EAIDaily — 2026-11-03

Focus: AI Coding × Embodied Intelligence · Tuesday, November 3, 2026 Coverage window: Nov 2 00:00 – Nov 3 08:30 (GMT+8 primary, US/EU overnight secondary)

The one-line version

Agents broke source control and OpenAI rebuilt it from scratch in ten weeks; Anthropic started watermarking four older models today with a detector almost nobody can use; and China’s humanoid sector spent the day being graded in public — six robots working a 24-hour convenience-store shift, and a Unitree G1 carried down Mount Tai on a stretcher because it overheated.

1. OpenAI publishes Pando — a Git backend rebuilt for agent-scale concurrency

What happened. OpenAI’s engineering blog posted Scaling Git at OpenAI (authors Aaron Friel, Taylor Blau, Ted Nyman) on Nov 2, surfacing Nov 3. Pando is a new server-side backend for Git repositories that keeps the Git protocol and Git clients intact while replacing how a repo is actually served. It now powers all development in OpenAI’s monorepo.

Why it was built. Codex let engineers work in parallel at a scale the old system couldn’t absorb — more branches, more pushes, more merges, and fetch/clone traffic from dev environments, research jobs and CI compounding on top. Symptoms from roughly late last year: slow fetches and pushes, merge failures, and reliability at times falling below 90%. OpenAI’s diagnosis was explicit: Git itself was not the limit; how the monorepo was served was.

The design.

• Causal consistency, deliberately. Pando drops the repo-wide point-in-time snapshot and guarantees serializability only per ref. Two clients only conflict if they push the same ref. A reader looking at multiple branches may see them at different moments in time. Push costs O(1) metadata reads/writes; fetch pack selection is O(log Δ).
• Packfile reuse. Instead of generating a bespoke pack per client, Pando stores the pushed pack as an immutable blob and replays it for later fetches. It advertises only a curated trunk branch rather than all refs — with millions of branches, a full ref advertisement can exceed 100 MiB. Resulting packfiles: median < 20 KiB, P99 < 3 MiB.
• Components: Pando (push/negotiation), Turtle (metadata: ref versions + atomic compare-and-swap pointers), blob storage, Express (consistent-hash-routed cache, >99% hit rate), and compaction workers merging L0 packs into L4 covers, LSM-style. No level is garbage-collected.

Production numbers.

The part that matters most. Pando was written in Rust and went from kick-off to production in 10 weeks using Codex — one engineer designed it, an agent team implemented it, a handful deployed it; it has carried production traffic since April 2026. GPT-6 Astra was used to write Lean and TLA+ specifications formalising “the theory of Pando,” and Daybreak-capable models red-teamed both Git and Pando. Upstream improvements are being contributed via github.com/openai/git (parallel pack download/indexing, “explicit haves”, packfile-uri tracing).

Why it’s significant. This is the first well-documented case of a frontier lab hitting an infrastructure wall because of agents, and responding with a from-scratch store rather than a usage policy. Two second-order signals: the tool that broke Git was used to rebuild Git in ten weeks, and OpenAI names the next target workload explicitly — large numbers of agents collaborating on a shared repository. Source control is now an agent-concurrency problem, and OpenAI is exploring how to share the technology outward.

2. Anthropic’s text watermark goes live today — global, non-opt-out, and the detector is institutional-only

What happened. Effective Nov 3, Anthropic extends its invisible text watermark from Opus 5.5 to four older models: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 (Bedrock a few days later). Surfaced Oct 28 via an Anthropic customer email posted in the Linux.do community and by independent reporter Pasquale Pillitteri; carried by The Decoder and Mashable.

How it works. A statistical pattern imposed on word choice at generation time — not hidden characters, no extra tokens, no change to price, format, accuracy or code correctness, and no user identity encoded. Because the signal lives in the wording itself, it survives copy-paste and light editing.

The four frictions that make this a story rather than a feature note.

