Monday's tape prints the Codex-third-day-Latest-hold Monday-reset: the day after Sun Oct 18's second dark-weekend harness day, Mon Oct 19 2026 prints the Codex-staging-extended-into-a-third-day tape. On the Codex tape, OpenAI Codex rust-v0.164.0 is not promoted to Latest on Sun Oct 18—the compare endpoint between rust-v0.163.0 (Thu Oct 15 21:30 UTC Latest) and main now reports 162 commits across 56 contributors and 1,181 files changed (up from Sat Oct 17 EOD's alpha.3 compare at 154 / 51 / 1,035), so Sun adds 8 commits + 5 contributors + 146 files to the staged change set without shipping a fourth pre-release alpha; rust-v0.163.0 remains the Latest-tag coding-agent for a third consecutive day, past the 72-hour mark. On the Claude Code tape, Anthropic ships Claude Code v2.1.303 at Sun Oct 18 22:26 UTC as the thinnest-note Latest cut of the quarter (release-note body: “Bug fixes and reliability improvements”), closing a ~48-hour first-party-harness-dark weekend at EOD. On the open-weights-Thursday tape, three Oct 15 open-weights drops anchor the week's lab-wave: StepFun Step 5 Preview publishes a 600B sparse MoE / 27B active / 1M context open-weights model aimed at long-horizon agentic execution and fans out across DeepInfra, OpenRouter, Vercel AI Gateway, Fireworks (FP8), Featherless and Haimaker; NVIDIA expands the Nemotron family with Nemotron 3.5 Super VL (120B / 12B active vision-language, 1M context, 26.5 on Artificial Analysis as the top US open-weight model) and Bolt 74B-A9B (24GB native 2-bit quantization for RTX laptops) — both on Hugging Face + OpenRouter + ModelScope + NIM; and Multiverse Computing — in collaboration with NVIDIA — ships Comet 60B, a CompactifAI-compressed version of Nemotron 3.5 Super VL that halves the parameter count to 60B while keeping the 12B active and three modalities and staying within 5% accuracy on most benchmarks. On the Fri Oct 16 enterprise-agent-liability tape, Charles Schwab CEO Rick Wurster tells Bloomberg Television that Schwab is blocking unauthorized AI agents, including Meta's Muse, that try to move client money — weekly blocks run in the “tens and twenties”, and Schwab points third-party AI tools at read-only MCP servers and its own Charley assistant; JPMorgan CEO Jamie Dimon says on the earnings call that he isn't going to allow any “outside agent to come in inside the company”. Also Fri Oct 16, China's Ministry of Commerce + Ministry of Science and Technology publish a draft export-control list adding 40 technologies — 23 of them AI items covering large language models, agents, chip designs and computing clusters, with model weights and source code captured above performance thresholds, API, website, app and SaaS access exempted, open-source and public-domain carved out, interim pre-export registration already in effect, and the consultation open through Nov 15. Throughline: Mon Oct 19 is the Codex-third-day-Latest-hold Monday-reset — the week's lab-wave backbone (StepFun Step 5 Preview + Nemotron 3.5 Super VL + Comet 60B) lands on Thu Oct 15, the enterprise-agent-liability wave (Schwab + JPMorgan + MOFCOM) lands on Fri Oct 16, the first-party-harness-dark weekend consumes Sat Oct 17 – Sun Oct 18, and the only cadence events across that 48h window are DeepSeek V4.1-Flash-1017 + Codex alpha.3 on Sat and Claude Code v2.1.303 bug-fixes-only Latest at Sun 22:26 UTC.

