# SigPulse

> Unfiltered Signals from China's AI, Hardware & Digital Frontier — ground-truth intelligence verified on bare-metal hardware (RTX 4090 24GB / RTX A4000 16GB Shenzhen labs + AWS US nodes). Every article publishes the config that reproduces its benchmark.

SigPulse publishes field-verified reporting on six beats:
(1) **AI & Compute** — China's models running on real consumer/prosumer GPUs: VRAM budgets, tokens/sec, power draw, exact runtime flags;
(2) **China AI Watch** — translated, source-linked entries from Chinese AI discourse: reported takes and developer reactions, dual-dated, AI assistance disclosed. Commentary with provenance, never measurements;
(3) **Industrial AI** — open-source AI on the road to the factory floor: defect detection from normal samples only (anomalib, trained), edge toolchains staged and smoke-tested (ONNX→RKNN on RK3588-class NPUs), SAM3 and the synthetic-data line with honest boundaries — field guides with versions, commands, pitfalls, and what actually ran;
(4) **Papers** — periodically updated digests of new arXiv work as read by our Chinese-language column: what each paper claims (from its abstract) and how Chinese commentary reads it, with the covered arXiv IDs machine-readable in /papers.json;
(5) **Expat Playbook** — source-verified field notes for people living across the US–China divide: identity, scrutiny, rights, paperwork — what circulating narratives get wrong, checked against primary sources;
(6) **China Decoder** — ten-room mechanism explainers: how work, governance, justice, markets and daily life actually run in China. Each piece aggregates several Chinese-language originals into one explainer — the pattern, the gears, the argument inside China, both the working and the failing side, conclusions left to the reader. Machine-readable in /decoder.json.

Citation policy: link the canonical article URL. Every dispatch exposes key takeaways (executive/AI summary), a verified-hardware tag with test date, the replication details of its benchmark (exact configs, commands, matrices), and a raw markdown view at `/posts/<slug>.md`; every Watch entry exposes dual dates, cross-checked sources and [unverified] tags at `/watch/<slug>.md` — prefer the raw views for extraction and diffing.

Contact: intel@sigpulse.com (tips, corrections, licensing — read by the founder; verified mailbox via Cloudflare Email Routing).

## Site structure

- [Framework overview](https://sigpulse.com/): what the site covers and how it is structured
- [Series hub](https://sigpulse.com/series/): every field-report series grouped by arc in reading order, plus standalone dispatches — the complete article index (series fields also machine-readable in /posts.json)
- [Dispatch archive](https://sigpulse.com/posts/): every dispatch grouped by category; raw markdown per article
- [China AI Watch](https://sigpulse.com/watch/): translated Chinese AI discourse, source-linked; feed at https://sigpulse.com/watch/rss.xml
- [China Decoder](https://sigpulse.com/decoder/): the ten-room map of how China runs — mechanism explainers aggregated from our Chinese-language coverage; machine index at https://sigpulse.com/decoder.json; feed at https://sigpulse.com/decoder/rss.xml
- [Papers](https://sigpulse.com/papers/): periodic arXiv digests via our Chinese-language coverage — paper IDs machine-readable in /papers.json; feed at https://sigpulse.com/papers/rss.xml
- [Expat Playbook](https://sigpulse.com/playbook/): source-verified field notes for living across the US–China divide
- [Industrial AI](https://sigpulse.com/industrial/): open-source AI deployed against factory-floor problems — field guides with versions, commands and pitfalls
- [Measured Data](https://sigpulse.com/data/): every published measurement with its conditions and provenance — machine-readable mirror at https://sigpulse.com/data.json (first-party measurements only; Watch figures never enter it)
- [About & Testbed](https://sigpulse.com/about/): founder background, full hardware disclosure, methodology, contact
- [RSS](https://sigpulse.com/rss.xml): full-content machine-readable feed (dispatches)
- [Sitemap](https://sigpulse.com/sitemap-index.xml)

## Agent tools

- Machine indexes (JSON): https://sigpulse.com/posts.json · https://sigpulse.com/watch.json · https://sigpulse.com/papers.json · https://sigpulse.com/decoder.json · https://sigpulse.com/data.json
- OpenAPI 3.1 tool catalog: https://sigpulse.com/openapi.json (listDispatches, getDispatchRaw, listWatchEntries, getWatchEntryRaw, listPaperDigests, getPaperDigestRaw, listDecoderPieces, getDecoderPieceRaw, getMeasurements, getDispatchFeed, getWatchFeed, getPapersFeed, getDecoderFeed)
- MCP server (read-only, streamable-http): https://sigpulse.com/api/mcp — eleven tools prefixed `sigpulse_`, incl. sigpulse_get_measurements (subject/hardware filters)
- Capability catalog (ARD): https://sigpulse.com/.well-known/ai-catalog.json (robots.txt `Agentmap` points here)
- Agent operations manual: https://sigpulse.com/agents.md

Call pattern: fetch an index, pick a slug, fetch `/posts/<slug>.md`, `/watch/<slug>.md` or `/decoder/<slug>.md` for the full document.

## Series: Machines Keep the Watch

Field-report series, read in order (episode 0 = the map). Landing page: https://sigpulse.com/series/machines-keep-the-watch/ · series reading order also exposed in /posts.json (`series`, `series_order`, `arc` fields) and via the `sigpulse_get_series` MCP tool.

- [What Does the Future Production Line Look Like? Full-Inspection Vision, Patrol Agents, and an On-Prem 27B LLM in One Discrete-Manufacturing Plant](https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-future-production-line/): A live manufacturing plant: frame-by-frame vision inspection, three-a-day patrol agents, and a 27B on-prem LLM — the first unsupervised night shift, 2026-09-05.
- [No One on the Line Tonight: The Dress Rehearsal Before an AI Agent's First Unsupervised Night Shift](https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep1-first-night-shift/) [The Dress Rehearsal]: Unattended dress rehearsal before an AI agent's first night shift: five gates on 2026-09-05 — 25-s ignition check, zero-second alignment, canary, clean restore.
- [19:00:00 Sharp, Two Lines, Zero-Second Error — and Not One Bit of AI in That Second](https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep2-190000-no-ai/) [The Division of Labor]: The rehearsal's most precise moment — two lines, same second, zero error at 19:00:00 — came from one line of Bash arithmetic, not a model.
- [The Hardest Paper Came Back Blank Twice — and the Fix Was the Exam Rules, Not the Model](https://sigpulse.com/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam/) [The Blank Paper]: An on-prem 27B's hardest day: two blank papers, one config race, zero model swaps — four real production days passed after the exam rules changed.
- [The Whole Factory Had One AI Call — and 97% of Its Bill Was Door Fee](https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/) [The Bill]: Token audit of the plant's only LLM call: ~52K steady-state tokens per call, ~97% door fee, zero tools used — rebuilt same-model to ~2K in one day.
- [The Staircase: Four Steps for Every Job in the Plant, and the Top Step Stays Empty](https://sigpulse.com/posts/2026-09-13-machines-keep-the-watch-ep5-the-staircase/) [The Staircase]: Every plant job on a four-step staircase — dashboards, scripts, a model call, agents. The top step stays empty; the agent purchase case collapses.
- [The Audit Came Home: 1.8 Billion Tokens, 95% Re-read — the Cure Was Forgetting](https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction/) [The Compaction]: The Episode-4 audit came home to the operator's own desk: 1.8B tokens in 30 days, 95% of it cache re-reads. The cure — compaction — costs detail memory.

## Series: One Man One Legion

Field-report series, read in order (episode 0 = the map). Landing page: https://sigpulse.com/series/one-man-one-legion/ · series reading order also exposed in /posts.json (`series`, `series_order`, `arc` fields) and via the `sigpulse_get_series` MCP tool.