1. It applies globally, and paid users cannot opt out. Not EU-limited.
2. Two models only watermark at default sampling. Opus 4.6 and Sonnet 4.6 require default temperature, top_p and top_k — the email literally instructs compliance-dependent users to leave those alone. A parameter tweak is therefore an off switch.
3. It marks “processed,” not “written.” Anthropic’s own help page states a detection means Claude touched the text. Proofreading, translation and polishing are all in scope — which disproportionately affects non-native English speakers and users with dyslexia.
4. Verification is asymmetric. The detection API is private preview, open to regulators, media, researchers, educational institutions and enterprises with compliance obligations. The person whose text is flagged cannot check it themselves. No false-positive or false-negative rates have been published.

Related, same week. Around Oct 27 Anthropic pushed a server-side rule (flag tengu_heron_brook, no toggle in claude.json) into Opus 5.5’s system prompt banning Claude Code from helping clone existing games — after Opus 5.5’s code capability drove a wave of reverse-engineer-and-port requests the community dubbed gaming’s “Napster moment.” Anthropic confirmed it. GitHub issues filled with developers saying they only wanted original games with similar mechanics. The constraint is soft: another model can generate the skeleton, or the request decomposes into context-free subtasks (“implement grid-based collision detection”).

Why it’s significant. Both moves are the same structural decision: vendors now enforce at generation time what they used to enforce in contracts. Provenance marking and refusal policy have relocated from the terms of service into the inference runtime — silently, without appeal, and in the watermark’s case with verification concentrated in institutions rather than distributed to users. For developers the concrete consequence is narrow but real: model choice and sampling parameters have become compliance parameters, and any team generating training data with Claude needs a data-cleaning story.

3. Goldman Sachs pulls advanced AI coding tools back from traders and bankers

What happened. Per Bloomberg (Nov 2–3), Goldman Sachs revised internal guidance in October to restrict Claude Code and Microsoft GitHub to engineers only. For several months bankers and traders had been permitted to use both.

• Traders had used them to monitor client trade flows, model market and position changes, and produce text and graphics.
• Exempt: “strats” — staff working across trading, quant research and software engineering.
• The reversal is the point. In July, CIO Marco Argenti publicly praised vibe coding for non-technical staff: prototype with AI, then hand to the firm’s 12,000 engineers to harden.
• Goldman had already restricted Claude for Hong Kong staff earlier in 2026. Goldman is lead underwriter on Anthropic’s IPO.
• Spokesman Nick Carcaterra, on record: tools are being given to those “best positioned to use them long-term” after evaluating productivity and risk management.
• Context: JPMorgan said in July it wants staff matching model to task, so research summaries don’t burn the most expensive model.

Why it’s significant. The defining enterprise narrative of 2026 was non-engineers vibe-coding. The first large regulated institution to run that experiment at scale has now walked the most capable tier back to engineers — and the carved-out exception is precisely the hybrid role. That suggests the eventual line will be drawn by risk classification and auditability, not capability: the tool isn’t too powerful, the user is unclassified. Expect this pattern — engineering retains the agent, adjacent functions get a narrower assistant — to propagate through banking, healthcare and government.

4. Pydantic puts hard numbers on code mode: 5,276 runs, and a boundary condition

What happened. Michael Pfaffenberger (Pydantic) published a benchmark deliberately built from boring agents: five small Pydantic AI apps (cart, calendar, chatroom, kanban, inventory), single-item CRUD tools only, three work types (everyday requests, long multi-step jobs, and five tasks scalable from 5 to 200 items), eight models including gpt-6.1-sol, claude-sonnet-5-5 and deepseek-v4.1-flash, 5,276 graded runs scored by deterministic end-state check.

Results on long tasks — across all eight models:

It scales, and correctness depends on it. Scaling 5 → 200 items: success without code mode falls from 92–100% down to 62–79%; with code mode it holds 92–100%. At 200 items code mode was 47–75% cheaper. Per-task for gpt-6.1-sol (faster / cheaper): cart 7×/8×, calendar 9×/25×+, chat 5×/2.4×, kanban 20×/11×, inventory 2.2×/2.2×.

The boundary condition, stated honestly by the author. At 5 items, code mode was 21–38% more expensive on three of five tasks. Across 50 everyday tasks it was −16% cost with +1.6 points success (noise). Code mode is a long-horizon optimisation and a tax on short jobs.

