting machines, fires, or electrons for Americans in harm's way; and capabilities that scale quickly and dramatically reduce cost-per-kill — the latter two to be fielded within three years. It states DIU is "not in the business of keeping quality science projects on life support without 4-star equivalent sponsorship," will measure and report everything, and will adopt "Fight Tonight" vigilance.

The 'Week 1' memo sets two obligations — stewardship of taxpayer money and objective evaluation of every company: "Winners are chosen on test performance not conference attendance." It requires continuous conflict-of-interest disclosure, safeguarding of vendor proprietary property, and written risk-versus-reward evaluation. Both are signed by the Director with no typed name. Note for the pipeline: both memos are image-only scanned PDFs with no text layer, so they are invisible to every text-extraction path and required OCR.

Anthropic released Claude for Teachers, giving verified US K-12 educators free access to premium Claude capabilities, a library of teaching skills, and a Learning Commons connector exposing academic standards across all 50 states — plus the finer-grained learning competencies beneath each standard and the order students typically learn them, so drafted lesson plans are scaffolded against standards. It pulls in curricular resources including OpenSciEd and IM v.360, and connects to an ecosystem of K-12 tools (ASSISTments, Brisk, Canva Education, Coteach, Diffit, Eedi, MagicSchool, Snorkl, TeachFX). It includes Claude Code and Cowork, so work can run unattended — the example given is a scheduled 4pm task reviewing exit tickets and adapting the next day's plan.

Educators only, consistent with Claude's 18-and-over policy; data is not used for training, and student information falls under a K-12 Data Processing Addendum written to comply with FERPA. Anthropic is working with the American Federation of Teachers on a privacy and safety "Gold Standard," will pilot an evaluation in the Detroit Public Schools Community District, and is releasing an open-source repository of the teaching skills plus a technical write-up of how they were evaluated. Free access for a full year if signed up by June 30, 2027.

Smartphone shipments fell 11% year-on-year in Q2 to the lowest level since 2013, as AI-driven memory shortages pushed prices higher. Entry-level phone prices are up more than 50% this year; on cheap handsets memory can be half the bill of materials. DRAM and NAND prices have soared as AI absorbs supply — 32GB kits that sold for $100 in October are reported around $350. The steepest decline was in China, where three of the five biggest suppliers saw sales slump, a fifth straight quarterly drop. Apple and Samsung held consumer prices steady and grew slightly. This is the AI buildout showing up as a second-order price shock in an adjacent consumer market.

TerraZero is a procedural driving simulator and self-play training stack aimed at the three constraints that usually conflict: fast enough for RL at scale, realistic enough to ground behavior in real map structure, and diverse enough to cover the safety-critical long tail logged data rarely contains. A configurable C engine runs the simulation, and training uses no demonstrations — agents bootstrap entirely through self-play. Surfaced across five feeds including cs.AI, cs.LG, and cs.RO.

Cohere Labs recapped Expedition Tiny Aya, a mentored program built on its open-weight Tiny Aya model (70+ languages, runs on-device). Several results are worth pulling out. On safety: evaluations produce markedly different results under code-mixed prompting — where conversations blend languages, dialects, and scripts — than in isolation, exposing blind spots single-language benchmarks miss. A separate team using mechanistic interpretability found evidence that misalignment-associated directions partially generalize across languages, with transfer strength varying significantly by language pair. On efficiency: a language-aware quantization study found compression does not degrade all languages equally, so treating multilingual compression as one uniform problem obscures language-specific tradeoffs; another team sustained multilingual JSON tool-calling on-device with 4-bit quantized variants. On multimodality: attaching existing visual encoders while keeping most parameters frozen nearly doubled performance in some experiments. On interpretability: sparse autoencoders trained on Tiny Aya's internal representations show most features are shared across languages, with a smaller set specializing by script or language, and no evidence that representation density differences explain model quality. Core Tiny Aya research was accepted to COLM 2026.

Diffusion LLMs hold a theoretical parallel-generation advantage over autoregressive decoding, but the survey's framing is that parallelism alone buys nothing in practice. Realizing it requires diffusion-aware caching and reuse, and inference efficiency has become a prerequisite for deployment rather than an optimization. A useful consolidation of where the dLLM inference stack actually stands versus where the theory says it should be.