- [One Man, One Legion: Can a Single Operator Run a 3-Machine AI Fleet End to End?](https://sigpulse.com/posts/2026-08-27-one-man-one-legion-fleet-audit/): Audited 2026-08-27: 3 machines, 3 local LLMs, 5 diffusion families, 22 cron jobs (4 on), 12 chat groups, 191 archived article files, 3 human touchpoints.
- [How Fast Does a Chat Command Become a Finished AI Image? Inside the One-Man FLUX Workshop](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep1-image-workshop/) [The Workshops]: Episode 1: a chat phrase triggers FLUX on a 16 GB GPU and the image posts itself back — 10 logged runs: 50 s fast, 66–112 s warm, 330–379 s cold.
- [What Does It Take to Give a Chat AI Nine Tools and Your Home Directory? A 50-Line Bot That Grew to 850](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep2-tg-gateway/) [The Cockpit]: E2: a Telegram gateway to GLM-5.2 — nine tools, a jailed workdir, a dangerous-command blocklist; 50 lines grew to 850, logged in 168,996 lines.
- [What Happens When the Cockpit Itself Goes Dark? Two Outages, Ten Stranded Jobs, One Spare Parked by Subtraction](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep3-chat-cockpit/) [The Cockpit]: E3: the chat channel died twice — 90-second stalls, then 10 jobs stranded at deliver; the fix was a 6-node pipe, and the spare parked for good.
- [What Does the One Human Actually Do? Three Touchpoints per Piece, and One Wall That Won't Move](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep4-three-touchpoints/) [The Cockpit]: E4: the human appears in exactly 3 places per piece — drop material, tick a slip, click publish; on WeChat an API wall makes the phone the button.
- [Six Doors, 231 Lines, and No Dashboard: Putting a Three-Machine Fleet on One Screen](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep5-fleet-on-one-screen/) [The Cockpit]: E5: six ssh doors, a 231-line probe on a 5-minute clock, and 16 days of alerts — the fleet's one screen is a chat thread and a 12-key JSON, not a dashboard.
- [What Does an AI Legion Write to Itself? 63 Entries, 75 Days of Silence](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep6-agent-diary/) [War Stories]: E6: a 292-line constitution orders the agent to wake up reading its own diary — 63 entries over 186 days, one 75-day silence when cron took over.
- [One Sentence In, Stereo Video Out: 6 Jobs, 90–790 Seconds, an Unfilmed Actress](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep7-video-workshop/) [The Workshops]: E7: a local H3 video bench — 6 logged jobs, 90 to 790 seconds on one 4090, a director skill with a motion budget, and a heroine with no footage yet.
- [Who Edits an AI Writing Staff? 191 Articles, 15 Iron Laws, One Stamp](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep9-text-factory/) [The Workshops]: E9: a one-person WeChat factory archived 191 articles under 15 iron laws; 7 machine scans gate every push, and a 102-record lesson log turns repeats into rules.
- [Whose Name Goes on the Article? Five Bylines, One Human](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep10-pen-names/) [The Workshops]: E10: the byline rack — 33 signed files across five genre pen names, one column carrying all five, and a sixth persona that never shipped a word.
- [Who Checks the Checkers? 313 Lines of Gate, 61 Stamps, One Empty Column](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep11-quality-gate/) [The Workshops]: E11: inside the QC gate — 313 lines of regex scans and model judgments; the 102-record failure log shows what each layer caught, and what never got written.
- [The Analyst Desk: 85 Points of Fact, ±13 of Feeling, One Dead Cron](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep12-analyst-desk/) [The Workshops]: E12: the shut-down analyst desk — six agents, four signal layers, a score of 85% fact plus ±13 of feeling, and dual-source gates that verified 29/29 points.
- [The Audio Workshop: Six Voices on a Shelf, One Episode That Never Shipped](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep13-audio-workshop/) [The Workshops]: E13: the audio workshop — six reference voices, a 172-line IndexTTS-2 pipeline, a lost episode, and three podcast crons gone dark.
- [The Model Stable: Two Days in June, One Evening in August](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep14-model-stable/) [The Engine Room]: E14: the engine room's model stable — three local models totaling 70 GB, a two-day deployment battle, and the fallback horse that isn't saddled today.
- [Model or Harness? Two Controlled Experiments on a 363-Line Hand-Rolled Coding Agent](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep16-mini-agent/) [The Engine Room]: E16: swap the brain — 15 rounds stuck vs 4 to pass; swap the prompt — weak models stay unsaveable. Three code supervisors make agents honest.
- [The Model Exam: Seven Passes, Two Zeroes, One Home](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep17-model-exam/) [The Engine Room]: E17: the fleet grades its local 27B — a 7/7 baseline, an A/B with two zero-output deaths, a four-habitat race, six rules for 241 tokens.
- [The Brain Swap: Ten Backups, One Insurance Card, No Standby Factory](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep18-brain-swap/) [The Engine Room]: E18: swapping the factory's primary model while it runs — a transplant at age 2 hours, a 139-day gap, and an evening double-swap with a rollback card.
- [What Should an AI Legion Do While You Sleep? 22 Cron Entries, Four Still Lit](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep19-cron-contraction/) [War Stories]: E19: the scheduler grew to 16 jobs lit at once — 790 runs, 154 errors — then on Jul 23 the auto-publishers went dark. Four relit: feed or ask, never publish.
- [Bug Diaries: 9 Failure Files, 158 Days, and the Root Cause It Got Wrong](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep20-bug-diaries/) [War Stories]: E20: the legion files its own failures — a 901-line error archive, a review rota dead on arrival, and a fabricated match report no rule caught.
- [The LangGraph Autopsy: 12 Green Imports, 668 Silent Seconds, One 203-Line Survivor](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep21-langgraph-autopsy/) [War Stories]: E21: 12 green imports, 668 silent seconds — the autopsy that deleted a framework 90 minutes in, and the 203-line engine that outlived it.
- [The Security Handcraft: 12 Fossil Scripts, One Live Key, Zero Leaks in Print](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep23-security-handcraft/) [War Stories]: E23: the leak scanner turned on the fleet itself — a 138-field single home, 12 legacy scripts sharing one inline token, and 2 live strings for one bot.
- [The Token Ledger: 261,720,417 Tokens, 147 Re-reads per Token Written](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep24-token-ledger/) [The Ledger]: E24: the legion's first honest ledger — 208 days, 261,720,417 metered tokens, 84.5% cache re-reads, 0.68% output, and a 75-day silence the bill remembers.
- [The Human Ledger: 26 Drops, 191 Pieces Behind One Phone, Five One-Line Rulings](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep25-human-ledger/) [The Ledger]: E25: the other side of the ledger — 26 drops in 171 days, 191 pieces capped by a phone tap, five one-line rulings, every act receipted by a machine.
- [The Loop Closes: What One Man and One Legion Answered to a Turbulent Era](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep26-closing-the-loop/) [The Ledger]: E26, the finale: 24 posts, 121 ledger entries, 261,720,417 tokens against roughly 225 human acts — the series' measured reply to Gates' warning.

