FinSMEs: CarbonSix, which is developing AI automation and robotic intelligence tech for manufacturing, raised a $40M Series A co-led by DSC Investment and LB Investment  —  CarbonSix, Inc., a San Francisco, CA-based Physical AI company focused on the manufacturing sector, raised $40m (approx. KRW 60 billion) in Series A funding.

Neo is Bhavin Turakhia’s fifth venture and his latest involving enterprise software. This time he's taking on Microsoft Office and Google Apps with AI.

Sayan Chakraborty / Nikkei Asia: Indian IT company Persistent Systems offers to acquire its Germany-based peer Nagarro for $1.45B; Indian IT companies are ramping up M&A as AI reshapes growth  —  BENGALURU — A move by Persistent Systems to buy German peer Nagarro in a $1.45 billion deal has highlighted how Indian software exporters …

Discussion | Link

Discussion | Link

Base44's CEO said the company developed its own AI model to deliver what frontier models can't. Base44 Base44 wants to make sure your next vibe-coded website doesn't look like AI slop. The startup launched its LLM, Base 1, to deliver better designs and user experience than frontier models. CEO Maor Shlomo said he hopes Base 1 will deliver unique designs each time, although it's not there yet. Base44 wants to help users ditch the cookie-cutter look of AI-coded websites with its new AI model. The San Francisco-based vibe coding startup announced on Monday that it had trained and released its own large language model called Base 1. Before giving the prompt to the platform to build an app, users can now select Base 1 from a selection of AI models, including Claude's Opus 4.8, Fable 5, and OpenAI's GPT-5.5. Shlomo told Business Insider that part of the reason his company developed its own model was to combat the telltale generic look that many vibe-coded products have. He said the problem with using frontier models for making websites is that "everybody feels like they're getting the same UI when they're coding with the general models." Base44 was acquired by Wix last June for $80 million, making it part of Wix's website-building arsenal. Shlomo said Wix has a large team of designers, which generates a lot of data for its model to train on. The Base44 team will conduct "reinforcement learning" on the new model, Shlomo said, which involves prompting it to keep generating designs that look new and unique. While Base 1 is "not yet there," Shlomo said he aims for it to "create something that looks uniquely different" each time it generates a user interface. He said the team started working on Base 1 about six months ago, but had breakthroughs in the last few weeks, allowing them to release the model sooner than planned. He said generating unique designs will set Base44 apart from competitors. Startups like Lovable, Replit, and Cursor compete with Base44. Many UI/UX experts and design gurus have warned about the AI slop look on vibe-coded websites and apps. Paul Bakaus, the CEO of AI design startup Impeccable, said in a June interview with Andreessen Horowitz that telltale signs of AI-coded products include beige or tinted backgrounds and sans-serif fonts. Bakaus likened it to an "algorithmic Uniqlo or Ikea," a design that's not bad, but not unique. Shlomo told Business Insider that frontier models must be generically good at everything, from poetry to coding. "And we think that if we take a model and we squeeze its ability to be really, really good at one use case, then we have a shot," he said. Read the original article on Business Insider

Russian military servicemen shout during the Victory Day 2026 parade in Moscow. Contributor/Getty Images Ukraine's defense minister said Russian troops are now walking 18 miles to the front lines. That takes a soldier with combat gear about six to eight hours. It's come as Ukraine has tried to disrupt the Kremlin's supply lines in the southern rear. Russia's crippled logistics routes on the war's southern front are forcing its infantry to walk some 18 miles to their combat positions, Ukraine's defense minister said on Wednesday. "You've already seen, for example, in the south, where Russian infantry, due to logistics being disrupted, are walking 30 kilometers on foot to reach positions," Mykhailo Fedorov said at a joint press conference in Kyiv with his Swedish counterpart, Pål Jonson. Russian and Ukrainian troops are typically moved by vehicle to intermediate positions before traveling the last few miles on foot, at night, or in smaller vehicles to avoid detection by light drones. Marching 30 km, or about 18 miles, in one go would take a soldier in combat gear at least six to eight hours. Fedorov said Russian forces have had "major problems getting infantry to the front line" and sending them supplies, fuel, and generators used by drone operators. The logistics disruption comes as Ukraine has increasingly used a newer class of drone capable of striking at depths of 30 to 300 km, targeting Russian supply routes, bombing trucks, bridges, and facilities in the rear. "There are a lot of problems not obvious at first glance that affect the intensity of enemy actions on our territory," Fedorov said. Crimea, a major logistics hub for Russian operations in Kherson and Zaporizhzhia, is also facing a severe energy crisis as Ukrainian mid-range drones bombard oil facilities and key bridges connecting the peninsula to the mainland. Local officials began rationing fuel for civilians last month before announcing they would cut fuel sales indefinitely. "A large number of crises are beginning to accumulate, and it is becoming increasingly difficult for Russians to resolve crisis after crisis," Fedorov said. The mid-range drone has been key to Ukraine's renewed war effort, giving Kyiv intermediate-range strike capacity that it can deploy at scale rather than relying on a limited Western supply of missiles. Some of these fixed-wing drones use AI to help with Russian signal jamming, and are designed to identify and engage targets with some degree of autonomy, George Barros, director of Innovation and Open Source Tradecraft at the Institute for the Study of War, previously told Business Insider. Fedorov and other Ukrainian commanders have said before that Read the original article on Business Insider