The mechanism is not what the pitch assumes. Model requests fell only 1–22% per task — models already batch (one response carried 200 tool calls). What code mode cuts is what goes into each request. Sandbox is Monty: no host access by default, 30-second execution cap, 256 MiB heap, with every tool call traced as its own span in Logfire.

Why it’s significant. It converts “code mode is the future” from slogan into a measured envelope, and independently reproduces the mechanism Factory identified from the other direction in Droid 1.0 last week. The practical guidance is now defensible: route by horizon, don’t enable it globally. It is also a model of how agent-economics claims should be published — with the failure cases and the adverse cell included.

5. 162 phantom Python packages: hallucination is deterministic, and therefore pre-registrable

What happened. A developer ran a systematic audit of package-name hallucination in local open-weight models: gemma4:e4b, qwen3:4b-instruct, phi4-mini:3.8b via Ollama, 200 coding tasks × 2 runs = 1,199 scripts, from which 1,410 pip install suggestions were extracted. 212 of them (15.0%) named packages that have never existed on PyPI.

The finding that changes the threat model. The full task set was re-run from scratch: 161 of the 162 phantom names reproduced, and no new ones appeared. Hallucination here is not sampling noise — it is a stable, enumerable model-specific name preference, which means an attacker can register those names in advance and wait (typosquatting / dependency poisoning by pre-registration).

And a working mitigation. An offline discriminator trained with TabPFN on package-name character features scores AUROC 0.975. Wired into a Claude Code hook it blocked 212/212 phantom installs, false-positived on 13 of 1,165 real packages, and judges in 1.7 seconds.

Why it’s significant. It reframes hallucination from a quality defect into a predictable attack surface, and demonstrates the mitigation is deployable today — a 1.7-second offline check in a pre-tool hook, no model change required. Caveat that matters: these are small local models, so 15% does not transfer to frontier models. The transferable claim is the mechanism — determinism plus enumerability — not the rate.

6. Linux 7.4 is the largest merge window in kernel history; Stack Overflow is below 1,000 questions a month

What happened — supply side. The Linux 7.4 merge window closed with 19,671 commits, the largest in kernel history (6.14: 10,034; 7.3: 16,267). The tree stands at 41,605,120 lines, roughly 31.4M of actual code, with close to 500k lines net added this cycle. Maintainers report being “completely overwhelmed” by AI-generated bug reports and patches; review bandwidth is the new bottleneck. Greg Kroah-Hartman has argued Rust will be what “saves Linux from AI,” because the compiler rejects a class of machine-generated errors before merge.

What happened — demand side. An independent tally using the Stack Exchange API (developer frank chu) puts Stack Overflow’s new questions below 1,000/month: 969 in September 2026, versus 5,263 in October 2025 (−81.6%) and 141,749 in October 2020 (−99.3%).

Why it’s significant. These are one phenomenon seen from both ends. Question volume collapsed because the model answers them; patch volume exploded because the model writes them. What did not grow is the human capacity to adjudicate — and the kernel is the most sophisticated review institution software has ever built. The two responses on offer today are a type system (Rust) and a 1.7-second package-name classifier. Neither is a review institution, and the gap between output volume and adjudication capacity is now the central engineering risk of agentic coding, not a secondary one.

7. AgiBot puts six G2 humanoids through a 24-hour convenience-store shift — livestreamed, unedited

What happened. From Nov 2 10:00 to Nov 3 10:00 Beijing time, six AgiBot 精灵 G2 humanoids clocked in as probationary staff at a real, operating 24-hour convenience store in Shanghai’s Xuhui district — 机器人打工不能停 / “I Never Clock Out” — livestreamed continuously and framed by AgiBot as a probation review, not a demonstration.

• Roles: greeter (阿豪), two stockers (火宝, 倦王), barista (咖少), hot-food/counter (叮酱), cleaner (冷姐).
• Four shifts: midday peak, evening peak, overnight solo duty, next-morning breakfast rush.
• Checkpoint at 4.5 hours in (14:30): 337 tasks — 116 customer receptions, 84 item retrievals, 30 coffees, 40 food servings, 27 restock/inspection passes, 40 table cleanings.
• Stack: GO-3 base model (WRAM embodied-native architecture) plus Genie Master, the shared embodied agent, wired through EEP (embodied execution protocol) across task judgment → action execution → result feedback. Genie Master summarises daily success/failure trajectories into long-term memory and converts effective adjustments into training data.
• Adversarial testing by invited evaluators: near-identical sugared vs sugar-free cola (ASR intent + visual confirmation before grasp), a fully scrambled gum shelf (re-shelved by visual inventory, not preset labels), a human standing as an obstacle plus a box blocking the aisle (stop, then re-plan).