Two reads. (1) A closed-weights Latest-tag coding-agent that holds rust-v0.163.0 for 72+ hours running across Thu Oct 15 EOD → Sun Oct 18 EOD, growing the staged main change set to 162 commits across 56 contributors and 1,181 files (+8 / +5 / +146 vs Sat EOD) without shipping a fourth pre-release alpha, is the operative signal that the honest 2026 coding-agent counter-position has moved from “Fri Oct 16 is one staging day and Sat Oct 17 is a second staging day with three empty-note alphas” to “Sun Oct 18 is a third staging day with no alpha and no Latest promotion, so rust-v0.163.0 is the Latest-tag coding-agent for 72+ hours running through the Monday-reset tape, longer than any Latest-tag hold on the rust train since 0.163.0 was itself promoted”. The +8-commits-and-+146-files-on-a-Sunday tell is the operative release-shape signal — continuing to land commits on main across a Sunday without shipping an alpha and without promoting Latest, rather than letting the stage freeze at Sat EOD's 154 / 51 / 1,035 or shipping a Sunday alpha.4, is the posture a coding-agent team takes when it has decided “the 0.164 Latest-stable will be one atomic promotion whenever it is ready, and the alpha train is only staged on days when a cut is required, so a quiet Sunday is a quiet Sunday even on the staging branch”, and anchors the Codex surface on “rust-v0.163.0 is still the Latest-tag coding-agent at Mon Oct 19 start-of-week; 0.164 is staged behind three empty-note alphas and 8 more Sun-merged commits across a 72h+ Latest-tag hold”. (2) The no-fourth-alpha-across-Sun-Oct-18 tell is the operative weekend-shape signal — on the Sun Oct 18 that follows the Sat Oct 17 third-empty-note alpha, Codex ships no new pre-release and no stable promotion, is the shape a coding-agent team takes when it has decided “the alpha train is a daily cadence only when a cut is needed; a quiet Sunday is a quiet Sunday”. Landing on the Mon Oct 19 reset slot after Sat Oct 17's alpha.3 (154 / 51 / 1,035) and the Sun Oct 18 harness-dark window broken only by Claude Code v2.1.303 (item 02), the Codex 72h+ Latest-tag hold becomes the reference “the closed-weights Latest-tag coding-agent holds rust-v0.163.0 for a third consecutive day while 8 commits + 146 files land on main without an alpha and without a Latest promotion, extending the stage to 162 / 56 / 1,181 across the open-weights Thursday + enterprise-agent-liability Friday + harness-dark weekend as a single 72h+ reading-the-tape window” primitive every subsequent Cursor Agent, Cognition Devin, Google Antigravity, Replit Agent and Perplexity Comet weekend-cadence print now has to price against.

Two reads. (1) An agent client that ships a Latest cut with the single-line release note “Bug fixes and reliability improvements” on Sun Oct 18 at 22:26 UTC, 49h33m after v2.1.302's 24-bullet release (Fri Oct 16 20:53 UTC), is the operative signal that the honest 2026 agent-client-cadence counter-position has moved from “a Claude Code Latest cut is a feature-named release with 3-10 changes every 24-72 hours” to “Claude Code can also ship a single-line Latest cut on a Sunday night to close a 48h first-party-harness-dark weekend without widening the surface, and that minimal-note posture is itself a reading-the-tape signal”. The thinnest-note-of-the-quarter + Sun-22:26-UTC tell is the operative release-shape signal — shipping a one-line bug-fix Latest at Sun EOD, rather than letting the weekend-dark window extend into Monday or batching the fixes into the next feature cut, is the posture an agent client takes when it has decided “a quiet weekend ends with a quiet Sunday-EOD Latest cut, not a Monday-morning feature drop, so operators who pulled v2.1.302 Friday get Sunday's reliability deltas before start-of-week”, and anchors the Claude Code surface on “v2.1.303 is a Sun EOD stability cut, not a feature cut; v2.1.302 remains the last feature-named release”. (2) The Sun-EOD-closes-the-dark-weekend-window tell is the operative weekly-rhythm signal — v2.1.303 at Sun 22:26 UTC closes the ~48-hour first-party-harness-dark window (Fri Oct 16 22:11 UTC's langgraph==1.2.16 → Sun Oct 18 22:26 UTC = 48h15m), is the shape the first-party agent-client ecosystem takes when it has decided “the client-stability layer closes the weekend-dark window on Sunday night, even if it means a bug-fix-only Latest, so Monday starts on the current stability tag”. Landing on the same Sun Oct 18 as the Codex third-day Latest-tag hold (item 01) and in the 48h+ first-party-harness-dark weekend (item 03), the Claude Code v2.1.303 bug-fix Latest becomes the reference “the agent-client ships the thinnest-note Latest cut of the quarter at Sun EOD to close a 48h+ harness-dark weekend window before the Monday-reset tape” primitive every subsequent Cursor, Zed, Windsurf, Continue and Aider Sunday-cadence print now has to price against.