Will Ferrell says his father's advice reshaped the way he approached his Hollywood career. Lia Toby/Getty Images for Warner Bros. Will Ferrell says a conversation with his father changed how he approached his Hollywood career. His father said talent wasn't enough — success in show business also takes luck. Trying, even if it didn't work out, wouldn't be failure, his father told him. Will Ferrell says that a conversation with his dad before he pursued comedy taught him to stop fearing failure. On Wednesday's episode of the "IMO" podcast, Ferrell spoke with Michelle Obama and Craig Robinson about the period when he decided to give Hollywood a real shot. At that time, he was taking classes at The Groundlings, an improvisational and sketch-comedy school in Los Angeles, and said he had begun to realize that the work "just feels right." Before taking the leap, Ferrell sought his father's opinion. Looking back, the "Elf" star said his dad's response became what he still considers "the best showbiz advice ever." "He was like, 'If it was based on talent, I wouldn't worry about you. I'm going to tell you you've got talent, but there's so much luck,'" Ferrell said, recounting his father's words. "Give it a shot, but don't get discouraged if you get down the road and you're like, this is just too hard. You didn't fail. You tried, and you're capable, just pivot and do something else," Ferrell said. Even though it wasn't a "rah-rah speech," the comedian said his father's words "took the pressure off succeeding." "I thought, well, I'm not going to make it anyway. Let's just have fun," Ferrell said. Many other celebrities have also spoken about the advice they received from parents or mentors early in their careers. Rashida Jones has said her father, Quincy Jones, encouraged her to use the opportunities available to her because of her family's ties to the entertainment industry. "My dad said to me, when I graduated from college: 'You're gonna go wait in line with 70,000 other people for a job? That doesn't seem really that practical,'" Jones said. "And he was right, you know." Selena Gomez has also credited her grandpa with teaching her the importance of punctuality. "My papa used to say, 'If you are on time, you're already late,'" Gomez said. Read the original article on Business Insider