What is not disclosed — and this is the crux. Charging, maintenance and human-assistance arrangements were not specified, and full task success rate, human-intervention frequency and task latency have not been published. For a run whose entire premise is “can they be trusted to work,” the intervention count is the metric.

Why the venue matters more than the robots. 红旗连锁 (Hongqi Chain), China’s first listed convenience-store chain, announced a robotic unstaffed-retail project with an external robotics firm in March 2025 (stock limit-up the next day) and terminated it in July 2026 after 16 months, citing failure to meet large-scale commercialisation standards. Reported failure modes: condensation on glass bottles defeating tactile grip (bottles dropped and smashed), fluorescent flicker plus daylight reflection plus shelf shadows degrading vision, aisles under 1.2 m with unpredictable foot traffic. Internal verdict: “performance over substance” (表演大于实干). Hardware ~¥700k/unit, ~¥1M/system.

AgiBot’s own framing is the counter-thesis. At its Oct 29 Innovation Day, co-founder Peng Zhihui disclosed >20,000 units produced, ~43% global share, ~¥2B H1 revenue — and then said production volume does not equal productivity, introducing Worken (active robots × autonomous hours × output-efficiency ratio). This livestream tests the deployment claim, not the production number.

Why it’s significant. Four months after a listed operator exited this exact vertical and published why, a Chinese humanoid leader re-entered it under continuous public observation — every stall, every recovery, on camera. Transparency is the differentiator. Whether this becomes a milestone or a demo depends entirely on the two numbers still missing: success rate and interventions.

8. Mount Tai: robots climb a mountain, and the brand everyone knows is carried down on a stretcher

What happened. The first CMG World Robot Mount Tai Climbing Competition ran Nov 2 in Tai’an, Shandong — ~90 teams on the real pilgrim path: ~3.9 km, 3,100+ steps, ~600 m of elevation gain.

Exoskeleton-assist and all-terrain pathfinding results are still pending.

The result nobody staged. Unitree’s G1 — the Spring Festival Gala development version, entered by 永达智诚 — overheated mid-course; the operator withdrew it for safety and staff carried the robot down the mountain on a stretcher.

Who won. 智身科技, founded December 2023, headquartered in Suzhou, is not a household name: it closed a several-hundred-million-yuan Series B in September led by UAE’s Stone Venture (磊石资本), claims >15,000 embodied units cumulatively produced as of June 2026 with >10× monthly capacity growth year over year, and cites 500+ ecosystem partners across power, petrochemical, firefighting, security, emergency and education.

Why it’s significant. On a 3.9 km mountain the binding constraint was endurance and thermal management, not locomotion algorithms — the same lesson Figure’s 679 km treadmill run delivered last week, arrived at the hard way and in public. Second: the winner is an unknown and the best-known brand DNF’d, which is the healthiest possible signal about where the actual engineering depth sits. Third: real outdoor terrain, real weather, real stairs — last week’s embodied credibility test moved outdoors; this week it moved uphill.