Two reads. (1) A 48h15m first-party-harness-dark window — across Pydantic AI, LangGraph, Vercel AI SDK, langchain-core, Microsoft Agent Framework, openai-agents-python and OpenAI Codex Latest — broken only by a Codex pre-release alpha, a DeepSeek lab-side open-weights drop, and a Sun-EOD Claude Code bug-fix Latest is the operative signal that the honest 2026 agent-harness-cadence counter-position has moved from “every weekday is a cadence day for at least one dominant OSS harness and even Saturdays print a bugfix tag” to “a Fri cadence day consumes the week's harness churn, so the Sat Oct 17 + Sun Oct 18 pair is a 48h15m first-party-harness-dark weekend window broken only by the Codex pre-release alpha train, the DeepSeek lab-side open-weights wave, and the Sun-EOD Claude Code bug-fix cut”. The only-three-cadence-events-in-48h tell is the operative weekend-shape signal — DeepSeek + Codex alpha + Claude Code bug-fix are the only cadence surfaces moving across Sat-Sun, is the posture the ecosystem takes when it has decided “first-party harness Latest cuts cluster on the Thu lab-wave + Fri client-and-harness-wave days, and the weekend is reserved for open-weights lab drops, closed-weights coding-agent staging, and a Sun-EOD client bug-fix close”, and anchors the weekend reading on “first-party-harness dark; Codex alpha.3; DeepSeek 1017; Claude Code 303”. (2) The dark-window-closes-at-Sun-22:26-UTC-not-at-Mon-00:00-UTC tell is the operative weekly-rhythm signal — the dark window is closed by Claude Code's thinnest-note Latest cut 1h34m before midnight UTC, is the shape the first-party harness ecosystem takes when it has decided “the client-stability layer closes the weekend-dark window on Sunday night so Monday opens on the current stability tag, not on Friday EOD's tag”. Landing in the Sat Oct 17 – Sun Oct 18 slot between the Fri Oct 16 client-and-harness cadence day and the Mon Oct 19 lab-platform-and-coding-agent next-cadence-day, the 48h15m first-party-harness-dark window becomes the reference “the dominant OSS agent-harness ecosystem goes dark for 48h on the weekend after a Fri cadence day, broken only by a Codex pre-release alpha, an open-weights lab's weekend-cadence model drop, and a Sun-EOD agent-client bug-fix close” primitive every subsequent LangChain Agent-Middleware, LlamaIndex, DSPy, Haystack and Semantic Kernel weekly-cadence read-off now has to price against.

