Run metadata:

• ran_at: 2026-07-29
• job_id: physical_ai_model_to_matter
• task: Task 1 — Physical AI / Embodiment Scan + Task 2 — Ethan Portfolio Engine
• scope: physical AI / embodiment / humanoid robotics / robotics integration / bionics-prosthetics / AI-driven materials / BCI / defense robotics / capital-flow signals
• web_current: TRUE
• source_basis: Sources fetched in this run only.

---

EXECUTIVE READ

The highest-signal shift in this run is physical AI moving from “demo race” into three partially investable conversion paths:

1. robotics infrastructure/platform layer — NVIDIA Cosmos / Isaac / Jetson / Omniverse;
2. deployment layer — Figure/BMW, Agility/GXO, UBTECH/Hitachi, Tesla Fremont Optimus tooling;
3. defense autonomy layer — Anduril / Shield AI / AeroVironment contract and capital acceleration. NVIDIA announced Cosmos 3 and Cosmos 3 Edge for physical AI and robot policy deployment, with Japanese robotics/manufacturing partners building on Cosmos / Isaac / Metropolis / Jetson. Source URL: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/default.aspx and https://investor.nvidia.com/news/press-release-details/2026/Japa…REDACTED/default.aspx (investor.nvidia.com)

The strongest deployment-confirmed humanoid signal is Figure at BMW: BMW says Figure 03 will start at Spartanburg on complex sequencing logistics after a successful Figure 02 deployment, and BMW’s own press attachment says Figure 02 supported production of more than 30,000 BMW X3 vehicles. Figure’s own update says Figure 03 arrived at BMW on June 30, 2026. Source URLs: https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-…REDACTED and https://www.figure.ai/news/f-03-at-bmw (press.bmwgroup.com)

The strongest capital-flow signal is not one round; it is clustering: Skild AI announced a $1.4B Series C at a valuation over $14B; Anduril raised $5B at a $61B valuation and is reportedly already exploring a new round around a $100B valuation; UK humanoid startup Humanoid raised $152M Series A at a $1.35B valuation; and Physical Intelligence was reported to be discussing a ~$1B raise at an $11B+ valuation. Source URLs: https://www.skild.ai/blogs/series-c, https://techcrunch.com/2026/05/13/andu…REDACTED/, https://www.axios.com/2026/07/24/andu…REDACTED, https://ca.marketscreener.com/news/robo…REDACTED, https://techcrunch.com/2026/03/27/phys…REDACTED/ (skild.ai)

SIGNAL_LAYER

Physical AI platform layer hardening: NVIDIA Cosmos → edge policy deployment

NVIDIA launched Cosmos 3, described as an open world foundation model for physical AI combining vision reasoning, world generation, and action prediction; NVIDIA also announced a Cosmos Coalition including robotics and AI participants such as Agile Robots, Generalist, Runway, and Skild AI. The portfolio-relevant signal is that NVIDIA is positioning itself not merely as a GPU supplier but as the operating infrastructure layer for robot training, synthetic data, robot evaluation, and edge deployment. Source URL: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/default.aspx (investor.nvidia.com)

NVIDIA’s July 15 Tokyo announcement added Cosmos 3 Edge for on-device vision reasoning and robot policy deployment on Jetson Thor, with Japanese robotics and manufacturing leaders building on Cosmos, Isaac, Metropolis, and Jetson. This is a conversion signal because “physical AI” is shifting from training/demo stack toward deployable edge inference and industrial integration. Source URL: https://investor.nvidia.com/news/press-release-details/2026/Japa…REDACTED/default.aspx (investor.nvidia.com)

Interpretation: Platform value is currently more investable than humanoid OEM value because public access is concentrated in compute, simulation, sensors, automation controllers, and defense/autonomy suppliers. NVIDIA remains the reference public name here.