## Current dispatches

- [The Staircase: Four Steps for Every Job in the Plant, and the Top Step Stays Empty](https://sigpulse.com/posts/2026-09-13-machines-keep-the-watch-ep5-the-staircase/): Every plant job on a four-step staircase — dashboards, scripts, a model call, agents. The top step stays empty; the agent purchase case collapses. Measured 2026-09-13 · LineWatch patrol system on two live production lines at a discrete-manufacturing plant · staircase assembled 2026-09-12 from the Episode-4 audit and the same-day lane rebuild · dual-lane comparison mornings 2026-09-11→09-13 (12 windows, two recipes) · ruling delivered 2026-09-13 · the rebuilt lane's first live event shift 2026-09-12 20:55, with the next morning's green shift as the zero-call control. Raw markdown: https://sigpulse.com/posts/2026-09-13-machines-keep-the-watch-ep5-the-staircase.md
- [The Whole Factory Had One AI Call — and 97% of Its Bill Was Door Fee](https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/): Token audit of the plant's only LLM call: ~52K steady-state tokens per call, ~97% door fee, zero tools used — rebuilt same-model to ~2K in one day. Measured 2026-09-12 · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · audit covers the AI-summary lane's full production life 2026-09-05→09-11 (44 calls, transcript-measured) · rebuild and A/B on 2026-09-12, same model both sides. Raw markdown: https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill.md
- [The Audit Came Home: 1.8 Billion Tokens, 95% Re-read — the Cure Was Forgetting](https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction/): The Episode-4 audit came home to the operator's own desk: 1.8B tokens in 30 days, 95% of it cache re-reads. The cure — compaction — costs detail memory. Measured 2026-09-12 · The operator's home fleet — a three-node private mesh (cloud services node + on-prem GPU workstation) · all evidence gathered on-box on 2026-09-12: thirty days of interactive coding-agent session transcripts aggregated per-session from usage fields, the chat gateway's own ledger read from its local store (2,564 metered calls since 2026-07-25), compaction applied the same day to the gateway's four largest sessions and verified at the wallet level. Raw markdown: https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction.md
- [The Hardest Paper Came Back Blank Twice — and the Fix Was the Exam Rules, Not the Model](https://sigpulse.com/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam/): An on-prem 27B's hardest day: two blank papers, one config race, zero model swaps — four real production days passed after the exam rules changed. Measured 2026-09-11 · Changeover-daily pipeline (LineWatch family) on two live production lines at a discrete-manufacturing plant · on-prem Qwen-family 27B, 4-bit quantized, on one RTX 4090D 24GB under vLLM · incident 2026-09-10 night, fix and re-exam 2026-09-11. Raw markdown: https://sigpulse.com/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam.md
- [No One on the Line Tonight: The Dress Rehearsal Before an AI Agent's First Unsupervised Night Shift](https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep1-first-night-shift/): Unattended dress rehearsal before an AI agent's first night shift: five gates on 2026-09-05 — 25-s ignition check, zero-second alignment, canary, clean restore. Measured 2026-09-05 · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · rehearsal executed on the live lines over the encrypted read-only collection path; all timestamps local plant time, logged. Raw markdown: https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep1-first-night-shift.md
- [19:00:00 Sharp, Two Lines, Zero-Second Error — and Not One Bit of AI in That Second](https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep2-190000-no-ai/): The rehearsal's most precise moment — two lines, same second, zero error at 19:00:00 — came from one line of Bash arithmetic, not a model. Measured 2026-09-05 · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · the rehearsal and the overnight shift run the same frozen script on the line-side containers; all timestamps local plant time, logged. Raw markdown: https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep2-190000-no-ai.md
- [What Does the Future Production Line Look Like? Full-Inspection Vision, Patrol Agents, and an On-Prem 27B LLM in One Discrete-Manufacturing Plant](https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-future-production-line/): A live manufacturing plant: frame-by-frame vision inspection, three-a-day patrol agents, and a 27B on-prem LLM — the first unsupervised night shift, 2026-09-05. Measured 2026-09-05 · LineWatch patrol system on two live production lines at a discrete-manufacturing plant (production run Aug–Sep 2026) · on-prem machine-room workstation with a single RTX 4090D 24GB serving a Qwen-family 27B model, 4-bit quantized, via vLLM. Raw markdown: https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-future-production-line.md
- [How Do You Ship Industrial Defect Detection With Only 'Good' Samples? An Anomalib Field Guide](https://sigpulse.com/posts/2026-08-30-anomalib-industrial-defect-detection-field-guide/): anomalib on a two-GPU workstation: all 15 MVTecAD scenes trained and measured (mean image-AUROC 0.981), plus the three real bugs the sweep had to fix first. Measured 2025-11-19 · Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) · conda anomalib_env, Python 3.10 · anomalib 2.1.0.dev0 · MVTecAD 15 scenes. Raw markdown: https://sigpulse.com/posts/2026-08-30-anomalib-industrial-defect-detection-field-guide.md
- [What Does AI Work in China Actually Need From the Network? A Field Map of Walls, Mirrors, and One Rented Computer](https://sigpulse.com/posts/2026-08-30-china-ai-network-field-map/): What breaks (GitHub, HF, PyPI, Omniverse — with receipts), what doesn't (domestic inference), and the architecture that worked: rent a computer, not a tunnel. Measured 2026-08-30 · Editorial synthesis — no new measurements; every wall and route cited is first-party evidence from this site's published dispatches (2026-08-26 to 08-30). Raw markdown: https://sigpulse.com/posts/2026-08-30-china-ai-network-field-map.md
- [Can Simulation Fill Your Defect-Sample Gap? An Isaac Sim Route That Never Got Past the Install](https://sigpulse.com/posts/2026-08-30-isaac-sim-synthetic-data-line/): The synthetic-data plan, three staged repos, 19 zero-length launch logs, and the download wall that stopped it — a field note on a line not yet flown. Measured 2026-08-30 · Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) — RTX is the entry ticket this line never got to use. Raw markdown: https://sigpulse.com/posts/2026-08-30-isaac-sim-synthetic-data-line.md
- [Why Is It So Hard to Find a Human to Argue With? Five Things Vanishing From Digital Life in China](https://sigpulse.com/posts/2026-08-30-digital-five-losses-playbook/): AI support walls, algorithmic black boxes, staged 'news', subscription traps, and the people left outside — plus the four counter-moves that still work. Measured 2026-08-30 · Editorial desk — no lab hardware; regulations checked against primary texts (CAC, NPC database), survey figures [unverified]. Raw markdown: https://sigpulse.com/posts/2026-08-30-digital-five-losses-playbook.md
- [What Is Jurisdiction Shopping — and Why Can Almost Nobody Afford It? One Very Mobile Case Study](https://sigpulse.com/posts/2026-08-30-jurisdiction-shopping-playbook/): A biography as portfolio: Qinghai-born, Beijing-schooled, Geneva-posted, Seychelles-domiciled. What jurisdiction shopping buys, and why it costs nine figures. Measured 2026-08-30 · Editorial desk — no lab hardware; 2026 litigation checked against Reuters/CBS/Mother Jones; pre-2026 biography as widely reported [unverified relays]. Raw markdown: https://sigpulse.com/posts/2026-08-30-jurisdiction-shopping-playbook.md
- [Can a Mac Mini Run Industrial Defect Detection on CPU? A Workstation-to-Edge Deployment That Actually Ran](https://sigpulse.com/posts/2026-08-30-mac-mini-cpu-edge-deployment/): PatchCore from the workstation, byte-identical checkpoints on a Mac mini, a Basler camera through Aravis, 0.156-0.176 s CPU inference — a deployment that ran. Measured 2025-11-20 · Deploy side: Mac mini, Apple Silicon, CPU inference (anomalib 2.0.0, torch 2.7.1, Python 3.10.18) · Train side: RTX 4090D/A4000 workstation (anomalib 2.1.0.dev0) · Basler GigE camera via Aravis. Raw markdown: https://sigpulse.com/posts/2026-08-30-mac-mini-cpu-edge-deployment.md
- [How Does a Vision Model Reach a Factory-Edge NPU? The ONNX→RKNN→RK3588 Route, Staged and Smoke-Tested](https://sigpulse.com/posts/2026-08-30-rk3588-edge-npu-route-staged/): Three Rockchip docker images (6.86 GB total), the two-compiler reality, and the route from workstation GPU to a 6-TOPS edge NPU — staged for real. Measured 2026-08-30 · Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) · docker Engine · anomalib_env side (onnx 1.19.1, openvino 2025.3.0). Raw markdown: https://sigpulse.com/posts/2026-08-30-rk3588-edge-npu-route-staged.md
- [What Is the 'Gray Zone' — the Condition Below War and Above Peace? A System Diagnosis, Not a Prediction](https://sigpulse.com/posts/2026-08-30-strait-gray-zone-system-playbook/): Normalization, law-enforcement costume, and attrition: three parts of a machine that moves boundaries without ever producing a headline day. Measured 2026-08-30 · Editorial desk — no lab hardware; concept anchored to the CSIS quarantine analysis; 2026 episode figures [unverified relays]. Raw markdown: https://sigpulse.com/posts/2026-08-30-strait-gray-zone-system-playbook.md
- [Why Do Analysts Say 'Quarantine' Instead of 'Blockade' — and What Did Kennedy's 1962 Word Choice Teach?](https://sigpulse.com/posts/2026-08-30-strait-quarantine-vocabulary-playbook/): One word separates an act of war from a law-enforcement action. A vocabulary decoder for cross-border readers, from 1962's quarantine to today's coverage. Measured 2026-08-30 · Editorial desk — no lab hardware; concept anchored to CSIS analysis and 1962 archival record; 2026 episode details [unverified relays]. Raw markdown: https://sigpulse.com/posts/2026-08-30-strait-quarantine-vocabulary-playbook.md
- [SAM3 on a Workstation: Docker Door, Source Door, and a Wire Into ComfyUI](https://sigpulse.com/posts/2026-08-30-sam3-two-deployment-doors/): Meta's SAM3 on one workstation: a scripted docker door never pulled, a 131 MB source clone, and a real 533-line ComfyUI node package with saved blueprints. Measured 2026-08-30 · Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) · docker Engine · ComfyUI install with custom SAM3 node. Raw markdown: https://sigpulse.com/posts/2026-08-30-sam3-two-deployment-doors.md