Two reads. (1) A Chinese open-weights lab that ships a 600B sparse MoE / 27B active / 1M context open-weights model tuned for long-horizon agentic execution with repeated tool calls under stepfun-ai/Step-5-Preview on the Thu Oct 15 open-weights Thursday, routed through DeepInfra + OpenRouter + Vercel AI Gateway + Fireworks (FP8) + Featherless + Haimaker at ~$1 per 1M input tokens, is the operative signal that the honest 2026 open-weights-lab counter-position has moved from “Chinese open-weights MoE labs are DeepSeek + Alibaba Qwen + Moonshot Kimi + Z.AI and StepFun is the Agent-OS lab from the WAIC phone demo” to “StepFun opens a 600B sparse MoE / 27B active / 1M context open-weights model under stepfun-ai/Step-5-Preview on the Thu lab-wave day, routed through the full third-party inference-provider stack at $1 per 1M input tokens, so the open-weights Chinese-MoE agent-execution surface now carries a StepFun layer alongside DeepSeek, Qwen, Kimi and Z.AI”. The long-horizon-agentic-execution + finance-strong tell is the operative training-shape signal — tuning a 600B MoE for repeated tool calls, software engineering, and finance-leaning professional-knowledge work, rather than for single-turn chat, is the posture an open-weights lab takes when it has decided “the open-weights MoE frontier is multi-step tool-use in professional domains, not conversation”, and anchors the StepFun surface on “Step 5 Preview is a 600B / 27B / 1M MoE under stepfun-ai/Step-5-Preview tuned for long-horizon agent execution”. (2) The Thu-Oct-15-open-weights-Thursday tell is the operative release-cadence signal — landing on the Thu lab-wave day alongside Anthropic's Independent Evaluators Tier framework (previously covered) and the week's Fri Oct 16 client-and-harness cadence day, is the shape an open-weights lab takes when it has decided “the Thu lab-wave day is where the week's major open-weights-cadence model drops; the Fri cadence day is for clients and harnesses; the Sat weekend is for Pareto-improvement checkpoints like DeepSeek-V4.1-Flash-1017”. Landing on the same Thu Oct 15 as the NVIDIA Nemotron 3.5 Super VL + Bolt open-weights drops (item 06) and the Multiverse Computing Comet 60B compression (item 07), the StepFun Step 5 Preview open-weights drop becomes the reference “the Chinese open-weights 600B MoE / 27B active / 1M context agent-execution model lands under stepfun-ai/Step-5-Preview on the Thu Oct 15 open-weights Thursday backbone alongside NVIDIA's US open-weights VL leader and Multiverse's CompactifAI-compressed derivative, so a single Thursday carries three open-weights-anchor drops across three continents” primitive every subsequent Alibaba Qwen4, Moonshot Kimi, Z.AI, Mistral and Meta Muse open-weights-cadence print now has to price against.

Two reads. (1) A Chinese open-weights MoE that lands on the full US-and-global third-party inference-provider stack on day zero — DeepInfra + OpenRouter + Vercel AI Gateway + Fireworks (FP8) + Featherless + Haimaker — is the operative signal that the honest 2026 open-weights-distribution counter-position has moved from “a Chinese open-weights MoE drops on Hugging Face and third-party routing providers catch up over the following week or two” to “a Chinese open-weights MoE drops on Hugging Face + ModelScope AND on DeepInfra + OpenRouter + Vercel AI Gateway + Fireworks (FP8) + Featherless + Haimaker in a same-day fan-out, so the open-weights-and-fully-routed-inference stack is a single Thursday on Day 0”. The same-day-across-six-providers tell is the operative distribution-shape signal — coordinated launch presence across the entire first-tier routing stack, rather than a staggered pickup, is the posture a Chinese open-weights lab takes when it has decided “the open-weights launch and the routing launch are one event, and the model has to be callable through every mainstream inference provider on day zero”, and anchors the StepFun distribution surface on “Step 5 Preview is on DeepInfra + OpenRouter + Vercel AI Gateway + Fireworks (FP8) + Featherless + Haimaker on launch day”. (2) The Fireworks-FP8-at-launch tell is the operative deployment-shape signal — Fireworks shipping an FP8 cut of a 600B / 27B MoE on launch day, rather than only BF16 or FP16, is the posture the inference-provider stack takes when it has decided “a 600B / 27B MoE is a workload deployment surface, not just a research artefact, so quantized FP8 and NVFP4 are first-day cuts”. Landing on the same Thu Oct 15 as the NVIDIA Nemotron 3.5 Super VL + Bolt open-weights drops on Hugging Face + OpenRouter + ModelScope + NIM (item 06) and the Multiverse Comet 60B on Hugging Face in BF16 + FP8 + NVFP4 cuts (item 07), the Step 5 Preview same-day fan-out becomes the reference “the Thu Oct 15 open-weights Thursday carries three major open-weights drops that each land on their full first-tier inference-provider stack on day zero, so an open-weights launch and a routed-inference launch are the same event” primitive every subsequent Mistral-Mamba, Qwen4-Flash, DeepSeek-R2, Moonshot-K4 and Meta-Muse-Open open-weights-cadence print now has to price against.