---

Humanoid deployment: BMW/Figure is still the cleanest “factory reality” proof point

BMW’s June 25 release says Figure 03 humanoid robots will start at BMW Group Plant Spartanburg, working on complex sequencing applications in logistics, after a successful Figure 02 deployment. Figure’s June 30 update says Figure 03 arrived in Hall 52 at BMW Spartanburg and frames the use case as dynamic material manipulation not easily solved by fixed automation. Source URLs: https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-…REDACTED and https://www.figure.ai/news/f-03-at-bmw (press.bmwgroup.com)

BMW’s own attachment says Figure 02 supported production of more than 30,000 BMW X3 vehicles. That number matters because it separates Figure/BMW from pure stage demos, though it still does not prove broad-margin commercial scalability. Source URL: https://www.press.bmwgroup.com/global/article/attachment/T0458778EN/651004 (press.bmwgroup.com)

Figure also claimed in April that BotQ had delivered over 350 Figure 03 robots and increased production from one robot per day to one per hour. Because this is company-reported and not independently audited in the fetched sources, treat it as a watch-positive internal production signal, not yet a portfolio-action proof point. Source URL: https://www.figure.ai/news/ramp…REDACTED (figure.ai)

Interpretation: Figure remains one of the few humanoid companies with a credible documented factory-learning loop: deployment → data → OTA/software → next-gen hardware → redeployment. But Figure is private; public exposure is indirect through partners and suppliers.

---

Tesla Optimus: production preparation signal, but current wording increases execution-risk watch

Tesla’s AI page still frames Optimus as a general-purpose bipedal autonomous humanoid robot for unsafe, repetitive, or boring tasks. Source URL: https://www.tesla.com/AI?redirect=no (tesla.com)

Tesla’s Q2 2026 reporting cycle produced mixed physical-AI signals. TechCrunch reported that Tesla removed prior language about Optimus reaching “volume production” and quoted Musk saying Optimus would be the hardest product Tesla has ever scaled because everything on the robot is new. Axios reported Tesla expects Optimus production to begin at Fremont later in 2026 after tearing out discontinued Model S/X assembly lines. Source URLs: https://techcrunch.com/2026/07/22/tesl…REDACTED/ and https://www.axios.com/newsletters/axio…REDACTED (techcrunch.com)

Interpretation: TSLA remains a call option on vertically integrated humanoid scale, but this run does not support treating Optimus as deployment-proven or revenue-proven. The action signal is “monitor tooling and real factory-hours,” not “assume humanoid revenue.”

---

Agility / GXO: logistics RaaS remains a benchmark for non-glamorous deployment

GXO and Agility announced a multi-year agreement to deploy Digit robots in logistics operations, with Agility Arc as the cloud automation platform for facility mapping, workflow definition, and operational management. Although the original commercial agreement is older, the fetched GXO investor-hosted PDF appeared in this run and reinforces RaaS deployment framing. Source URLs: https://www.agilityrobotics.com/content/gxo-…REDACTED and https://investors.gxo.com/node/9401/pdf (agilityrobotics.com)

Interpretation: Digit/GXO is less sensational than humanoid race narratives but more relevant as a conversion model: narrow logistics tasks + service contract + deployment platform. GXO is not a pure-play robotics stock, but it is a useful adoption proxy.

---

China / humanoid industrialization: Unitree IPO + UBTECH industrial deployment

ChinaDaily reported Unitree Robotics received regulatory approval to proceed with its Shanghai STAR Market IPO, potentially making it China’s first publicly listed humanoid robotics company. Xinhua reported Unitree passed the STAR Market IPO review on June 1 and expected H1 2026 revenue between RMB 1.052B and RMB 1.128B. Source URLs: https://subsites.chinadaily.com.cn/ezhejiang/2026-07/07/c_1195677.htm and https://english.news.cn/20260601/a855…REDACTED/c.html (subsites.chinadaily.com.cn)

UBTECH announced a strategic partnership with Hitachi on July 13, 2026, saying Hitachi had introduced UBTECH’s Walker S2 into select manufacturing environments for operational testing and application-scenario validation. UBTECH also said more than 1,000 full-sized embodied intelligent humanoid industrial robots were delivered and revenue-recognized during the year, primarily in industrial scenarios. Source URL: https://www.ubtrobot.com/en/about/news/826577556529221 (ubtrobot.com)

Interpretation: China’s physical-AI ecosystem may reach investable public-market form earlier than U.S. private leaders. The risk is source transparency and geopolitical/liquidity access; the signal is that low-cost hardware + state/industrial support is compressing the humanoid timeline.