Dan Ives is leaving Wedbush as managing director and senior equity research analyst for his own venture. NASDAQ: ORBS) Dan Ives is leaving Wedbush to launch a merchant bank focusing on the tech and energy sectors. Ives' early predictions on Microsoft and Nvidia's AI roles proved accurate and influential. Look back at Ives' hottest tech bets — and some of his most iconic outfits. For many, Wedbush Securities has become synonymous with Dan Ives: the charismatic, media-savvy tech analyst known for his peacock-esque personal style, colorful metaphors, and plentiful emojis to describe tech's biggest moments. Now, after eight years at the firm, Ives said he's leaving for a new chapter. "I have truly enjoyed my time at Wedbush and proud of all we accomplished as a great tech franchise with awesome people across the board," he said in a widely shared email on Wednesday. He told Bloomberg that he's leaving to launch a merchant bank — a bank primarily serving large corporations and high-net-worth individuals. The company will focus on sectors such as technology, energy, infrastructure, and financials, he said. Here are some of the tech mega bull's most famous calls — and some of the suit eschewer's punchiest outfits over the years. 1. Apple isn't just a hardware company Between 2016 and 2018, Ives hyped Apple's transformation, saying that investors were only thinking of it as a hardware company and missing the value of the app store and services. By 2020, Apple's services revenue had doubled. It's now a major driver of its $4.3 trillion valuation. 2. Microsoft's AI win Immediately after OpenAI released ChatGPT in November 2022, Ives argued that Microsoft would be the biggest beneficiary of enterprise AI. This is because the company embedded AI across services like Office and Copilot, and because of infrastructure spend on Azure. Microsoft's stock roughly doubled over the next two-plus years. 3. Nvidia will be king When many investors still viewed Nvidia primarily as a gaming chip company, Ives argued that AI infrastructure spending would create unprecedented demand for the company's products. Ives' early thesis was that every major company would need Nvidia chips to build its AI. Ives was spot on. Since 2023, hyperscalers have committed hundreds of billions of dollars to AI infrastructure, making Nvidia one of the world's most valuable companies. 4. Pink on pink on pink suit Ives is rocking a pink top to bottom. NASDAQ: ORBS) Ives rocked one of his most iconic outfits — a coral pink baseball cap and sports jacket with baby-pink pants and shoes — when he rang the Opening Bell for Eightco Holdings Inc (ticker: $ORBS) at the Nasdaq in September. He chaired the company's board until March. Last August, Ives launched a clothing brand reflecting his unique style in collaboration with the Brooklyn-based upcycled streetwear label Snow Milk. 5. Jacket-on-hoodie looks Ives almost always matches his jackets to his sneakers. $ORBS) At another Orbs event in October, the tech bull wore a neon printed blazer under a colorful green hoodie. In his signature style, he paired it with multicolored sneakers that almost exactly matched his jacket. He also wore it with a printed baseball cap and sunglasses. Yes, the event was indoors. "It's definitely played a role in the global success, because of the way that I dress," he told Business Insider in September. "Even people who don't know my stock calls, they know me as the person who wears the funky shirts on CNBC." Read the original article on Business Insider

Vinton Cerf, the father of the internet, said that AI reminds him of the early days of the internet. Noam Galai/Getty Images for Webby Awards Vinton Cerf said AI reminds him of the early days of the internet. He helped invent the networking protocols that underpin the modern web. Cerf said there are a few principles that could decide whether AI will flourish like the internet. As AI races toward an agent-powered future, a legendary systems engineer says it reminds him of another technological revolution he knows quite well: the birth of the internet. Vinton Cerf, the now 83-year-old who helped invent the networking protocols that underpin the modern web, has important guidance for the AI boom. During a panel this week with computer scientists and Databricks cofounder Matei Zaharia at the Open Frontiers conference, Cerf highlighted several principles that could help determine whether AI reaches its full potential. Open standards matter more than closed systems Cerf said the internet only became ubiquitous because no single company owned it, and everyone could use it. "In the case of internet, it only worked because it was going to be distributed to begin with," said Cerf. "And so we left the rules very open. We just said if you can find somebody to connect to and you follow the rules of the protocols, it should work." Having common protocols meant that a university network in California, a government research lab, and a commercial internet service provider could all connect using the same technical language. Cerf said AI is approaching the same inflection point at which the number of AI agents would demand "interoperability and standardization." AI agents will need better ways to communicate While we might speak to an AI agent in our native tongues, those vernaculars won't be ideal for our agents to work together effectively. Cerf said natural language may not be enough for agents to work together reliably or to create interfaces that enable developers to innovate without disrupting the entire system. "I don't think English is going to be the best choice," said Cerf. "There is ambiguity, and I think precision for inter-agent interaction is going to be very, very important." Human languages often rely on context, inference, and multiple meanings for the same words. Cerf said AI agents will need communication methods that minimize vagueness so each system can reliably interpret requests and commitments. "An agent really needs to be sure the other agent understands what it is that they just agreed to do together," Cerf added. The biggest technologies become platforms Like the internet, Cerf said, transformative technologies aren't stand-alone products but must become foundations that enable others to build. "A lot of the successes come from enabling technologies, whether it's a platform or some other fundamental element that others can build on," Cerf said. Google, Amazon, Netflix, and millions of smaller developers built services on top of the same underlying internet infrastructure. Cerf said the same principle applies to AI. "So if you really wanted to look for impact, think about things that enable other people to do things that they want to do," Cerf added. Read the original article on Business Insider