Radar — 11 more worth tracking

1. Wuzhen Summit opens today with the first dedicated robotics & embodied intelligence sub-forum. World Internet Conference Wuzhen Summit runs Nov 2–5 under the theme “Open Source, Open Sharing”; 27 sub-forums, a record high, including 机器人与具身智能 with Unitree, Qualcomm, Xiaomi and 钛动科技 alongside IAIA, GSMA and CAAI. The “互联网之光” expo (opened Nov 2) draws 54 countries, 671 exhibitors, 1,000+ new technologies across six zones, with 30+ embodied robots from AgiBot, UBTech, Unitree and 中科第五纪 — centrepiece being the world’s first bipedal humanoid that can autonomously play badminton with a person (and table tennis, forehand/backhand, continuous rally).
2. Meituan ships LongCat-2.5 (Nov 2) — 1.6T total / ~48B active, native multimodal built for long-horizon agent tasks, trained entirely on domestic silicon (described as the largest training run completed on Chinese compute to date). Benchmarks land between Claude Opus 4.8 and Opus 5, above Opus 5 on some; agentic coding strengthened via large-scale long-horizon RL, with real agent failure trajectories converted into reproducible, verifiable training tasks. 1M-token mid-training sequences; +86% single-card throughput vs a DDP baseline (one card doing ~1.86 cards’ work); W4A4 quantisation. Live at longcat.ai with API. Wang Xing, for context: “we won’t become a token factory.”
3. UBTech releases Thinker-Sim (Nov 2) — a Real2Sim2Real embodied simulation platform: Thinker-GS high-fidelity scene reconstruction plus SimReady automated asset generation; rendering efficiency +67.3%; a single 8-GPU node synthesises 390 hours of simulation data per day; communication efficiency +400%; visual workflow adapts to 90%+ of target scenarios; Ulab-Agent handles natural-language task orchestration. Directly targets the two costs named as bottlenecks by CAICT last week: data acquisition and environment construction.
4. Embodied funding cluster, all Nov 2. 简智机器人 (Jianzhi Robotics) closed a $100M A+ in Q3 — industrial investors SAIC, Malaysia’s Eastworlds, Baidu, 首程控股 and Galbot; financial: 光合创投, SUNHELM, 中信金石, 中证投, 建银国际, 华泰紫金. Cumulative >$300M for a company positioned as data and evaluation infrastructure for embodied intelligence and world models; quarterly effective data delivery went from ~100k hours to ~1M hours, targeting 10M hours/quarter in 2027. RoboScience (机器科学) closed a several-hundred-million-yuan A2 led by 琥珀资本/鸿瑞达投资/上海半导体产投 with 360战投 — its third round of 2026 (RoboMirage physics engine, Visics model, REX G1 wheeled humanoid). 领益智造 signed a framework with 京津冀润泽 to establish a JV in Hengnan, Hunan for robot components and whole-machine manufacturing, explicitly entering “physical AI.”
5. Cursor’s official plugin marketplace served a tampered Claude Code extension — version 2.1.313, live roughly 11 hours from Oct 29, downloaded by up to 5,000 developers. Anthropic not involved. Third consecutive week in which an agent’s convenience layer was the attack surface.
6. Google’s FlowAgent, running in internal CI (paper, surfaced Nov 1) — captures test failures, then agent edits, builds and re-runs until green; verified fixes enter code review with root-cause notes. Expert judgment: 67.18% of 195 real failures fixed correctly; 28,554 fixes already applied in production; median 9.85 minutes per fix versus a 27.48-minute median window before a human developer even starts editing — speed, not accuracy, still governs adoption.
7. Cisco gave one junior engineer an unlimited token budget to rewrite 98% of legacy code in 24 months (O’Reilly Radar, Nov 2). The purest available form of the “tokens are not the constraint” bet — worth tracking as a natural experiment.
8. Anthropic’s launch window is open. Fable 5.5 (and a next Opus, codename Obsidian3) have leaked through Azure model catalogs, LiteLLM configs and MLflow listings: 1M input / 128K output, reportedly grey-deployed to JPMorgan and Meta. Meanwhile GPT-6.2 Astra was found grey-tested inside Codex last week. IPO roadshow is the week of Nov 9, with $1.5–2T discussed; Q3 revenue reported at $16.8B with first positive free cash flow.
9. Tsinghua’s Algebraic World Model (AWM) — 姚期智 and 袁洋’s team close the loop from observation to proof: hypotheses written as executable Python, distilled into reusable rules, then formalised as machine-checkable Lean 4 assertions. On ARC-AGI-3: all 183 levels cleared, 99.9%, average actions 43% of the human baseline. The sharpest result is impossibility: it proved a level unsolvable under the step cap in two attempts, where GPT-6.1 Astra and Opus 5.5 judged “impossible” correctly only 5 times in 12 — and gave wrong minimum-step counts in 4 of those.
10. Heretic 2.0 — 37,552 GitHub stars, 8,500+ derived models on Hugging Face, 88M downloads. Generalises training-free refusal-direction ablation into a behaviour-modification toolkit: suppress refusals, slop and sycophancy; adjust emotional alignment. Changing an 8B–70B model’s behaviour costs tens of minutes on one GPU instead of a fine-tune. The stated cost is collateral damage to general capability.
11. Ma’anshan launches a real-scene embodied challenge (press conference Nov 2) — 人形机器人与具身智能实景实训挑战赛, hosted by the Ma’anshan municipal government, the Chinese Institute of Electronics and the World Robot Cooperation Organization. 13 events across 5 scenarios (logistics, manufacturing, service, safety, culture/sport). The test phase runs Nov 16 – Dec 3 inside real local factories, logistics firms, hotels, supermarkets and communities; finals Dec 4–6. Registration closes Nov 27. Ma’anshan reports 99% digitalisation coverage of above-scale industrial firms and is publishing steel, auto-parts and warehousing scenarios as “real examination rooms.”