---

Defense robotics: autonomy is crossing from prototype to procurement

AeroVironment announced a three-year $500M IDIQ contract supporting JIATF-401’s Domestic Shield Program. The company describes itself as deploying autonomous systems, loitering munitions, counter-UAS, space-based platforms, directed energy, cyber, and electronic warfare capabilities. Source URL: https://investor.avinc.com/news-releases/news-release-details/av-a…REDACTED (investor.avinc.com)

Shield AI announced a U.S. Air Force production contract for Collaborative Combat Aircraft mission autonomy on June 17, 2026, saying Hivemind is flying aboard Anduril’s YFQ-44A and that the contract preserves competition by enabling autonomy integration across the broader CCA ecosystem. Source URL: https://shield.ai/shie…REDACTED/ (shield.ai)

Anduril announced on June 17, 2026 that the U.S. Air Force selected it to bring mission autonomy to Collaborative Combat Aircraft, and it framed affordable mass as requiring mission autonomy. Source URL: https://www.anduril.com/news/u-s-…REDACTED (anduril.com)

Anduril also has undersea autonomy momentum: its March 2026 DIU/U.S. Navy XL-AUV announcement says Anduril autonomous undersea vehicles had accumulated over 42,355 km and 6,752 mission hours and that Anduril would complete a long-duration operationally representative Dive-XL demonstration within four months of award. Source URL: https://www.anduril.com/news/diu-…REDACTED (anduril.com)

Interpretation: Defense autonomy is ahead of civilian humanoids on revenue conversion because procurement, geopolitical urgency, and operational tolerance for limited autonomy are stronger. Public exposure is better through AVAV and primes/suppliers; Anduril and Shield AI remain private.

---

AI materials / autonomous labs: “model → matter” is becoming infrastructure, not just discovery papers

Nature Communications Materials published a July 8, 2026 article on managing autonomous materials labs with multi-agent AI, discussing AI/autonomous experimentation and the implications for scientific efficiency. Source URL: https://www.nature.com/articles/s43246-026-01219-5 (nature.com)

Johns Hopkins announced a $20M NSF-backed AIMD-Net materials effort on July 22, 2026, focused on AI-powered materials discovery, shared datasets, metadata standardization, and autonomous materials research. Source URL: https://engineering.jhu.edu/news/buil…REDACTED/ (engineering.jhu.edu)

The University of Chicago announced a $20M NSF grant for PoLARIS, a remotely operated robotic lab with more than 20 autonomous stations across five labs at UChicago and Argonne, calling it a distributed operating system for scientific discovery. Source URL: https://www.computerscience.uchicago.edu/news/remo…REDACTED/ (computerscience.uchicago.edu)

DARPA’s April 29, 2026 RFI sought robotics materials capable of intermixed sensing, adapting, and acting in real time without continuous external computation or communications links. Source URL: https://www.darpa.mil/news/2026/rethinking-robotics (darpa.mil)

Interpretation: AI materials is a real “model-to-matter” pillar, but most signals are still research/federal-infrastructure rather than public-company revenue. Watch for lab automation suppliers, materials informatics startups, defense materials programs, and semiconductor/energy materials pull-through.

---

BCI / human-machine boundary: meaningful progress, but low near-term public equity access

Neuralink’s official pages describe PRIME and CAN-PRIME as investigational medical-device trials for a fully implantable wireless BCI, and CONVOY as a study exploring use of the Link to control assistive devices such as a robotic arm. Source URLs: https://neuralink.com/updates/a-year-of-telepathy/ and https://neuralink.com/trials/device-control/ (neuralink.com)

Synchron’s site says its trial has completed 12 months of safety follow-up under FDA IDE oversight, and its current study page frames device control around Apple-device interaction. Source URLs: https://synchronbci.com/ and https://synchron.com/join (synchronbci.com)

A July 2026 arXiv landscape analysis reported that, as of June 2026, China had five BCI-related products approved by the National Medical Products Administration, including a semi-invasive implantable BCI and non-invasive EEG-based rehabilitation systems. Source URL: https://arxiv.org/abs/2607.07185 (arxiv.org)

Interpretation: BCI remains strategically important to the human-machine boundary but is not yet a clean portfolio engine through liquid public equities. Treat as long-horizon watch unless linked to public medtech, neuromodulation, robotics-assisted rehab, or FDA-clearance milestones.