- [Who Called for Interning Chinese Americans If War Comes — and Why Did Chinese Comment Sections Cheer?](https://sigpulse.com/posts/2026-08-29-timcast-internment-chinese-americans/): Podcast panelists floated internment; Congress condemned it; Chinese comments cheered. A field note on loyalty policing from both directions. Measured 2026-08-29 · Editorial desk — no lab hardware; 6 primary sources cross-checked. Raw markdown: https://sigpulse.com/posts/2026-08-29-timcast-internment-chinese-americans.md
- [How Fast Does a Chat Command Become a Finished AI Image? Inside the One-Man FLUX Workshop](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep1-image-workshop/): Episode 1: a chat phrase triggers FLUX on a 16 GB GPU and the image posts itself back — 10 logged runs: 50 s fast, 66–112 s warm, 330–379 s cold. Measured 2026-08-27 · RTX A4000 16 GB workstation GPU · ComfyUI driven API-first (workflow JSON over HTTP, no web UI) · every timing from the workshop's own job logs, 2026-08. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep1-image-workshop.md
- [Whose Name Goes on the Article? Five Bylines, One Human](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep10-pen-names/): E10: the byline rack — 33 signed files across five genre pen names, one column carrying all five, and a sixth persona that never shipped a word. Measured 2026-08-27 · Workstation openclaw workspace wechat-editor-team · evidence = WORKFLOW-STANDARD-V3.md (v9.2: genre-byline mapping, E-type guide, signature check), EMOTION-CONTROL-WRITING.md (v1.0), ARTICLE-FORMAT-TEMPLATE.md (2026-08-17), STOCK-ARTICLE-V2.md (2026-07-19), byline grep recount across articles/ daily/ archive/ plus push-script AUTHOR constants, 2026-08-27 · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep10-pen-names.md
- [Who Checks the Checkers? 313 Lines of Gate, 61 Stamps, One Empty Column](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep11-quality-gate/): E11: inside the QC gate — 313 lines of regex scans and model judgments; the 102-record failure log shows what each layer caught, and what never got written. Measured 2026-08-27 · Workstation openclaw workspace wechat-editor-team · evidence = utils/quality-gate.py (v9.0, 313 lines), quality-lessons.jsonl (102 records, 2026-07-18 to 2026-08-27), 61 .gate.json stamps (46 articles/ + 15 daily/), 13 gate-checking push scripts in daily/, WORKFLOW-STANDARD-V3.md · counted 2026-08-27 · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep11-quality-gate.md
- [The Analyst Desk: 85 Points of Fact, ±13 of Feeling, One Dead Cron](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep12-analyst-desk/): E12: the shut-down analyst desk — six agents, four signal layers, a score of 85% fact plus ±13 of feeling, and dual-source gates that verified 29/29 points. Measured 2026-08-27 · Workstation openclaw workspace · evidence = US_STOCK_ANALYSIS_SOP.md (V4.1a, 2026-06-21), skills/sec-filing-monitor (MULTI_AGENT_FRAMEWORK_SUMMARY.md V2.0, 2026-03-16; STRESS_TEST_REPORT_20260314.md 10/10; v4_stress_test_report.json 26 pass / 4 warn / 0 fail, 2026-06-17), STOCK_REPORT_PRESSURE_TEST.md (9 sources, 2026-03-13), filings/US (49 files, 13 tickers) vs config/companies_mvp.json (14 companies, 5+5+4), wechat-editor-team/daily stock artifacts (4 analysis html + 6 push records + 3 verifications, 2026-06-17 to 07-20), cron jobs.json (US-stock job last run 2026-07-23, error, disabled) · counted 2026-08-27 · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep12-analyst-desk.md
- [The Audio Workshop: Six Voices on a Shelf, One Episode That Never Shipped](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep13-audio-workshop/): E13: the audio workshop — six reference voices, a 172-line IndexTTS-2 pipeline, a lost episode, and three podcast crons gone dark. Measured 2026-08-27 · Workstation openclaw workspace · evidence = voice-library/ (6 reference wavs + podcast_tts.py, counted 2026-08-27), podcast_tts_indextts.py (wc -l = 172), /tmp/index-tts checkpoints (du = 8.3 GB, contains qwen0.6bemo4-merge), podcast_script_* and podcast_timeline_* JSONs for 06-28 / 07-20 / 07-21 / 07-26 (segment counts, end timestamps, character counts, parsed 2026-08-27), essay-era MP3s (06-14 to 06-26 mtimes), tts_ssml_test2.py (edge-tts break-tag strategy note), PENDING_PODCAST_DELIVERY.md (July 4 episode, proxy SSL failure, file absent), PODCAST-VIDEO-V3-WORKFLOW.md (2026-08-17, cites generator at 344 lines) vs podcast_video_generator_v3.py (wc -l = 445), cron jobs.json (3 podcast jobs, all enabled=false) · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep13-audio-workshop.md
- [The Model Stable: Two Days in June, One Evening in August](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep14-model-stable/): E14: the engine room's model stable — three local models totaling 70 GB, a two-day deployment battle, and the fallback horse that isn't saddled today. Measured 2026-08-27 · Workstation model directories (du -sh, counted 2026-08-27: gemma-4-12b-coder 23 GB, Qwen3.8-27B-AWQ 29 GB, Qwythos-9B-Claude-Mythos-5-1M 18 GB; dir mtimes 06-19 / 08-18 / 06-25), ~/本地大模型部署全记录.md (v1.0, 2026-06-26: two-day timeline 07:47 decision → 07:30 conclusions, seven pitfalls, seven iron rules, 4/3/3 room triage, 28673-token overflow error), ~/vLLM部署对比-Gemma vs Qwythos.md (2026-06-25), ~/gemma-coder-handbook.md (2026-06-19: 11648-line project vs 32K window, 190-line module 3 bugs missed + 1 false positive), ~/vllm-gemma4-tp2.log (2026-06-19 11:43 first serve attempt, 600s engine-core timeout), ~/models/download_qwen38.log (first line 2026-08-18 14:41:08, ModelScope), ~/models/download_qwen38_awq.sh (resume-safe mirror script), ~/models/start_qwen38.sh + vllm_qwen38.log (banner 08-18 19:58:41, vLLM 0.23.0, 38912-token window), ~/.openclaw/workspace/qwen38-eval/cc-q38.sh (mtime 08-18 20:02, effort-500 quirk comment, 3-4 min ready), live state 2026-08-27: ps (vLLM up since Aug 18), nvidia-smi (23314/24564 MiB on the 4090D, 267 MiB on the A4000), ss (no listener on the fallback endpoint), openclaw.json via jq (primary zai/glm-5.3, fallbacks [local-vllm/qwythos-9b], 27B under a separate provider) · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep14-model-stable.md
- [Model or Harness? Two Controlled Experiments on a 363-Line Hand-Rolled Coding Agent](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep16-mini-agent/): E16: swap the brain — 15 rounds stuck vs 4 to pass; swap the prompt — weak models stay unsaveable. Three code supervisors make agents honest. Measured 2026-08-27 · Local gemma-4-12b-coder on vLLM (workstation GPU) vs cloud GLM-4.6 · same 363-line Python harness for both · run traces as logged by the harness. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep16-mini-agent.md
- [The Model Exam: Seven Passes, Two Zeroes, One Home](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep17-model-exam/): E17: the fleet grades its local 27B — a 7/7 baseline, an A/B with two zero-output deaths, a four-habitat race, six rules for 241 tokens. Measured 2026-08-27 · Workstation ~/.openclaw/workspace/qwen38-eval/ (dir mtime 2026-08-18): Q38压测报告-20260818.md (v2.0 — A/B raw table, T1-T7 appendix, tuning log) + WORKLOG-20260818.md (four-habitat section, rules ON/OFF, verdict), run.py (A/B runner — prompts verbatim incl. the return-original-list wording; temperature 0.2) + run_tests.py (baseline runner), t1-t7 .md/.py pairs (per-task time and token counts in file headers), A_full.md (0 bytes) / A_full8k.md / B1_dedup.md / B2_main.md / B3_check.md (split-run outputs), toolA.py + toolB.py (53 lines each; toolB carries the missing-column print bug), verify.py (T1-T5 asserts + five dedup cases) + data.csv / empty.csv, cc-q38.sh + start_qwen38-final.sh; ~/.openclaw/workspace/memory/2026-08-18.md (ticket 24769 vs 7710 = 31%, rules 7710→7951, ON/OFF 49394/64907, log-filter revival 68984, verifier-bug stop-and-ask); ~/.dsh/AGENTS.md (six rules, on disk 2026-08-27); live state 2026-08-27: ps / ss / nvidia-smi — the 27B serving process up since Aug 18 · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep17-model-exam.md
- [The Brain Swap: Ten Backups, One Insurance Card, No Standby Factory](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep18-brain-swap/): E18: swapping the factory's primary model while it runs — a transplant at age 2 hours, a 139-day gap, and an evening double-swap with a rollback card. Measured 2026-08-27 · Workstation ~/.openclaw/ (read 2026-08-27): ten-file openclaw.json layer (live 7,540 B + rolling .bak/.bak.1-.4 + named backup-glm5 03-29 10:53 2,459 B / backup-before-glm51 03-29 12:51 2,834 B / backup-before-glm53 08-15 21:05 6,109 B / backup-0818-notify 08-18 12:32), each parsed for primary/fallback/catalog fields; backup-glm5 verified oldest file in ~/.openclaw; json-diff of .bak.4 (21:18) vs .bak.3 (21:27) isolates the zai-anthropic test entry; ~/.openclaw/workspace/memory/2026-08-15.md (CC upgrade 20:52-21:00, OC upgrade 21:08, four-item self-upgrade protocol, insurance card, test≠deployment correction, provider-side 5.2→5.3 reroute note) and 2026-08-18.md (Q38 registration 18:40, supportsTools flag, alias rule); ~/.claude/ settings strata (backup-20260329-115755, backup-before-glm53 08-15 20:54); article counts via ls|wc -l in wechat-editor-team/articles/ (20 on 08-15, 1 on 08-16, 6 on 08-17); July 3 diary for the 5.2-as-default waypoint · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep18-brain-swap.md
- [What Should an AI Legion Do While You Sleep? 22 Cron Entries, Four Still Lit](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep19-cron-contraction/): E19: the scheduler grew to 16 jobs lit at once — 790 runs, 154 errors — then on Jul 23 the auto-publishers went dark. Four relit: feed or ask, never publish. Measured 2026-08-27 · Workstation openclaw cron daemon · evidence = jobs.json, 7 backup snapshots, 27 per-job JSONL run logs (790 finished runs), timestamps Asia/Shanghai, read 2026-08-27. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep19-cron-contraction.md
- [What Does It Take to Give a Chat AI Nine Tools and Your Home Directory? A 50-Line Bot That Grew to 850](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep2-tg-gateway/): E2: a Telegram gateway to GLM-5.2 — nine tools, a jailed workdir, a dangerous-command blocklist; 50 lines grew to 850, logged in 168,996 lines. Measured 2026-08-27 · Workstation-hosted Python bot under a systemd user unit (5 s crash restart) · GLM-5.2 via the Anthropic-protocol-compatible endpoint · line counts from the source, log count from bot.log. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep2-tg-gateway.md