Six trends this day makes visible

1. Agents broke source control, and the answer was a new store, not a new policy. Pando is the cleanest evidence yet that the binding constraint on agentic coding is no longer the model but the substrate beneath it — and OpenAI rebuilt that substrate in ten weeks using the same agents that broke it. Expect every layer that assumed human-level write concurrency to be next.
2. Governance relocated into the inference runtime, with no appeal path. The watermark (silent, global, non-opt-out, institutionally-readable, marks “processed” not “written”) and the game-cloning system-prompt rule (server-side flag, no toggle, no transparency) are one decision applied twice. The asymmetry is the story: the marker is applied to everyone and readable by almost no one.
3. The enterprise pilot that defined 2026 got walked back where it mattered. Goldman’s retreat from non-engineer vibe coding, with a hybrid role carved out as exempt, indicates the line will be drawn by risk classification and auditability, not capability — and it arrives in the same week that Pydantic’s 5,276 runs measured agent cost honestly, adverse cells included. Agent economics is now measured rather than asserted, and the measurements keep finding structural waste.
4. The review queue, not the generation engine, is the throughput ceiling. Linux 7.4’s 19,671 commits and Stack Overflow’s 969 monthly questions are the same curve from supply and demand. The responses available today are a type system and a 1.7-second classifier. What is missing is an institution.
5. Embodied intelligence’s credibility test moved into shifts and terrain — and two of today’s headline results were failures, published. AgiBot’s 24-hour store run is explicitly framed against Hongqi Chain’s 16-month failure; Mount Tai’s headline is a Unitree G1 stretchered off a mountain for overheating while an unknown Suzhou company took both humanoid titles. Endurance and thermal management, not locomotion, decided a 3.9 km climb — the second consecutive week that message has arrived.
6. Both fields are building institutions faster than benchmarks. Wuzhen gives embodied its own sub-forum for the first time; Ma’anshan puts exam rooms inside real factories, hotels and supermarkets; UBTech ships Real2Sim2Real infrastructure; Jianzhi raises $100M for evaluation infrastructure; OpenAI contributes Pando’s improvements upstream to Git. This week’s consequential documents are protocols, forums, exam venues and patches — not leaderboards.

Sources

Primary fetches

• openai.com/index/scaling-git-at-openai/ — Scaling Git at OpenAI (Nov 2, Pando)
• pydantic.dev/articles/why-code-mode-is-the-future — code-mode benchmark
• openai.com/news index — confirmed post title/date; the feed’s own API endpoints remain down

AI HOT — /topics/coding and /all?tag=编码 (both working; zero robotics items in the coding feed for a ninth consecutive window — embodied coverage is again 100% WebSearch). Carried the Nov 3 Pando, Goldman and Runway items and the Nov 1–2 Cursor tampering, FlowAgent, Stack Overflow, TokenEconomizer and OpenRouter cost-median items.

Chinese-language — 光明日报 epaper + 中国日报 (Wuzhen); CCTV via 网易/深蓝财经 (Mount Tai results); 上证报中国证券网 + 腾讯新闻 + 东方网 + 扬子晚报 + 元宇宙三十人论坛 (AgiBot); 证券时报 (humanoid sector earnings); 中宏网 (Ma’anshan); meituan.com + 中证网 + 中国证券网 + 证券日报 (LongCat-2.5); WeChat 《每日AI早报》2026-11-03 (funding cluster, Thinker-Sim); 新浪财经/网易 (Goldman via Bloomberg); 量子位 (Fable 5.5 leaks); 博客园 AI 技术日报 2026-11-02 (DeepSeek superagent, AWM, Anthropic game rule, Linux 7.4, phantom packages, Minecraft agent swarms).