---

CAPITAL_FLOW_MAP

Confirmed / reported capital flow

Capital-flow interpretation

Capital is bifurcating: very large rounds are going to robot brains and defense autonomy, while smaller but strategically meaningful funding is going into AI materials/autonomous labs and infrastructure robotics. The market is paying for “generalized control layers” before broad humanoid revenue is proven. That creates upside but also bubble risk in private valuations.

---

CONVERSION_LAYER

Conversion path A — Platform picks and shovels

• NVIDIA / NVDA: strongest public platform exposure; Cosmos/Isaac/Jetson/Omniverse stack is being explicitly positioned for physical AI training, simulation, reasoning, evaluation, and edge deployment. Source URLs: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/default.aspx and https://investor.nvidia.com/news/press-release-details/2026/Japa…REDACTED/default.aspx (investor.nvidia.com)
• Conversion evidence: partner ecosystem + model releases + edge deployment framing.
• Bottleneck: robotics revenue materiality may remain small relative to data-center AI.

Conversion path B — Industrial deployment

• Figure/BMW: strongest factory proof point, but Figure is private and BMW is a diversified automaker. Source URL: https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-…REDACTED (press.bmwgroup.com)
• Agility/GXO: logistics RaaS structure; lower glamour but cleaner operational model. Source URL: https://investors.gxo.com/node/9401/pdf (investors.gxo.com)
• UBTECH/Hitachi: China/HK-listed humanoid deployment signal; access and governance risk need separate review. Source URL: https://www.ubtrobot.com/en/about/news/826577556529221 (ubtrobot.com)

Conversion path C — Defense autonomy

• AVAV: current public-contract signal from $500M IDIQ. Source URL: https://investor.avinc.com/news-releases/news-release-details/av-a…REDACTED (investor.avinc.com)
• Anduril / Shield AI: highest private defense-autonomy momentum; not directly public, but they shape valuation and procurement comps. Source URLs: https://www.anduril.com/news/u-s-…REDACTED and https://shield.ai/shie…REDACTED/ (anduril.com)

Conversion path D — Model → Matter / AI materials

• Current conversion is grant/research infrastructure, not immediate public equity revenue. Johns Hopkins AIMD-Net and UChicago PoLARIS are important because they standardize datasets, autonomous stations, and lab automation workflows. Source URLs: https://engineering.jhu.edu/news/buil…REDACTED/ and https://www.computerscience.uchicago.edu/news/remo…REDACTED/ (engineering.jhu.edu)

---

TIME_MAP

0–3 months

• Watch whether Tesla Fremont Optimus production moves from preparation language to measurable units, factory-hours, internal tasks, or external deployment. Current signal is tooling/preparation and expectation of production later in 2026, not volume proof. Source URLs: https://www.axios.com/newsletters/axio…REDACTED and https://techcrunch.com/2026/07/22/tesl…REDACTED/ (axios.com)
• Watch Anduril’s reported $100B valuation round and whether terms confirm or reset defense-tech comps. Source URL: https://www.axios.com/2026/07/24/andu…REDACTED (axios.com)
• Watch Unitree IPO pricing/bookbuilding after China regulatory approval. Source URL: https://subsites.chinadaily.com.cn/ezhejiang/2026-07/07/c_1195677.htm (subsites.chinadaily.com.cn)

3–12 months

• Watch Figure 03 at BMW for repeated, quantified deployment metrics: hours worked, parts moved, downtime, safety incidents, gross margin, and number of robots in production. Current BMW/Figure sources confirm the program but not full economics. Source URLs: https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-…REDACTED and https://www.figure.ai/news/f-03-at-bmw (press.bmwgroup.com)
• Watch Shield AI / Anduril CCA mission-autonomy contract execution and Air Force autonomy milestones. Source URLs: https://shield.ai/shie…REDACTED/ and https://www.anduril.com/news/u-s-…REDACTED (shield.ai)
• Watch AI materials/autonomous labs for supplier pull-through: robotics arms, lab automation, metrology, cloud data infrastructure, and materials IP. Source URL: https://www.computerscience.uchicago.edu/news/remo…REDACTED/ (computerscience.uchicago.edu)

12–36 months

• Expect public-market availability to widen through IPOs/SPACs/listings: Unitree first; possibly private U.S./European leaders later if capital markets stay open.
• Expect robotics revenue to remain uneven: defense autonomy and warehouse/logistics likely monetize earlier than general home humanoids.
• Expect platform companies to capture value before most robot OEMs prove durable margins.