Multimodal emotion recognition gains have come bundled with parameter growth — typically 7B and up — and the attendant compute cost. This paper asks whether the scale is load-bearing, testing sub-1B multimodal emotion models that jointly model video, audio, and language while retaining interpretable description generation. The question generalizes past emotion recognition: it is a probe of how much of MLLM benchmark performance is task-necessary capacity versus inherited scale.

Sparse MoE inference efficiency depends heavily on expert activation patterns, and existing speculative-decoding draft selection optimizes acceptance likelihood alone. This work makes draft selection cost-aware: a draft token that is likely to be accepted but activates many experts can be worse than a slightly less likely draft that activates few. Optimizing acceptance times cost rather than acceptance is the right objective for large-scale MoE serving.

Tool access alone does not make LLM empirical reasoning governable — accepted outputs need not descend from attested evidence, and accepted deductions need not survive formal scrutiny. EG-VAR is a Lean 4 tool-calling architecture in which the Lean kernel is the sole minter of Verified claims, via tool-attestation axioms and declared assumptions. The design point is that verification authority is centralized in the kernel rather than distributed across the agent's own judgment.

GRPO is now run routinely on a supervised checkpoint in the hope of a stronger agent. This controlled study asks whether it adds skill to a 4B-8B web agent or mostly reshapes behavior the supervised model already has — and reports a null, gated by learning rate. Controlled nulls with an identified mechanism are scarcer and more useful than another positive result; the finding suggests some reported GRPO gains at small scale may be reallocation rather than acquisition.

The SGLang team reports more than 500 tokens per second on 8xB300 at batch size 1 for GLM5.2 in NVFP4, reached over two weeks. The contributing changes: sync-free speculative decoding for GLM 5.2 multi-token prediction, built-in IndexShare MTP with Spec V2, a 2.33x faster TopK-V2 at 80k input sequence length, indexer prologue fusion, and GEMM kernel work. Batch-size-1 throughput is the number that matters for single-user agentic loops, where latency rather than aggregate throughput is the binding constraint.

LLM agents rarely ask how much effort a task requires, defaulting to a maximum-context-first strategy — re-reading files and dependencies already seen — which turns a one-line edit into a small code-base audit. The paper argues the missing capability is task-aware execution-scope estimation. This is the cost story behind most agentic coding bills, and framing it as a predictable scope-estimation problem rather than a prompting problem is the contribution.

Elenchos measures abductive inference — inferring latent hypotheses that explain observed behavior — as a structural inverse problem, rather than testing pattern recognition or generation. The title carries the result: models detect the evidence of a hidden cause while failing to recover the cause itself. A generative evaluation framework rather than a fixed benchmark, which makes it harder to saturate.

VLA models perform well on embodied tasks but deploying them on low-power onboard hardware like the Jetson Orin is bounded by inference latency and low control frequency. Asynchronous inference partially masks latency by parallelizing action execution against computation; Jetson-PI adds foresight alignment so the actions executed during the masked window remain consistent with what the model would have chosen. The gap between VLA benchmark results and onboard control frequency is where most embodied deployments actually fail.

Self-evolving agent systems create, revise, and retire their own skills, but every such loop rests on a hidden assumption: that a reliable evaluation metric already exists. Often it does not. The paper evolves the metric itself, searching compositions of small drawback detectors under a full evolutionary lifecycle trained to agree with reference judgments. If the grader is learned alongside the skills, the failure mode is co-adaptation — which is exactly the risk the evolutionary lifecycle is meant to bound.

Failure attribution — identifying which step in a failure trajectory caused the task to fail — is critical for debugging agentic systems. Existing approaches either use expensive prompting pipelines or require post-training on failure trajectories with step-level error annotations that are costly to collect. This work derives attribution from the structure of successful runs instead, sidestepping the annotation bottleneck.

Existing long-term memory benchmarks evaluate almost exclusively through downstream QA, scoring only final-answer correctness — a black-box formulation that conflates heterogeneous causes of memory failure, such as missing the information versus retrieving it and reasoning over it wrongly. MemOps decomposes the memory lifecycle into operations and benchmarks them separately. Attribution rather than aggregate accuracy is the contribution.