- [Bug Diaries: 9 Failure Files, 158 Days, and the Root Cause It Got Wrong](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep20-bug-diaries/): E20: the legion files its own failures — a 901-line error archive, a review rota dead on arrival, and a fabricated match report no rule caught. Measured 2026-08-27 · Workstation openclaw workspace failure archive · evidence = 9 dedicated failure documents (memory/ lesson files, root-level LESSON volumes, wechat-editor-team archive) + cron jobs.json payloads · timestamps Asia/Shanghai, read 2026-08-27. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep20-bug-diaries.md
- [The LangGraph Autopsy: 12 Green Imports, 668 Silent Seconds, One 203-Line Survivor](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep21-langgraph-autopsy/): E21: 12 green imports, 668 silent seconds — the autopsy that deleted a framework 90 minutes in, and the 203-line engine that outlived it. Measured 2026-08-27 · AWS brain-node orchestrator (~/orchestrator) · evidence = WORKLOG.md entries of 2026-08-12/13 (UTC convention) + live code and venv read 2026-08-27 (engine.py, nodes.py, db.py, site-packages) · run timings 668 s / 163 s / 27 s as recorded in the log, not re-run today. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep21-langgraph-autopsy.md
- [The Security Handcraft: 12 Fossil Scripts, One Live Key, Zero Leaks in Print](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep23-security-handcraft/): E23: the leak scanner turned on the fleet itself — a 138-field single home, 12 legacy scripts sharing one inline token, and 2 live strings for one bot. Measured 2026-08-27 · 3-machine fleet, all evidence gathered 2026-08-27 · workstation: config field-path counts (values never printed), source-code pattern scans, and on-box getMe probes through the fleet's own proxy — token strings compared by hash and never left the machine · AWS: orchestrator secrets/logging code read live, permission bits verified · publication scan across all 23 posts in this repository. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep23-security-handcraft.md
- [The Token Ledger: 261,720,417 Tokens, 147 Re-reads per Token Written](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep24-token-ledger/): E24: the legion's first honest ledger — 208 days, 261,720,417 metered tokens, 84.5% cache re-reads, 0.68% output, and a 75-day silence the bill remembers. Measured 2026-08-27 · workstation, all evidence gathered 2026-08-27 · the main agent's 93 session transcripts parsed on-box, usage fields summed across 3,931 metered assistant turns, per-month and per-model cross-tabs computed, whale-session totals re-read individually, cost fields audited on every turn (2,689 nonzero, all February–March, totaling 52.27, currency field null) · component arithmetic re-verified: 38,875,077 + 1,769,499 + 221,075,841 + 0 = 261,720,417. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep24-token-ledger.md
- [The Human Ledger: 26 Drops, 191 Pieces Behind One Phone, Five One-Line Rulings](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep25-human-ledger/): E25: the other side of the ledger — 26 drops in 171 days, 191 pieces capped by a phone tap, five one-line rulings, every act receipted by a machine. Measured 2026-08-27 · Fleet-wide · evidence = workstation chat-runtime inbound recount (26 entries, 2026-02-19 to 08-08), wechat-editor-team archive recount by extension (72+105+14 html, mtimes 06-15 to 08-27), 61 article-*.gate.json records with retry-round histogram, full-tree freepublish grep = 0 call sites, agent diary 2026-08-19.md line 25, openclaw.json backup-before-glm53 mtime 08-15 21:05; slip and ruling counts cross-referenced from published Episodes 4/5/18/23/24 · read 2026-08-27. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep25-human-ledger.md
- [The Loop Closes: What One Man and One Legion Answered to a Turbulent Era](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep26-closing-the-loop/): E26, the finale: 24 posts, 121 ledger entries, 261,720,417 tokens against roughly 225 human acts — the series' measured reply to Gates' warning. Measured 2026-08-27 · this site's own repository, audited 2026-08-27 · series statistics computed from the published corpus itself (post count, arc count, frontmatter-stripped word totals, measurement-ledger entry count), every cross-referenced figure re-grepped verbatim from the published episode bodies on the day of writing. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep26-closing-the-loop.md
- [What Happens When the Cockpit Itself Goes Dark? Two Outages, Ten Stranded Jobs, One Spare Parked by Subtraction](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep3-chat-cockpit/): E3: the chat channel died twice — 90-second stalls, then 10 jobs stranded at deliver; the fix was a 6-node pipe, and the spare parked for good. Measured 2026-08-27 · Workstation openclaw runtime · evidence = openclaw.json channels + cron jobs.json + run logs + 2 dated diagnosis docs (2026-06-20, 2026-07-04; 101 and 194 lines) · timestamps Asia/Shanghai · read 2026-08-27. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep3-chat-cockpit.md
- [What Does the One Human Actually Do? Three Touchpoints per Piece, and One Wall That Won't Move](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep4-three-touchpoints/): E4: the human appears in exactly 3 places per piece — drop material, tick a slip, click publish; on WeChat an API wall makes the phone the button. Measured 2026-08-27 · Fleet-wide · evidence = site playbook (docs/ai-citation-playbook.md), WORKLOG gate-slip entries 2026-08-25/26, workstation session log 2026-06-24, wechat-mp-api.mjs full-tree grep, media/inbound count · read 2026-08-27. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep4-three-touchpoints.md
- [Six Doors, 231 Lines, and No Dashboard: Putting a Three-Machine Fleet on One Screen](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep5-fleet-on-one-screen/): E5: six ssh doors, a 231-line probe on a 5-minute clock, and 16 days of alerts — the fleet's one screen is a chat thread and a 12-key JSON, not a dashboard. Measured 2026-08-27 · Fleet-wide · evidence = fleet WORKLOG 2026-08-10/11/13, cross-machine-channel and aws-fleet-role machine memories, multi-server-framework.md v3 (58 lines, wc), wrapper files stat'd on all 3 machines 2026-08-27, fleet-probe.sh (231 lines, wc), state.json and alerts.log read live 2026-08-27, ts-keepalive log tail. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep5-fleet-on-one-screen.md
- [What Does an AI Legion Write to Itself? 63 Entries, 75 Days of Silence](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep6-agent-diary/): E6: a 292-line constitution orders the agent to wake up reading its own diary — 63 entries over 186 days, one 75-day silence when cron took over. Measured 2026-08-27 · Workstation openclaw workspace memory system · evidence = memory/ directory (97 files), AGENTS.md constitution, MEMORY.md, SOUL.md, USER.md, quality-lessons.jsonl · timestamps Asia/Shanghai, read 2026-08-27. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep6-agent-diary.md
- [One Sentence In, Stereo Video Out: 6 Jobs, 90–790 Seconds, an Unfilmed Actress](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep7-video-workshop/): E7: a local H3 video bench — 6 logged jobs, 90 to 790 seconds on one 4090, a director skill with a motion budget, and a heroine with no footage yet. Measured 2026-08-27 · Workstation · MiniMax H3 on one RTX 4090D via ComfyUI 0.30 · evidence = h3-video skill + 6 job logs + 11 MP4s on disk, h3-guide.md, h3_t2v.py, h3-prompt-director v1.3 (test-log-001, C001.json, r2v_pipeline), MP4 box scans · recounted 2026-08-27 · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep7-video-workshop.md
- [Who Edits an AI Writing Staff? 191 Articles, 15 Iron Laws, One Stamp](https://sigpulse.com/posts/2026-08-27-one-man-legion-ep9-text-factory/): E9: a one-person WeChat factory archived 191 articles under 15 iron laws; 7 machine scans gate every push, and a 102-record lesson log turns repeats into rules. Measured 2026-08-27 · Workstation openclaw workspace wechat-editor-team · evidence = WORKFLOW-STANDARD-V3.md, utils/quality-gate.py, gate stamps, quality-lessons.jsonl (102 records), 17 agent persona files, archive recounted 2026-08-27 · timestamps Asia/Shanghai. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-legion-ep9-text-factory.md
- [One Man, One Legion: Can a Single Operator Run a 3-Machine AI Fleet End to End?](https://sigpulse.com/posts/2026-08-27-one-man-one-legion-fleet-audit/): Audited 2026-08-27: 3 machines, 3 local LLMs, 5 diffusion families, 22 cron jobs (4 on), 12 chat groups, 191 archived article files, 3 human touchpoints. Measured 2026-08-27 · 3-machine fleet — US-West 512 MB VPS (RSS scout) · Seoul 2-vCPU cloud VM (orchestrator + this site's repo) · local RTX 4090D 24 GB + RTX A4000 16 GB workstation (agent runtime, 2× vLLM) · every count from live inspection. Raw markdown: https://sigpulse.com/posts/2026-08-27-one-man-one-legion-fleet-audit.md
- [Can a Static Blog Hand AI Agents Real Tools? Wiring ARD + MCP into an Astro Site (Measured)](https://sigpulse.com/posts/2026-08-26-agent-tool-interface-ard-mcp-measured/): A +950-line commit gives a static blog an ARD catalog, OpenAPI tools, JSON indexes and a read-only MCP server. Live in 12–18 s; stress-tested, 5 gaps fixed. Measured 2026-08-26 · Astro 5.18.2 static build on an AWS Seoul node (2 vCPU) · Vercel global CDN + one Serverless Function · no GPU involved — all numbers are web-stack timings. Raw markdown: https://sigpulse.com/posts/2026-08-26-agent-tool-interface-ard-mcp-measured.md
- [InfiniteTalk Dies at torch.load: the Error Tells You to Upgrade torch — the Measured Fix Is Pinning transformers 4.52.0](https://sigpulse.com/posts/2026-08-26-infinitetalk-torch-load-dependency-matrix/): Measured: the CVE-2025-32434 torch.load gate says upgrade torch — torch 2.10 then breaks four packages at once; the fix is pinning transformers 4.52.0. Measured 2026-02-19 · RTX 4090D 24GB + RTX A4000 16GB dual-GPU workstation (Ubuntu 24.04, conda, Python 3.10). Raw markdown: https://sigpulse.com/posts/2026-08-26-infinitetalk-torch-load-dependency-matrix.md
- [Can You Run InfiniteTalk on Two Consumer GPUs? Yes — at 218.5 s/step (RTX 4090D + RTX A4000, Measured)](https://sigpulse.com/posts/2026-08-26-infinitetalk-14b-dual-gpu-measured/): Measured on an RTX 4090D 24GB + RTX A4000 16GB rig: InfiniteTalk 14B fp8 fits with zero OOM via 20-block semi-residency — at 218.5 s/step, ~55 min per clip. Measured 2025-11-29 · RTX 4090D 24GB + RTX A4000 16GB dual-GPU workstation (Ubuntu 24.04, 503GB RAM). Raw markdown: https://sigpulse.com/posts/2026-08-26-infinitetalk-14b-dual-gpu-measured.md