---

BOTTLENECK_INDEX

---

THESIS_STATE

Maintained

• Physical AI remains one of the most important AI-next-leg themes.
• Deployment > demos remains the correct filter.
• Robot-brain companies are strategically important but mostly private.
• NVIDIA is the clearest public infrastructure exposure.

Strengthened

• Defense autonomy is now a more investable near-term conversion path than civilian humanoids because contracts are becoming production/procurement-linked.
• Figure/BMW strengthens the thesis that humanoids first monetize in logistics/material handling inside existing industrial environments.
• China’s public-market path for humanoid robotics is advancing faster than the U.S. pure-play public-market path.

Weakened / caution

• Tesla Optimus should not be portfolio-modeled as near-term volume revenue based on this run. The evidence supports production preparation and long-term option value, not current commercial proof.
• Private humanoid valuations are moving faster than audited deployment economics.

New thesis edge

The physical-AI market may split into:

1. hardware OEM race — risky, capital-intensive, margin uncertain;
2. robot-brain / policy layer — high strategic value, mostly private;
3. simulation/world-model/edge platform layer — public, investable, NVIDIA-led;
4. defense autonomy — contract-backed, investable through AVAV and later private liquidity events;
5. AI materials/autonomous labs — long-horizon model-to-matter infrastructure.

---

STRATEGIC INTERPRETATION

The practical investment thesis is not “buy humanoids.” It is: own the layers that convert physical AI into revenue before general-purpose humanoids are broadly profitable. Those layers are compute/simulation, edge inference, defense autonomy, logistics automation, and eventually autonomous-lab infrastructure.

Humanoids are entering factories, but the bottleneck remains useful work per dollar. BMW/Figure is the most useful civilian signal because it is attached to a real plant and real production contribution. Tesla has much larger theoretical upside, but current evidence is weaker on deployment economics.

Defense autonomy is the most urgent revenue pathway. CCA, counter-UAS, XL-AUV, and loitering/autonomous systems all have procurement demand, national-security urgency, and less need to solve the full human-world dexterity problem.

AI materials and autonomous labs are strategically important for “model → matter,” but the equity signal is still indirect. Watch for public lab automation, metrology, materials software, semiconductor materials, and industrial automation suppliers that start reporting AI/autonomous-lab pull-through.

---

SOURCES

1. NVIDIA — “NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI”
   URL: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/default.aspx (investor.nvidia.com)

2. NVIDIA — “Japan’s Robotics and Manufacturing Leaders Build on NVIDIA Cosmos to Advance Physical AI Frontier”
   URL: https://investor.nvidia.com/news/press-release-details/2026/Japa…REDACTED/default.aspx (investor.nvidia.com)

3. NVIDIA — “NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots”
   URL: https://investor.nvidia.com/news/press-release-details/2026/NVID…REDACTED/default.aspx (investor.nvidia.com)

4. Figure AI — “F.03 Arrives at BMW”
   URL: https://www.figure.ai/news/f-03-at-bmw (figure.ai)

5. BMW Group — “BMW Group advances the use of Physical AI in production with Figure 03 project in Spartanburg”
   URL: https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-…REDACTED (press.bmwgroup.com)

6. BMW Group attachment — Figure 02 production contribution detail
   URL: https://www.press.bmwgroup.com/global/article/attachment/T0458778EN/651004 (press.bmwgroup.com)

7. Figure AI — “Ramping Figure 03 Production”
   URL: https://www.figure.ai/news/ramp…REDACTED (figure.ai)

8. Agility Robotics — GXO multi-year agreement
   URL: https://www.agilityrobotics.com/content/gxo-…REDACTED (agilityrobotics.com)