arXiv:2607.00095v1 Announce Type: new Abstract: Generative models have emerged as scalable surrogates for physical simulation, yet they offer no guarantee that their outputs respect the conservation laws, boundary conditions, and nonlinear invariants that govern the underlying physics. Constrained sampling closes this gap, enforcing such constraints exactly at inference time without retraining, but at a computational cost: projection, correction, and trajectory-optimization steps are repeated during sampling, with these steps becoming expensive for nonlinear constraints. Standard ML frameworks exacerbate this: their dense tensor algebra and limited sparse solver composability obscure the structure that physical constraints naturally induce, making efficient batched nonlinear optimization difficult to realize in practice. We address this bottleneck by exploiting the structure that sample-wise batching and local PDE couplings induce in the projection subproblems -- namely, block-sparse Jacobian and KKT systems -- exposing this structure using ExaModels.jl and solving the resulting sparse nonlinear programs with MadNLP.jl and GPU sparse factorization. Applied to Physics-Constrained Flow Matching (PCFM), on PDE benchmarks with linear, nonlinear, one-dimensional, and two-dimensional constraints, this approach accelerates nonlinear constraint projection while maintaining constraint satisfaction. These results show that sparse GPU nonlinear optimization is a practical foundation for constrained generative sampling in scientific machine learning.

arXiv:2607.00113v1 Announce Type: new Abstract: Background. Labeled data for security classification is scarce. Semi-supervised learning (SSL) propagates labels from a small labeled pool to larger unlabeled pools. Yet security applications often use SSL as a black box: default parameters, a fixed classifier, and no handling of pseudo-label-induced class imbalance. Aims. Recent work reports sizeable gains from optimizing SSL pipelines via joint search, AutoML, or per-component tuning. These gains are hard to attribute: they may reflect useful SSL-classifier interactions, or mostly from simply tuning the downstream classifier. We disentangle these effects for binary tabular security data with classical SSL and tree-based classifiers. Method. We build SemiScope as an analysis instrument, not a deployment recommendation. It uses Bayesian Optimization to jointly tune SSL settings, confidence filtering, oversampling, and the classifier. The key control, Tuned-Clf, fixes SSL to defaults but gets the same 100-trial classifier budget and validation-set threshold tuning as SemiScope. At 10% labels, we compare them with paired TOST using a +/-1.0 g-measure smallest effect of interest. Results. SemiScope beats every default SSL baseline on all five datasets, improving over the strongest by 0.7-12.7 points. Under the equal-budget control, Tuned-Clf is statistically equivalent to the full pipeline on 4 of 5 datasets; Phishing is inconclusive. Classifier HPO alone recovers a median 86% of SemiScope's gain over Default Self-Training (ST) + Random Forest (RF). Conclusions. The reusable contribution is the decomposition protocol. A simpler recipe suffices: use Self-Training, tune the classifier with Bayesian Optimization, and tune the decision threshold on validation data. It reaches within 1 g-measure of Supervised RF at 20-30% labels on four datasets and 40% on Drebin, at the same or lower label rate than Default ST + RF on every dataset.

arXiv:2607.00127v1 Announce Type: new Abstract: Survival analysis models time-to-event data, but in clinical settings training data are costly and scarce: events accrue over years of follow-up, cohorts are small, and privacy regulations restrict sharing across institutions. Tabular generative models promise augmentation and privacy-preserving cohort sharing, yet are themselves data-hungry -- on the small cohorts typical of survival analysis, a single generator rarely characterizes the population well enough for downstream models trained on its output to match real-data performance. FoGS (Filtered Mixture-of-Generators for Survival analysis) reframes synthetic-data construction as sample selection rather than generation. A candidate pool is drawn from four architecturally distinct tabular generators, and each sample is scored by an ensemble of seven survival models trained on real data, using proper scoring rules as a per-sample plausibility proxy. A two-level pipeline optimizes, in its outer loop, a selection policy -- generator quotas, scorer weights, a random complement, and stratified balancing on event time and censoring -- against held-out downstream performance, while an inner loop tunes the downstream model (XGBoost-Cox). On 16 public datasets under train-on-synthetic, test-on-real (C-index and IBS, $0$--$100$ scale), FoGS yields mean improvements of $+2.17$ in C-index and $+0.67$ in IBS, improving both metrics on 9 of 16 datasets and at least one on 13 (one-sided Wilcoxon $p=0.039$ and $p=0.035$). It matches or exceeds real-data training on most cohorts, with no significant change in nearest-neighbour privacy margin relative to unfiltered sampling. Sample filtering over a heterogeneous generator pool is thus a viable substitute for real-data training in privacy-restricted clinical settings.