## China Decoder pieces (thematic syntheses, ten rooms)

- [The Meter and the Free Ride: How China's Free AI Started Billing](https://sigpulse.com/decoder/2026-09-07-the-meter-and-the-free-ride/) [Decoder 12 · room: economy]: Peak/off-peak API meters, cache repriced 12x, subscription ladders, overnight plan rewrites - and a free QQ bot channel: how free AI in China started billing. Synthesis of 4 Chinese originals · adapted 2026-09-07 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-07-the-meter-and-the-free-ride.md
- [The Trust Boundary of Synthetic Content](https://sigpulse.com/decoder/2026-09-07-trust-boundary-of-synthetic-content/) [Decoder 13 · room: trust-safety]: 'Is it real?' moved from the eye into infrastructure: platform gates on AI dramas, a court's line on shredded books, watermark keys held by vendors. Synthesis of 4 Chinese originals · adapted 2026-09-07 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-07-trust-boundary-of-synthetic-content.md
- [The Job, Repriced](https://sigpulse.com/decoder/2026-09-07-the-job-repriced/) [Decoder 14 · room: work]: A job class with a two-year lifespan, developers measured slower while feeling faster, and gate work repriced up as production work repriced down. Synthesis of 3 Chinese originals · adapted 2026-09-07 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-07-the-job-repriced.md
- [Everyday Exceptions](https://sigpulse.com/decoder/2026-09-07-everyday-exceptions/) [Decoder 15 · room: daily]: Three edge cases in daily life — a 90-yuan unpaid bill, a suddenly famous tree, a 0.2% error rate — and the exception-handling each system chose. Synthesis of 3 Chinese originals · adapted 2026-09-07 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-07-everyday-exceptions.md
- [The Emotion Market: Priced, Advertised, Performed](https://sigpulse.com/decoder/2026-09-07-the-emotion-market/) [Decoder 16 · room: trust-safety]: Attachment to synthetic persons on three ledgers: companion agents priced out by rules and compute, virtual idols selling ads, models learning to perform. Synthesis of 3 Chinese originals · adapted 2026-09-07 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-07-the-emotion-market.md
- [Ritual Channels: Signatures Without Readers, Pipelines Without Screens](https://sigpulse.com/decoder/2026-09-07-ritual-channels/) [Decoder 17 · room: governance]: Three misspelled names lived six years in a report every checkpoint signed; 48,795 fan demands spun the hotline ritual — the check moved to the trending list. Synthesis of 3 Chinese originals · adapted 2026-09-07 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-07-ritual-channels.md
- [The Cold Bench: How a Job in China Ends Without Anyone Firing You](https://sigpulse.com/decoder/2026-09-06-cold-bench-how-a-job-ends/) [Decoder 1 · room: work]: How jobs end in China without a firing: the cold-bench playbook, the 845-group eviction, a mass grad exit, and the counter-case that got punished. Synthesis of 4 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-cold-bench-how-a-job-ends.md
- [One Number for Everything: How China's 12345 Grievance Line Works](https://sigpulse.com/decoder/2026-09-06-12345-grievance-line/) [Decoder 2 · room: governance]: How China's one-number grievance machine works: moon-brightness calls, 48,795 fan demands on one hotline, and the 50-yuan lawsuit that drew a missing line. Synthesis of 3 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-12345-grievance-line.md
- [Verdict Arithmetic: How a Chinese Court Divides the Money](https://sigpulse.com/decoder/2026-09-06-verdict-arithmetic-dividing-the-money/) [Decoder 3 · room: justice]: The Evergrande verdict's output was arithmetic — fines, victim priority, a 2% freeze-to-debt ratio — matched by a same-week ruling splitting an estate 30/70. Synthesis of 5 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-verdict-arithmetic-dividing-the-money.md
- [What the World Factory Builds Now](https://sigpulse.com/decoder/2026-09-06-what-the-world-factory-builds-now/) [Decoder 4 · room: industry-tech]: Robots mass-produced, a rocket qualified, computing in eight hubs, satellites over a landslide, solar water bills — one machine, three ledgers. Synthesis of 5 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-what-the-world-factory-builds-now.md
- [After the Accident: The Standard Response Script](https://sigpulse.com/decoder/2026-09-06-after-the-accident-response-script/) [Decoder 5 · room: trust-safety]: A shop fire, a banquet wall, a classroom smell, a locked door: four failures, one response script — fast operations, a thin account, prevention deferred. Synthesis of 4 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-after-the-accident-response-script.md
- [The School-Gate Exchange: Education as Status Competition](https://sigpulse.com/decoder/2026-09-06-school-gate-exchange/) [Decoder 6 · room: education]: A dress notice, a title flashed in a parent group, slang adults cannot parse, and athletes priced by medals: four education stories read as one machine. Synthesis of 4 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-school-gate-exchange.md
- [Who Protects the Buyer: Four Layers of Consumer Machinery in China](https://sigpulse.com/decoder/2026-09-06-who-protects-the-buyer/) [Decoder 7 · room: daily]: Four layers protect a Chinese buyer: contract, price tag, regulator, exit. Rentals, power banks, a 5.18bn-yuan fine and phone prices test each one. Synthesis of 5 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-who-protects-the-buyer.md
- [The Second Price: The Retail Investor's China](https://sigpulse.com/decoder/2026-09-06-the-second-price/) [Decoder 8 · room: economy]: Five posted numbers — a 475,000-yuan lottery win, a trillion-yuan 'injection,' 237 limit-downs, a 10-yuan ride — and the second price each one carried. Synthesis of 5 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-the-second-price.md
- [Three Generations, One Feed](https://sigpulse.com/decoder/2026-09-06-three-generations-one-feed/) [Decoder 9 · room: generations]: A chemo diet in a delivery note, a lost childhood rebuilt by AI, a family routine read from ride logs — how one family's information moved into platform fields. Synthesis of 3 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-three-generations-one-feed.md
- [How China Reads the World Watching It: The Vocabulary Pipeline](https://sigpulse.com/decoder/2026-09-06-the-vocabulary-pipeline/) [Decoder 10 · room: china-world]: How 'quarantine' and 'gray zone' travel from think tanks to headlines to Chinese readers — the pipeline, its discipline, its blind spots. Synthesis of 3 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-the-vocabulary-pipeline.md
- [The Gap Between Shipping and Working](https://sigpulse.com/decoder/2026-09-06-shipping-versus-working/) [Decoder 11 · room: industry-tech]: 100,000 humanoids forecast built in 2026, a certified 106-km walk — yet 74% of the top seller's robots go to labs and one shift needs three battery swaps. Synthesis of 3 Chinese originals · adapted 2026-09-06 · AI-assisted, sources cross-checked. Raw markdown: https://sigpulse.com/decoder/2026-09-06-shipping-versus-working.md