9. GXO investor-hosted Agility agreement PDF
   URL: https://investors.gxo.com/node/9401/pdf (investors.gxo.com)

10. Tesla AI & Robotics
    URL: https://www.tesla.com/AI?redirect=no (tesla.com)

11. TechCrunch — Tesla Q2 2026 production timeline / Optimus language
    URL: https://techcrunch.com/2026/07/22/tesl…REDACTED/ (techcrunch.com)

12. Axios — Tesla Fremont Optimus production expectation
    URL: https://www.axios.com/newsletters/axio…REDACTED (axios.com)

13. Skild AI — Series C announcement
    URL: https://www.skild.ai/blogs/series-c (skild.ai)

14. TechCrunch — Anduril raises $5B at $61B
    URL: https://techcrunch.com/2026/05/13/andu…REDACTED/ (techcrunch.com)

15. Axios — Anduril seeks $100B valuation
    URL: https://www.axios.com/2026/07/24/andu…REDACTED (axios.com)

16. Reuters via MarketScreener — Humanoid raises $152M Series A
    URL: https://ca.marketscreener.com/news/robo…REDACTED (ca.marketscreener.com)

17. TechCrunch — Physical Intelligence reported raise talks
    URL: https://techcrunch.com/2026/03/27/phys…REDACTED/ (techcrunch.com)

18. AeroVironment — $500M IDIQ for JIATF-401 Domestic Shield
    URL: https://investor.avinc.com/news-releases/news-release-details/av-a…REDACTED (investor.avinc.com)

19. Shield AI — Air Force production contract for CCA mission autonomy
    URL: https://shield.ai/shie…REDACTED/ (shield.ai)

20. Anduril — U.S. Air Force selects Anduril for CCA mission autonomy
    URL: https://www.anduril.com/news/u-s-…REDACTED (anduril.com)

21. Anduril — DIU and U.S. Navy select Anduril for XL-AUV program
    URL: https://www.anduril.com/news/diu-…REDACTED (anduril.com)

22. ChinaDaily — Unitree IPO approval
    URL: https://subsites.chinadaily.com.cn/ezhejiang/2026-07/07/c_1195677.htm (subsites.chinadaily.com.cn)

23. Xinhua — Unitree clears STAR Market listing review
    URL: https://english.news.cn/20260601/a855…REDACTED/c.html (english.news.cn)

24. UBTECH — Hitachi strategic partnership
    URL: https://www.ubtrobot.com/en/about/news/826577556529221 (ubtrobot.com)

25. Nature Communications Materials — autonomous materials labs with multi-agent AI
    URL: https://www.nature.com/articles/s43246-026-01219-5 (nature.com)

26. Johns Hopkins — $20M AIMD-Net materials discovery grant
    URL: https://engineering.jhu.edu/news/buil…REDACTED/ (engineering.jhu.edu)

27. UChicago — $20M PoLARIS autonomous robotic lab
    URL: https://www.computerscience.uchicago.edu/news/remo…REDACTED/ (computerscience.uchicago.edu)

28. DARPA — Rethinking robotics with physical intelligence
    URL: https://www.darpa.mil/news/2026/rethinking-robotics (darpa.mil)

29. UVC Partners — alqem raises €8M for AI-driven materials discovery
    URL: https://www.uvcpartners.com/blog/endl…REDACTED (uvcpartners.com)

30. Gritt — $32.4M physical AI infrastructure robotics launch
    URL: https://www.gritt.ai/news/gritt-launches-physical-ai (gritt.ai)

31. Neuralink — A Year of Telepathy
    URL: https://neuralink.com/updates/a-year-of-telepathy/ (neuralink.com)

32. Neuralink — Device Control trials
    URL: https://neuralink.com/trials/device-control/ (neuralink.com)

33. Synchron — BCI home / trial status
    URL: https://synchronbci.com/ (synchronbci.com)

34. Synchron — Join study
    URL: https://synchron.com/join (synchron.com)

35. arXiv — Clinical Translation of Brain-Computer Interface in China
    URL: https://arxiv.org/abs/2607.07185 (arxiv.org)