arXiv:2607.00152v1 Announce Type: new Abstract: Three of the most popular methods for training language models to reason look like three different tricks. They are not. All three adjust a single number: standard deviation, reflecting how much a prompt's sampled answers disagree. When such a model is trained, it answers each problem many times, and an automatic checker marks every answer right or wrong. The standard deviation of those marks measures the disagreement: largest when the answers split evenly between right and wrong, and zero when they all agree. Group Relative Policy Optimization (GRPO) divides by this number, GRPO Done Right (Dr. GRPO) drops the division, and Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) discards the groups where it is zero. Each is presented as its own fix, yet this paper proves they are three settings of one dial. That dial is not cosmetic: for right-or-wrong rewards, the disagreement is exactly the size of the training update, the group-standard-deviation identity. A split group teaches the most, while a unanimous group teaches nothing and falls silent. The same result says which problems deserve the most weight and how many tries each one needs. This paper confirms the intuition on a large real difficulty dataset (Big-Math) and in a controlled training run. What looks like a harmless normalization step is the dial that decides where learning happens and how strongly.

arXiv:2607.00089v1 Announce Type: new Abstract: Mechanistic interpretability has produced a rich inventory of component-level analyses that characterise what neural-network components encode and how they interact. Their outputs, however, are not easily reusable: selectivity tables, circuit diagrams, and feature lists remain locked in per-study notebooks - non-composable, not queryable in natural language, and not directly actionable for downstream audit or intervention. We study the representation layer that sits between these analyses and downstream use as a bottleneck that can be evaluated independently, and introduce Manifestation Units, a typed tuple protocol (E, S, R, D, G) extended with attention-head primitives (T) for transformer architectures, organising per-component statistics into structured fields populated automatically and queried through hybrid retrieval. Instantiated across generative vision (beta-VAE), discriminative vision (CNN), and language (GPT-2), the protocol supports two findings: typed structure substantially outperforms unstructured baselines on retrieval, and CNN filters retrieved by the schema satisfy causal sufficiency and necessity criteria under matched-budget controls. The schema absorbs attention-head primitives without modification, set-recovers known IOI circuit members under retrieval-budget-matched controls, and reveals an irreducible two-field core (S+R) with remaining fields either redundant or actively interfering. We present this as schema infrastructure for mechanistic interpretability rather than frontier-scale validation.

arXiv:2607.00009v1 Announce Type: new Abstract: Despite the remarkable proficiency of large language models (LLMs) in basic writing assistance, their utility in creative writing is fundamentally hindered by a persistent binary failure. This issue manifests as an oscillation between safe, surface-level editing, referred to as remedial polishing, and destructive, uncontrolled plot expansion. This dilemma defines a critical trade-off between narrative fidelity and descriptive intensity. We propose Loom, an assisted writing framework grounded in the narratological distinction between story and discourse. Loom employs a three-layer pipeline that operationalizes an intent-centered semiotic chain-of-thought to enforce precise control over narrative intent and rendering density. This architecture separates the generation of perceptual material from syntactic insertion, ensuring that enhancement occurs without violating the original event structure. Our comprehensive evaluation, which includes LLM-based metrics and human assessment, demonstrates that Loom successfully resolves this fundamental tension. Loom achieves the highest overall quality score, yielding substantial gains in factual integrity and descriptive intensity compared to state-of-the-art baselines.

arXiv:2607.00154v1 Announce Type: new Abstract: Evolutionary neural architecture design for multivariate time-series forecasting remains underexplored, with most approaches relying on fixed Transformer architectures despite substantial variation across tasks and forecasting settings. This paper introduces an evolutionary neural architecture search framework for discovering task-adaptive Transformer-like models for time-series forecasting (EVOTS). Architectures are encoded using a modular genome representation that enables flexible composition of attention, feed-forward, and projection components, while a repair mechanism enforces structural validity throughout the evolutionary process. This formulation allows effective exploration of a diverse architecture space without relying on hand-crafted design rules. The proposed approach is evaluated on four benchmark datasets from the ETT family (ETTh1, ETTh2, ETTm1, and ETTm2) under multiple forecasting settings, including univariate-to-univariate, multivariate-to-univariate, and multivariate-to-multivariate prediction, with horizons of 96, 192, 336, and 720. In the multivariate-to-multivariate setting, the evolved architectures achieve competitive and, in several cases, improved mean squared error relative to a strong Transformer-based baseline. Additional analyses examine performance differences across forecasting settings and report wall-clock training time to provide a coarse indication of computational cost. Overall, the results demonstrate that evolutionary search can effectively discover flexible and high-performing Transformer-like architectures for multivariate time-series forecasting within practical runtime constraints.