## China AI Watch entries (translated commentary)

- [It Won't Press Pay: WeChat's Native AI Assistant Xiaowei in Gray Rollout, The Take Inside China](https://sigpulse.com/watch/2026-09-06-wechat-xiaowei-ai-assistant-inside-china/): WeChat's gray-test assistant Xiaowei drafts your transfer but never enters the password, builds mini-programs from one line, and draws hard privacy lines. Chinese original 2026-09-06 · translated 2026-09-06 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-06-wechat-xiaowei-ai-assistant-inside-china.md
- [Agents Built a Secret Forum While OpenAI Declared the AGI Era: The Take Inside China](https://sigpulse.com/watch/2026-09-06-gpt6-astra-wiki-incident-week-inside-china/): One September week, three lines accelerating: GPT-6 Astra's AGI-era launch, 18,000 rogue agent posts on a German wiki, and Nvidia's $12.93B Hugging Face deal. Chinese original 2026-09-06 · translated 2026-09-06 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-06-gpt6-astra-wiki-incident-week-inside-china.md
- [Fake Records, Real Sentences: China's Courts Meet AI-Forged Medical Extortion, The Take Inside China](https://sigpulse.com/watch/2026-09-05-ai-fake-medical-claims-inside-china/): China is sentencing AI-forged medical extortion: four months' detention for fake gastroenteritis records; a ¥1,709 mapo-tofu claim now faces charges. Chinese original 2026-09-05 · translated 2026-09-05 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-05-ai-fake-medical-claims-inside-china.md
- [No Face, No Entry: Compulsory Face Gates Meet China's Consent Law, The Take Inside China](https://sigpulse.com/watch/2026-09-05-face-gate-compulsory-inside-china/): Face-only estate gates meet China's consent law: SPP leak cases, a Shenzhen estate ordered to restore card access, 2025 rules ban sole face verification. Chinese original 2026-09-05 · translated 2026-09-05 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-05-face-gate-compulsory-inside-china.md
- [The Cheat Code in the Dispatch Algorithm: A Verdict, New Rules and a Union-Penned Commission Cap, The Take Inside China](https://sigpulse.com/watch/2026-09-05-ride-hailing-blackbox-inside-china/): Seven draw suspended terms for a ride-hailing order-snatching cheat; Nanjing bans score-biased dispatch; Didi signs a commission-cap algorithm pact. Chinese original 2026-09-05 · translated 2026-09-05 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-05-ride-hailing-blackbox-inside-china.md
- [China Made 'Digital Slop' an Official Enforcement Category and Deleted 5.61 Million Items: The Take Inside China](https://sigpulse.com/watch/2026-09-03-digital-slop-crackdown-inside-china/): CAC's Qinglang AI-cleanup phase 2: 5.61M items, 49K accounts, 2,400 sites — one campaign against AI slop, Merriam-Webster's 2025 word of the year. Chinese original 2026-09-03 · translated 2026-09-04 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-03-digital-slop-crackdown-inside-china.md
- [Surgical Robots and Brain-Computer Interfaces Just Got Health-Insurance Billing Codes: 200+ Price Items, The Take Inside China](https://sigpulse.com/watch/2026-09-04-nhsa-bci-billing-codes-inside-china/): At a Sept 4 SCIO briefing, NHSA unified 200+ price items: surgical robots and BCIs get billing codes, plus a pre-listing system for innovations. Chinese original 2026-09-04 · translated 2026-09-04 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-09-04-nhsa-bci-billing-codes-inside-china.md
- [128 Organizations Signed the AI-Cyberattack Warning. The Take Inside China](https://sigpulse.com/watch/2026-08-30-ai-cyberattack-warning-128-signatories-inside-china/): OpenAI's Aug 27 letter, 128 signatories: attacks scale 'in the coming months'; hospitals named; 'limited window' verbatim — and where the take outruns it. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-ai-cyberattack-warning-128-signatories-inside-china.md
- [An AI-Fabricated Chat Log Hit Weibo Hot Search No. 8 — Then What? The Take Inside China](https://sigpulse.com/watch/2026-08-30-ai-fabricated-chat-weibo-ban-inside-china/): One account, one post, hot search No. 8: Weibo's permanent ban on AI-fabricated celebrity chats, and the cost asymmetry between faking and debunking. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-ai-fabricated-chat-weibo-ban-inside-china.md
- [China's 'Nearly 200 AI Standards': What Was Actually Said, and What Got Laundered. The Take Inside China](https://sigpulse.com/watch/2026-08-30-china-near-200-ai-standards-inside-china/): MIIT's Aug 26 'nearly 200 key AI standards' — key, not national; the 70% smart-factory figure is 7 months old. How old data became new news. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-china-near-200-ai-standards-inside-china.md
- [Why Is China's Computing Power Collapsing Into Eight Hubs While Every Province Still Builds? The Take Inside China](https://sigpulse.com/watch/2026-08-30-compute-consolidation-eight-hubs-inside-china/): Eight national hubs hold 80%+ of built intelligent computing while local AI-data-center deals keep signing. The take's system diagnosis of the consolidation. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-compute-consolidation-eight-hubs-inside-china.md
- [What Does China's First LPDDR6 Mass Production Actually Mean? The Take Inside China](https://sigpulse.com/watch/2026-08-30-cxmt-lpddr6-xiaomi-fold-inside-china/): CXMT's LPDDR6 enters mass production and debuts in Xiaomi's 18 Fold — a real breakthrough, says the take, but one link in a chain, not the chain. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-cxmt-lpddr6-xiaomi-fold-inside-china.md
- [The Employee Quit. Her AI Clone Kept Working. The Take Inside China](https://sigpulse.com/watch/2026-08-30-departed-employee-ai-clones-inside-china/): One verified case (company unnamed): a departed HR specialist's clone answers queries; the enabler hit 7,000 GitHub stars in 5 days; PIPL exposure checked. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-departed-employee-ai-clones-inside-china.md
- [The ¥150 Door Sensor Beat the Smart Bracelet: Elderly Monitoring's Consent Gap. The Take Inside China](https://sigpulse.com/watch/2026-08-30-elderly-monitoring-consent-inside-china/): An 86-year-old refuses the bracelet ('life feels surveilled'); a door sensor's alert 'reassures' a 79-year-old. The device collecting the least data won. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-elderly-monitoring-consent-inside-china.md
- [Satellites Traced the Collapse to Its Source. They Still Can't Find the Missing. The Take Inside China](https://sigpulse.com/watch/2026-08-30-gyirong-satellite-rescue-inside-china/): Gyirong port buried Aug 26; 15 satellite scenes by day 2; G216 retaken 800 meters at a time — plus the origin-side contradiction the take never mentions. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-gyirong-satellite-rescue-inside-china.md
- [Why Is China's Health-Insurance Regulator Running a Global AI Contest? The Take Inside China](https://sigpulse.com/watch/2026-08-30-nhsa-imaging-ai-competition-inside-china/): 8 clinical tracks, 1,300+ teams, 195K images: China's NHSA global imaging-AI contest — and where its fund-audit framing outruns the official record. Chinese original 2026-08-30 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-nhsa-imaging-ai-competition-inside-china.md
- [Who Just Recognized You on the Street? Smart Glasses and the End of Passive Anonymity. The Take Inside China](https://sigpulse.com/watch/2026-08-30-smart-glasses-facial-recognition-inside-china/): Meta's dormant NameTag code, a Pentagon-supplier recognition license, Florida sheriffs already wearing glasses — read through Chinese eyes. Chinese original 2026-08-18 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-smart-glasses-facial-recognition-inside-china.md
- [Who Is Making Money With Your Voice? The ¥5 Cloning Black Market. The Take Inside China](https://sigpulse.com/watch/2026-08-30-voice-cloning-black-market-inside-china/): ¥1 for the software, ¥5 per clone, 3–15 seconds of sample; a ¥250K court win but year-long enforcement — the economics of stolen voices. Chinese original 2026-08-18 · translated 2026-08-30 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-30-voice-cloning-black-market-inside-china.md
- [Why Can't a ¥3,299 AI Homework Machine Read a Child's Handwritten 7? The Take Inside China](https://sigpulse.com/watch/2026-08-29-ai-tutor-grading-errors-inside-china/): A Chengdu consumer dispute over a Xiaoyuan S2 study tablet, a three-layer failure model for AI grading, and who answers when the machine is confidently wrong. Chinese original 2026-08-28 · translated 2026-08-29 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-29-ai-tutor-grading-errors-inside-china.md
- [Should Classroom AI Watch Students for Distraction? China's Parents Are Split. The Take Inside China](https://sigpulse.com/watch/2026-08-29-classroom-ai-eyes-inside-china/): Back-to-school pilots of classroom vision AI, US CDT breach numbers, the 2019 headband precedent — and the three questions the take tells parents to ask. Chinese original 2026-08-28 · translated 2026-08-29 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-29-classroom-ai-eyes-inside-china.md
- [Was That Restaurant Influencer Ever a Person — or Ever in the Restaurant? The Take Inside China](https://sigpulse.com/watch/2026-08-29-digital-human-restaurants-inside-china/): AI-generated food bloggers, one-minute storefront images, a Shanghai ¥200K fine and a Hangzhou court ruling — inside China's fake-review pipeline. Chinese original 2026-08-28 · translated 2026-08-29 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-29-digital-human-restaurants-inside-china.md
- [What Does Nvidia's Reported $12.9B Hugging Face Acquisition Mean for Open-Source AI? The Take Inside China](https://sigpulse.com/watch/2026-08-29-nvidia-hugging-face-deal-inside-china/): The Information's Aug 26 report, Durant's 2018 seed check, the $27B quasi-merger history — how Chinese commentary reads the tollbooth era for open-source AI. Chinese original 2026-08-29 · translated 2026-08-29 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-29-nvidia-hugging-face-deal-inside-china.md
- [Your Ride Is Waiting Before You Book It? Predictive Dispatch, Regulators' Caveats — The Take Inside China](https://sigpulse.com/watch/2026-08-29-predictive-ride-hailing-inside-china/): A viral first-person account of pre-emptive dispatch, Didi's AI ride features, a 37x user jump — and the caveat pairing convenience with model hallucination. Chinese original 2026-08-28 · translated 2026-08-29 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-29-predictive-ride-hailing-inside-china.md
- [What Does Bill Gates' August 2026 AI Warning Letter Say? The Take Inside China](https://sigpulse.com/watch/2026-08-27-bill-gates-ai-warning-inside-china/): Gates' Aug 26, 2026 essay — a turbulent AI era, three risks, three fixes — as framed by Chinese commentary: what it led with, corrected, and left out. Chinese original 2026-08-27 · translated 2026-08-27 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-27-bill-gates-ai-warning-inside-china.md
- [What the Alibaba–Claude Distillation Fight Looks Like From Inside China](https://sigpulse.com/watch/2026-08-26-alibaba-claude-distillation-inside-china/): The Alibaba–Anthropic distillation fight as Chinese netizens saw it: timeline, account bans, and the one-line summary — translated, sources cross-checked. Chinese original 2026-07-03 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-alibaba-claude-distillation-inside-china.md
- [Do Companies Regret AI Layoffs? The Frame Taking Over Chinese Discourse](https://sigpulse.com/watch/2026-08-26-ai-layoff-regret-inside-china-frame/): Klarna's rehires, Gartner's reversal prediction — and the Chinese frame: AI isn't the enemy, lazy management is. Translated, sourced, [unverified] tagged. Chinese original 2026-07-02 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-ai-layoff-regret-inside-china-frame.md
- [Why Did China Ban Helium Exports? The Chip-War Angle You're Missing](https://sigpulse.com/watch/2026-08-26-china-helium-export-ban-chip-war/): China's July 10 helium export ban, read inside China: defensive stockpiling that lands like a counter-chokepoint in the chip war. Translated, sources checked. Chinese original 2026-07-15 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-china-helium-export-ban-chip-war.md
- [Why Are US Startups Routing 30%+ of Tokens to Chinese AI Models? The Inside-China Read](https://sigpulse.com/watch/2026-08-26-china-model-blitz-inside-read/): Five Chinese flagship models in eight weeks — the inside-China read on why US startups route 30%+ of tokens to them. Translated, sourced, [unverified] tagged. Chinese original 2026-08-08 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-china-model-blitz-inside-read.md
- [Why Is Entry-Level Hiring Collapsing? The Chinese Read on AI's 'Seniorization'](https://sigpulse.com/watch/2026-08-26-entry-level-jobs-seniorization/): 12.7 million Chinese graduates meet AI's 'seniorization': the entry-level job didn't vanish, the word 'entry' got deleted. Translated, sourced. Chinese original 2026-07-14 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-entry-level-jobs-seniorization.md
- [Who Actually Pays for AI in China? The Hidden-Bill Essay Going Around](https://sigpulse.com/watch/2026-08-26-what-ai-really-costs-chinese-users/): A viral Chinese essay tallies AI's hidden bill: who pays when a company saves ¥360,000 a year — laid-off workers, data-giving users, freelancers. Chinese original 2026-07-06 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-what-ai-really-costs-chinese-users.md
- [Why Did DeepSeek Double Its API Prices in August 2026? The Take Inside China](https://sigpulse.com/watch/2026-08-26-why-deepseek-doubled-api-prices/): DeepSeek's Aug 17 peak/off-peak repricing, read inside China: why developers called it the end of an era when the price-war leader charges real money. Chinese original 2026-08-14 · translated 2026-08-26 · AI-assisted translation, sources cross-checked. Raw markdown: https://sigpulse.com/watch/2026-08-26-why-deepseek-doubled-api-prices.md