Two reads. (1) A US chip-and-model vendor that ships two open-weights Nemotron 3.5 cuts on the Thu Oct 15 open-weights Thursday — a 120B / 12B active vision-language flagship with 1M context scoring 26.5 on the Artificial Analysis Intelligence Index as the top US open-weight model, and a 74B-A9B everyday-agentic-tasks RTX-laptop model with native 2-bit quantization in a 24GB cut — plus open datasets covering agentic tool use, finance, instruction following, long context, fairness and VL, is the operative signal that the honest 2026 US-open-weights counter-position has moved from “US-side open-weights MoE models are small Llama / Mistral derivatives and the frontier US stack is closed-weights” to “NVIDIA ships Nemotron 3.5 Super VL as the top US open-weight model by AA Index AND Bolt 74B-A9B with native 2-bit quantization for everyday RTX-laptop agentic workloads on the same Thu Oct 15 open-weights Thursday, so the open-weights-US-leader surface moves from a small-model community project to a vendor-shipped VL flagship + a laptop-local agentic-work model + a bundled dataset layer”. The top-US-open-weight-model-by-AA-Index + native-2-bit-RTX-laptop-cut tell is the operative open-weights-stack-shape signal — pairing a VL flagship with a laptop-local agentic-work model under one Nemotron 3.5 banner on a single day, rather than letting the two live on separate release cycles, is the posture a vendor takes when it has decided “the open-weights US stack is a coordinated two-surface drop — a server-side VL flagship AND a laptop-local agentic-work companion — and the DGX Spark / RTX hardware is the on-ramp”, and anchors the Nemotron surface on “Super VL is the top US open-weight VL on AA Index; Bolt 74B-A9B is the native-2-bit RTX-laptop agentic-work companion”. (2) The same-Thu-Oct-15-as-StepFun-Step-5-Preview tell is the operative open-weights-wave-coupling signal — Nemotron 3.5 Super VL + Bolt on Thu Oct 15 landing alongside StepFun Step 5 Preview + the Multiverse Comet 60B CompactifAI compression, is the shape the open-weights ecosystem takes when it has decided “the Thu lab-wave day carries both the Chinese open-weights MoE agent-execution anchor (StepFun) and the US open-weights VL leader (Nemotron 3.5 Super VL) and the compression derivative (Comet 60B) as a three-continent same-day lab-wave”. Landing on the same Thu Oct 15 as the StepFun Step 5 Preview open-weights drop (item 04) and the Multiverse Comet 60B compression (item 07), the Nemotron 3.5 Super VL + Bolt pair becomes the reference “NVIDIA ships the top US open-weight VL model by AA Index AND a native-2-bit RTX-laptop agentic-work companion AND agentic + finance + VL + long-context + fairness datasets on the Thu Oct 15 open-weights Thursday alongside a Chinese 600B MoE and a CompactifAI-compressed 60B VL derivative” primitive every subsequent AMD Instella, Intel OpenVino, Qualcomm, Mediatek, Apple MLX and Meta Llama open-weights-cadence print now has to price against.