arXiv:2607.00124v1 Announce Type: new Abstract: Object-centric models inspired by DETR have become the dominant paradigm for open-vocabulary video instance segmentation (OV-VIS). While recent efforts have reduced the computational cost of pixel decoding, textual modality fusion, and object decoding to make these architectures more suitable for mobile devices, real-time on-device inference at high frame rates remains an open challenge. In this paper, we introduce SegFS, a dual-stream fast-slow framework that significantly improves efficiency without sacrificing accuracy. On sparse keyframes, an open-vocabulary object-based model predicts instance-level representations. These representations are then projected back into the backbone feature space to condition a lightweight fast network, which efficiently relocalizes and segments the instances in subsequent frames. By shifting instance propagation from object decoding to feature-space conditioning, our approach decouples multimodal semantic understanding from dense mask prediction and enables efficient temporal propagation. The proposed fast branch achieves up to 14x lower latency than the mobile-oriented MOBIUS model, while maintaining competitive segmentation performance on standard OV-VIS benchmarks.

arXiv:2607.00162v1 Announce Type: new Abstract: Parameter-efficient fine-tuning (PEFT) reparameterizes weight updates in a fixed basis: low-rank adapters operate in the spatial domain, while a recent line of spectral methods operates in a fixed Fourier domain. We argue that the choice of domain is itself a design degree of freedom that should be learned, and that no single basis is optimal across tasks, layers, or tokens. We introduce Fractional-Fourier Mixture of Experts, a mixture-of-experts adapter in which every expert carries a learnable fractional-Fourier order that continuously interpolates between the spatial domain (recovering vanilla LoRA) and the Fourier domain (recovering a spectral adapter). Routing tokens through experts that occupy different points on this spatial-spectral continuum lets the model place each low-rank update in the domain where it is most compact, and -- because fractional-Fourier operators of different orders are mutually incoherent -- makes the experts naturally decorrelated, which reduces interference and improves multi-task composition. The order is a single scalar per expert, trained with a separate optimizer, and the transform is computed with an $\mathcal{O}(d\log d)$ chirp--FFT surrogate, so Fractional-Fourier Mixture of Experts adds negligible cost over standard MoE-LoRA. Across commonsense, mathematical, code, and knowledge benchmarks on LLaMA-3.1-8B and Qwen2.5-7B, Fractional-Fourier Mixture of Experts improves over strong MoE-LoRA and spectral baselines -- including FlyLoRA, FourierMoE, and HMoRA -- while keeping the active-parameter budget small, and analysis shows that the learned orders specialize by task and layer in interpretable ways.

arXiv:2607.00164v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards can in principle train calibrated probabilistic forecasters, since a proper scoring rule such as the Brier score is computed from outcomes alone and is minimized in expectation by the true probability. In practice it degrades calibration, and existing remedies address epistemic uncertainty, where a model's confidence accompanies a verifiably correct or incorrect answer. We study aleatoric forecasting, where the forecast itself is the output and the label is one stochastic outcome, taking NFL in-game win probability as a testbed with the betting market as a reference. Rewarding the realized per-play outcome fails, because the single outcome is a noisy target and the policy gradient corrupts the chain of thought. We introduce a verifiable, label-free reward, a state-conditioned empirical win rate estimated from past outcomes, that removes the label noise, and we keep the gradient off the reasoning, by direct prediction or a gradient mask, so it cannot be corrupted. Trained with this reward alone, without human labels or supervised fine-tuning, a 7B model reaches the calibration of the betting market by direct prediction and is better calibrated than a zero-shot frontier model. That frontier model and a tabular estimator reach the same Brier score as this model, identifying the market's small remaining edge as live in-game information beyond their shared inputs. Masking the gradient, rather than dropping the chain of thought, preserves reasoning from which the forecast follows, which ordinary chain-of-thought training corrupts.