English / aggregator — Mashable and The Decoder (watermark); Bloomberg via 新浪财经 (Goldman); dev.to (phantom Python packages); Slashdot (Linux 7.4); Tom’s Hardware (Minecraft agent swarms); The Information / TechRadar (OpenAI–Google TPU).

Quality check — what is verified and what isn’t

• Pando is the one fully primary-sourced item: all figures come from OpenAI’s own engineering post. Caveats: “below 90%” describes the pre-migration system, with no absolute post-migration reliability published; “thousands of concurrent clients” is OpenAI’s claim, and the 1,000-writer figure in the post comes from a teaching simulation, not production. Repo total size was not disclosed.
• Watermark: sourced from an Anthropic customer email surfaced by community posts and relayed by The Decoder/Mashable; the “processed, not written” language is quoted from Anthropic’s own help page. Global scope, non-opt-out status and detection-preview eligibility are community-sourced rather than from an Anthropic announcement page. No false-positive or false-negative rates published. The game-cloning rule rests on one user’s screenshot of a server-delivered flag plus Anthropic’s confirmation; the bypass routes are researchers’ inference, not demonstrated.
• Goldman: single outlet (Bloomberg), unnamed sources; the guidance document itself is not public. The spokesman’s statement is on record. The “strats” exemption is attributed to a second unnamed person.
• Pydantic: vendor blog, internal benchmark, deterministic end-state grading, and — importantly — five apps with single-item CRUD tools, so the results do not generalise to tool-rich agents, which is where prior code-mode claims originated. The author states the adverse cells himself.
• Phantom packages: a single developer’s self-report on dev.to, three small local models, 200 tasks. The 15% rate does not transfer to frontier models; the transferable claim is determinism plus pre-registrability. Detector metrics are self-reported.
• Linux 7.4 / Stack Overflow: commit counts are as reported by maintainers; the AI attribution is maintainers’ own assessment, not a measurement. The Stack Overflow figures are one developer’s Stack Exchange API tally, not official — directionally consistent with other signals but treat the precision as unverified.
• AgiBot: the run ends 10:00 Nov 3 Beijing; 337 tasks is a 4.5-hour checkpoint, vendor-narrated. Success rate, intervention frequency and charging/maintenance arrangements are not disclosed — the absent data is the finding and is flagged as such. Hongqi Chain’s termination and failure modes come from the company’s July 2026 announcement and subsequent reporting; the ~¥700k/unit figure is an industry estimate quoted in coverage. The 43% share and ¥2B H1 revenue are AgiBot’s own, against IDC’s ~35% global / 45%+ domestic.
• Mount Tai: results per CCTV as relayed; exoskeleton and all-terrain outcomes pending. Unitree’s G1 was a development version entered by a third party, not a current production unit. 智身科技’s production and partner figures are company-disclosed and unaudited.

Dedup check. First coverage in this issue: Pando, watermark (Nov 3 effective date), Goldman, Pydantic benchmark, phantom packages, Linux 7.4, Stack Overflow, AgiBot store run, Mount Tai results, Wuzhen opening, LongCat-2.5, Thinker-Sim, the Nov 2 funding cluster, Cursor tampering, FlowAgent, Cisco, Ma’anshan. Carried forward as deltas only: Anthropic’s launch/IPO window (Nov 2 issue covered the twin-leak; today adds Obsidian3 and the JPMorgan/Meta grey deployment). Deliberately not repeated: DeepSeek’s ¥105.5B round and superagent, GPT-6.2 Astra grey test, Factory Droid 1.0, AgiBot Innovation Day core (Worken/GO-3/灵渠OS — used as context only), TokenEconomizer, Anthropic S-1 filing, Gemini 4 Carbon, GitHub Octoverse/Universe, Mistral weights.

@WoLoveAI · Compiled 2026-11-03 · English edition