Two reads. (1) A model-compression startup in collaboration with NVIDIA that halves the parameter count of Nemotron 3.5 Super VL from 120B to 60B while keeping the same 12B active, the same three modalities, within 5% accuracy on most benchmarks, and outperforming the original on tau2-Bench telecom + BabyVision + AutomationBench-AA, shipped in BF16 + FP8 + NVFP4 on Hugging Face on the same Thu Oct 15 as the original NVIDIA release, is the operative signal that the honest 2026 open-weights-compression counter-position has moved from “a model-compression startup publishes a research paper and shows a demo on a popular open-weights model after the fact” to “a model-compression startup ships a vendor-collaborated 60B compression of a 120B VL flagship ON LAUNCH DAY, in three quantization formats on Hugging Face, within 5% accuracy on most benchmarks and outperforming on some, as a first-class same-day complement to the original open-weights drop”. The CompactifAI-same-day-with-NVIDIA tell is the operative compression-shape signal — a quantum-inspired classical compression framework shipping a vendor-collaborated 60B derivative on launch day, rather than as a follow-on paper weeks later, is the posture a compression startup takes when it has decided “the open-weights launch includes the compression on day zero and the compressed model is a first-class product, not a research artefact”, and anchors the Multiverse surface on “Comet 60B is Nemotron 3.5 Super VL CompactifAI-compressed to 60B, in BF16 + FP8 + NVFP4 on HF, within 5% on most benchmarks, outperforming on tau2-Bench telecom + BabyVision + AutomationBench-AA”. (2) The vendor-collaboration + three-format-launch tell is the operative open-weights-wave-coupling signal — NVIDIA's own blog naming Comet 60B as a launch-day complement, rather than Multiverse releasing it independently a month later, is the shape the open-weights ecosystem takes when it has decided “an open-weights VL flagship launches with its vendor-collaborated compression on the same day, so operators can pick between the 120B flagship and the 60B derivative on day zero”. Landing on the same Thu Oct 15 as the StepFun Step 5 Preview open-weights drop (item 04) and the NVIDIA Nemotron 3.5 Super VL + Bolt pair (item 06), the Multiverse Comet 60B compression becomes the reference “a quantum-inspired compression startup ships a vendor-collaborated 60B derivative of the top US open-weight VL on the same Thu Oct 15 open-weights Thursday as the 120B flagship, in BF16 + FP8 + NVFP4 on Hugging Face, within 5% accuracy on most benchmarks and outperforming on some, so the open-weights launch and the compression launch are a single event” primitive every subsequent AWQ, GPTQ, SpQR, GGUF, Unsloth, DeepCompressor and Microsoft NNI open-weights-compression print now has to price against.

Two reads. (1) A major US retail broker that publicly blocks unauthorized AI agents, including Meta's Muse, from moving customer money on Bloomberg Television, reports tens-and-twenties weekly blocks, questions who's liable when a password-handed-to-an-agent moves money the wrong way, and offers read-only MCP servers + the Charley assistant as the sanctioned alternative, alongside a major US bank CEO saying “no outside agent inside the company”, is the operative signal that the honest 2026 enterprise-agent-liability counter-position has moved from “AI agents log in to financial accounts with the client's password and act on their behalf; liability is implicitly on the client” to “major US financial institutions publicly block unauthorized AI agents from moving customer money, name a specific blocked agent (Meta Muse), publish a weekly block rate (tens and twenties), question US-side AI-agent liability on live television, and point third-party agents at read-only MCP servers + a first-party agent as the sanctioned alternative”. The name-Meta-Muse-as-blocked + offer-read-only-MCP-server tell is the operative counter-posture signal — publicly naming a competitor product as a blocked agent and offering a read-only MCP server as the sanctioned alternative, rather than keeping the blocklist internal, is the posture a major retail broker takes when it has decided “AI-agent-moves-money is a liability regime we don't want and MCP read-only is the compatible surface, and we'll say so on live TV”, and anchors the Schwab surface on “blocked AI agents (incl. Meta Muse) from moving money; read-only MCP servers for research + balances; Charley is the first-party agent”. (2) The Dimon-earnings-call tell is the operative category-coupling signal — JPMorgan's CEO saying on the Thu Oct 15 Q3 earnings call that he isn't going to allow ‘any outside agent to come in inside the company’, a day before Wurster goes on Bloomberg TV, is the shape US financial-services takes when it has decided “no-outside-agent is a public policy, not an internal risk memo, and the broker and the bank say so in the same 24h window”. Landing on the Fri Oct 16 enterprise-agent-liability-Friday slot alongside the MOFCOM + MOST AI export controls proposal (item 09) and the Oct 15 ServiceNow AI Passport launch + Opsin $21M round (prior-covered), the Schwab + JPMorgan AI-agent-moving-money block becomes the reference “US retail broker + US bank publicly block AI agents from moving customer money on Bloomberg TV + the Q3 earnings call in the same 24h window the first big-vendor cryptographic agent-identity primitive launches and a $21M agent-governance round closes, so enterprise-agent-liability is a four-anchor public-posture tape in a single week” primitive every subsequent Fidelity, Vanguard, Bank of America, Wells Fargo and Goldman Sachs AI-agent-liability print now has to price against.