Asked what experts believe about a contested claim versus what it believes, a trustworthy model should answer in two different epistemic registers — neutral attribution in the first case, stance expression in the second. This work probes whether that shift occurs and whether it occurs coherently, which is not directly assessed by standard evaluations. A clean operationalization of a property usually discussed only qualitatively.

GraphRAG methods add structured entity-relation graphs to improve retrieval and reasoning but treat knowledge graphs as static structures built offline and queried in a single pass. EvoGraph-R1 makes the hypergraph self-evolving and multimodal, updating structure across retrieval episodes rather than freezing it at build time.

An embodied foundation model targeting agents that must reason about actions, adapt to evolving situations, and interact with the physical world — not just perceive and understand multimodally. The report positions efficiency alongside capability, which is the live constraint for anything running on a robot rather than in a datacenter.

Most agent systems run on desktops or servers. Mobile devices are arguably the more important agent environment — widely accessible, holding user data and sensors — but constrain compute and memory hard. PalmClaw is a native on-device agent framework for phones, executing multi-step tool-calling loops locally. Reads naturally against Bonsai 27B: the model side and the harness side of the same on-device bet.

LLMs are trained to perceive and resonate with user emotion, but whether induced emotion shifts their own sequential decision-making is unclear. This study uses the Iowa Gambling Task — a standard instrument for probing decision-making under uncertainty in humans — to test it. Relevant as LLMs move into high-stakes autonomous roles where an affective context window should not move the policy.

Reaction mechanisms are step-by-step sequences of elementary reactions explaining chemical transformations, and stepwise mechanistic deduction aligns naturally with reasoning-model paradigms. Current chemical LLMs largely do not learn mechanism logic, predicting products without the intermediate structure. Teaching the mechanism rather than the mapping is the bet.

Robot learning increasingly relies on dense progress signals to evaluate intermediate states, guide policy learning, and detect completion. Since dense labels are rare at scale, normalized time within a demonstration is the usual substitute — later frames treated as higher progress. That proxy is wrong whenever a demonstration stalls, backtracks, or idles. UR-VC corrects the time-derived value signal without supervision.

Anthropic is committing $10 million CAD to Canadian research institutions, partnering with the three regional AI institutes — Amii (Edmonton), Mila (Montréal), and the Vector Institute (Toronto) — plus CHEO, CAMH, Université Laval, University of Toronto, and the University of Saskatchewan. Work spans reinforcement learning and trust and safety at Amii; responsible AI, health, multi-agent systems, and robotics at Mila; computational mental health and large-scale fairness evaluations of psychiatric AI systems at CAMH's Krembil Centre; and LLM behavior across cultural contexts and low-resource languages, including Quebec French and Indigenous languages, at Laval. Amii, Mila, and Vector join Anthropic for Startups, with affiliated Canadian startups receiving at least $5,000 USD each in API credits. Anthropic also published a Canadian country brief from its Economic Index: Canada ranks eighth worldwide in Claude.ai use and second by AI Usage Index, accounting for 2.6% of global consumer use at more than four times the per-capita rate population predicts. Translation requests concentrate in provinces with more government employment — New Brunswick, Nova Scotia, and Québec.

Reliable RL agents must hold up under sensor malfunction, dynamic disturbance, and slow environmental shift, which makes detecting when observations, transitions, or trajectory dynamics depart from training assumptions pivotal. OOD-RL-Bench provides a common framework for evaluating OOD detection across these three distinct deviation types, which prior work tends to conflate.

Complex swarm behaviors emerge from simple rewards with no explicit aggregation incentive, but the black-box nature of neural policies makes the mechanism opaque and limits multi-robot deployment. This work connects the emergent collective behavior back to the reward structure that produced it, which is a prerequisite for trusting MARL swarms in the field.

Self-repair literature measures retry-after-failure without placebo controls. PoPE treats a failed program as a conjecture and an execution counterexample as an oracle-relative refutation, then asks whether the information guiding the retry is the content of the feedback or merely its form. Preregistered and placebo-controlled — methodologically stronger than most of the self-repair results it is auditing.

Superparamagnetic nanodot ensembles driven by strain-induced magnetoelectric coupling are candidates for ultra-low-energy reservoir computing substrates, but their dynamics are governed by temperature, so real-world deployment demands controlling that sensitivity. This work addresses reproducibility under environmental variation — the gap between an unconventional-computing demo and a device.