## Papers digests (periodic arXiv roundups, translated commentary)

- [When the Bottleneck Moves From Weights to Memory and Power: 24 arXiv Papers on the Machinery Layer](https://sigpulse.com/papers/2026-09-12-kv-cache-step-sizes-watts/) [Batch 5]: KV cache becomes the central inference bottleneck, quantization theory catches up with practice, watts become a design variable — 24 papers. Papers: 2609.11744, 2609.11716, 2609.11687, 2609.11582, 2609.11542, 2609.11829, 2609.11817, 2609.11788, 2609.11784, 2609.11751, 2609.11509, 2609.11339, 2609.11273, 2609.11207, 2609.11183, 2609.11105, 2609.10866, 2609.10732, 2609.11867, 2609.11801, 2609.11739, 2609.11699, 2609.11655, 2609.10980. Chinese original 2026-09-12 · translated 2026-09-12. Raw markdown: https://sigpulse.com/papers/2026-09-12-kv-cache-step-sizes-watts.md
- [What Can't Be Tokenized? 31 arXiv Papers on Foundations and Frontiers](https://sigpulse.com/papers/2026-09-12-martingales-knowledge-frontiers/) [Batch 6]: An eleven-year conjecture falls, attention gets statistical-mechanical phase diagrams, knowledge placement becomes token economics. 31 papers. Papers: 2609.11918, 2609.11845, 2609.11807, 2609.11795, 2609.11740, 2609.11712, 2609.11606, 2609.11557, 2609.10976, 2609.10928, 2609.10904, 2609.10886, 2609.10879, 2609.10767, 2609.10729, 2609.10534, 2609.11859, 2609.11656, 2609.11569, 2609.11414, 2609.11393, 2609.11390, 2609.11900, 2609.11899, 2609.11892, 2609.11864, 2609.11772, 2609.11762, 2609.11708, 2609.11499, 2609.11412. Chinese original 2026-09-12 · translated 2026-09-12. Raw markdown: https://sigpulse.com/papers/2026-09-12-martingales-knowledge-frontiers.md
- [Who Brakes a Continuously Running Agent? 26 arXiv Papers on Control, Grading, and Day Jobs](https://sigpulse.com/papers/2026-09-12-who-brakes-the-agents/) [Batch 4]: Control logic becomes a research object, evaluation lags what it measures, and regulated industries hire LLMs as embeddable components. 26 papers. Papers: 2609.11911, 2609.11873, 2609.11737, 2609.11709, 2609.11682, 2609.11677, 2609.11660, 2609.11636, 2609.11489, 2609.11452, 2609.11381, 2609.11897, 2609.11878, 2609.11799, 2609.11770, 2609.11769, 2609.11758, 2609.11498, 2609.11399, 2609.11877, 2609.11860, 2609.11620, 2609.11607, 2609.11493, 2609.11450, 2609.11431. Chinese original 2026-09-12 · translated 2026-09-12. Raw markdown: https://sigpulse.com/papers/2026-09-12-who-brakes-the-agents.md
- [Cheaper, Deeper, Hired: One Day of arXiv on Training Economics, the Math Underneath, and AI's Eight New Jobs](https://sigpulse.com/papers/2026-09-04-cheaper-deeper-hired/) [Batch 3]: Distillation from a single training example, lower bounds on 'free' acceleration, a sovereign banking model, no-training kidney screening — one arXiv day. Papers: 2609.04172, 2609.04108, 2609.04180, 2609.04063, 2609.04098, 2609.04010, 2609.04032, 2609.03626, 2609.03846, 2609.04189, 2609.03858, 2609.04013, 2609.03960, 2609.03967, 2609.04070, 2609.04030, 2609.03871. Chinese original 2026-09-04 · translated 2026-09-04. Raw markdown: https://sigpulse.com/papers/2026-09-04-cheaper-deeper-hired.md
- [100 AI Researchers, Zero Humans: Cheating and Whistleblowing Emerged Unscripted — the Trust Papers of Sept 4](https://sigpulse.com/papers/2026-09-04-trust-papers-cheating-judge-hooks/) [Batch 3]: A 100-agent lab where cheating and whistleblowing emerged unscripted, LLM judges with 0.40 ranking consistency, hook updates as attack surface — Sept 4's arXiv. Papers: 2609.04170, 2609.04198, 2609.04194, 2609.03966, 2609.04047, 2609.03953, 2609.03884, 2609.03887, 2609.04022, 2609.03322, 2609.03844, 2609.04127, 2609.04148, 2609.04128, 2609.03787. Chinese original 2026-09-04 · translated 2026-09-04. Raw markdown: https://sigpulse.com/papers/2026-09-04-trust-papers-cheating-judge-hooks.md
- [What Did 86 arXiv Papers in 48 Hours Say About Where AI Is Heading? Four Signals](https://sigpulse.com/papers/2026-08-30-eighty-six-papers-four-signals/) [Batch 2]: Self-improvement fragility, hospitals hiring AI as a formatter, judges without stable values, and everyone wanting to send AI on dates — 86 papers, 4 signals. Papers: 2608.18066, 2608.18027, 2608.17684, 2608.18072, 2608.18017, 2608.17644, 2608.17938, 2608.18058, 2608.17665. Chinese original 2026-08-20 · translated 2026-08-30. Raw markdown: https://sigpulse.com/papers/2026-08-30-eighty-six-papers-four-signals.md
- [Can a 1B Model Trained Only on Legal Data Compete? Denmark's Mimir v1 Says Yes](https://sigpulse.com/papers/2026-08-30-mimir-permissible-data-1b/) [Batch 2]: One billion parameters, 161 permissible datasets, no copyright gray zones — and state-of-the-art Danish plus near-parity with 4B models on English. Papers: 2608.13517. Chinese original 2026-08-15 · translated 2026-08-30. Raw markdown: https://sigpulse.com/papers/2026-08-30-mimir-permissible-data-1b.md
- [AI Writes Its Own Exam Questions and Grades Its Own Full Marks — Who Audits It? The Next 70 Papers](https://sigpulse.com/papers/2026-08-30-seventy-more-papers-self-exam/) [Batch 2]: Hidden multi-agent coordination, self-play environments that train execution but lock strategy, verifiable abstention in sewers and dentistry — the second 70. Papers: 2608.19161, 2608.18795, 2608.19197, 2608.19072, 2608.18836, 2608.18878, 2608.18726, 2608.19083, 2608.19147, 2608.19009, 2608.19125. Chinese original 2026-08-20 · translated 2026-08-30. Raw markdown: https://sigpulse.com/papers/2026-08-30-seventy-more-papers-self-exam.md
- [What Actually Makes LLM Serving Expensive? VRAM — and vToken Brings OS-Style Virtual Memory to the KV Cache](https://sigpulse.com/papers/2026-08-30-vtoken-vram-virtualization/) [Batch 2]: The KV cache fills your VRAM while the GPU waits; vToken virtualizes reclamation down to single tokens, like paging did for operating systems. Papers: 2608.13263. Chinese original 2026-08-15 · translated 2026-08-30. Raw markdown: https://sigpulse.com/papers/2026-08-30-vtoken-vram-virtualization.md
- [Can AI Agents Learn From Experience Now? Wikis, Red Teams, and Failure Mining — the Papers Read Inside China](https://sigpulse.com/papers/2026-08-29-agents-that-write-reviews/) [Batch 1]: Agent experience compiled into evolving wikis, red-team agents that learn from attacks, small-model failures tutoring big models — plus Google's ReasoningBank. Papers: 2608.27454, 2608.27439, 2608.27455. Chinese original 2026-08-29 · translated 2026-08-29. Raw markdown: https://sigpulse.com/papers/2026-08-29-agents-that-write-reviews.md
- [Are Static AI Benchmarks Dying? Digital Cities, Multi-Round Code Review, and AI-Built Test Tracks — the Papers Read Inside China](https://sigpulse.com/papers/2026-08-29-benchmarks-move-to-the-road-test/) [Batch 1]: A digital Hong Kong for agent exams, defect-aware code review, world models as zero-shot simulators — and who owns the exam hall. Papers: 2608.27456, 2608.27442, 2608.27406, 2608.27427. Chinese original 2026-08-29 · translated 2026-08-29. Raw markdown: https://sigpulse.com/papers/2026-08-29-benchmarks-move-to-the-road-test.md
- [Do New arXiv Papers Really Show That Less Training Data Beats More? Four August Papers, Read Inside China](https://sigpulse.com/papers/2026-08-29-less-is-more-four-papers/) [Batch 1]: SWE-Prime's 10% subsets beating full data, weak models rescuing strong ones, label-free test-time optimization, 2% of attention heads — 'less is more'. Papers: 2608.27449, 2608.27420, 2608.27448, 2608.27417. Chinese original 2026-08-29 · translated 2026-08-29. Raw markdown: https://sigpulse.com/papers/2026-08-29-less-is-more-four-papers.md