Two reads. (1) A Chinese government proposal that adds 40 technologies to the export-control list, of which 23 are AI items covering large language models, agents, chip designs and computing clusters, with model weights and source code captured above designated performance thresholds, while API, website, app and SaaS access are exempted and open-source + public-domain are carved out, triggering a pre-export registration requirement already in effect even for in-flight shipments, is the operative signal that the honest 2026 US-China-AI-regulation counter-position has moved from “the US tightens chip export controls and China indicates it will respond at some point” to “China publishes a draft export-control list that captures AI model weights and source code above performance thresholds for 23 of 40 named technologies, exempts API + SaaS, carves out open-source and public-domain, triggers an interim pre-export registration already in effect even for in-flight shipments, and runs a public consultation through Nov 15”. The weights-and-source-captured + API-SaaS-exempted + open-source-public-domain-carved-out tell is the operative regulatory-shape signal — treating model weights and source code as controlled technology above thresholds while leaving API access and open-source alone, rather than banning cross-border AI entirely or doing nothing, is the posture China takes when it has decided “the controlled surface is the weights + source + chip designs + clusters at the frontier, not the API or the open-source baseline, so the regime is cost-free for the existing open-weights lab tape and bites on the frontier-closed-weights transfer”, and anchors the MOFCOM surface on “23 of 40 items are AI; weights + source above thresholds captured; API + SaaS exempted; open-source and public-domain carved out; Nov 15 consultation; interim registration already in effect”. (2) The same-day-Fri-Oct-16-as-Schwab-and-JPMorgan tell is the operative category-coupling signal — a Chinese government AI-export-controls proposal landing on the same Fri Oct 16 as a US retail broker and a US bank publicly blocking AI agents from moving customer money, is the shape the enterprise-agent-liability-Friday reading takes when it has decided “the agent-at-the-financial-infrastructure boundary and the model-at-the-national-border boundary are two faces of a single Friday enterprise-agent-liability tape, and both the US private-sector and the Chinese public-sector surfaces print on the same day”. Landing on the same Fri Oct 16 as the Schwab + JPMorgan AI-agent-money-block (item 08) and in the same week as the Thu Oct 15 open-weights Thursday backbone (items 04, 06, 07), the MOFCOM + MOST AI export controls proposal becomes the reference “China publishes a draft export-control list that captures AI model weights + source code above performance thresholds for 23 of 40 technologies on the same Fri Oct 16 that Schwab and JPMorgan publicly block AI agents from moving customer money, so the enterprise-agent-liability + national-security-AI-regulation axes print as a single Friday counter-tape to the Thu open-weights-Thursday backbone” primitive every subsequent US Executive Order, EU AI Act implementing act, UK AI Safety Institute and India AI Advisory AI-controls print now has to price against.