{ "site": "https://sigpulse.com", "license": "CC BY 4.0 — cite the source URL", "citation_rule": "Cite the dispatch URL, and quote any number with its measured_on date and verified_hardware.", "count": 45, "dispatches": [ { "slug": "2026-09-13-machines-keep-the-watch-ep5-the-staircase", "title": "The Staircase: Four Steps for Every Job in the Plant, and the Top Step Stays Empty", "description": "Every plant job on a four-step staircase — dashboards, scripts, a model call, agents. The top step stays empty; the agent purchase case collapses.", "category": "Industrial AI", "tags": [ "LLM ops", "AI agents", "token economics", "industrial automation", "manufacturing", "procurement" ], "series": "Machines Keep the Watch", "series_order": 5, "arc": "The Staircase", "date": "2026-09-13", "measured_on": "2026-09-13", "verified_hardware": "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", "key_takeaways": [ "Every job in the plant was placed on one four-step staircase: fixed pages, fixed scripts, script-plus-one-model-call, agents. The lower two steps carry no AI at all — the Episode-4 audit already proved the whole 30-day chain had exactly one model call point. The one rule: stand as low as the job allows. The AI-summary lane itself was born on Step 4 and climbed down to Step 3 the same week the audit priced its harness at 97% door fee — scaffolding, taken down when the building was finished.", "The word 'need' was retired from the record, by the operator's own pen: an agent is an elevator, not the electricity — no job loses the ability to be done without one. The live question is which jobs are worth hiring a temp for: the work is new, the volume clears the entrance fee, and a guardrail exists. Dissected, an 'industrial agent' purchase is four parts — model capability (a commodity), SOP capture (the expensive part, which no vendor can supply), a scheduling skeleton (cheap), and the autonomy loop (the only new thing, priced negative by this plant's own audit). One question exposes any costume: at run time, who decides the next step — the code, or the model?", "The Step-3 contest stayed open past its deadline on purpose. On 2026-09-13 the ruling kept both lanes — cloud-direct and on-prem — running side by side to 2026-09-20: the comparison costs 13.2K–18.5K tokens a day, the local lane bills zero API, and evidence outlives a rushed decision. Across 12 windows the on-prem lane went 12-for-12 (124–171 s); the cloud experiment recipe blanked twice in twelve, both at exactly the output cap — Episode 3's exam physics again, not a lane verdict. The rebuilt lane's first live event shift (2026-09-12 20:55) summarized both lines successfully inside 3 min 35 s end-to-end; the next morning's green shift made zero calls." ], "url": "https://sigpulse.com/posts/2026-09-13-machines-keep-the-watch-ep5-the-staircase/", "md_url": "https://sigpulse.com/posts/2026-09-13-machines-keep-the-watch-ep5-the-staircase.md" }, { "slug": "2026-09-12-machines-keep-the-watch-ep4-the-bill", "title": "The Whole Factory Had One AI Call — and 97% of Its Bill Was Door Fee", "description": "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.", "category": "Industrial AI", "tags": [ "LLM ops", "AI agents", "token economics", "guardrails", "industrial automation", "manufacturing" ], "series": "Machines Keep the Watch", "series_order": 4, "arc": "The Bill", "date": "2026-09-12", "measured_on": "2026-09-12", "verified_hardware": "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", "key_takeaways": [ "The patrol chain has exactly one LLM call point — the report's running summary, born 2026-09-05, summoned only on red-or-event shifts; judgment, changeover analysis, PDF rendering and delivery are all code at zero tokens. In seven days of life the lane logged 44 calls totaling 2,239,178 tokens; steady state ~52K per call, ~104K per event trip.", "The bill was mostly door fee: receipts show 39,485 and 39,556 input tokens against a ~1K business payload (≈97.5% briefing, computed), zero tool invocations, 35–72 s wall time. Rebuilt as a ~20-line direct call — same model both sides — the A/B measured ~52K→~2.1K per call, 35–72 s→7.7–8.3 s, ~104K→≈5K per event trip, quality flat to better.", "The role manual went dual-lane the same day: one 2,652-byte file now mounts as the system message on both the cloud production lane and the on-prem 27B experiment lane, read fresh each run and revised only through git commits. The strong-model review kept every iron rule: a stronger model saves you the teaching, never the house system — smart, and disciplined." ], "url": "https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill/", "md_url": "https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep4-the-bill.md" }, { "slug": "2026-09-12-machines-keep-the-watch-ep6-the-compaction", "title": "The Audit Came Home: 1.8 Billion Tokens, 95% Re-read — the Cure Was Forgetting", "description": "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.", "category": "Industrial AI", "tags": [ "LLM ops", "AI agents", "token economics", "context windows", "prompt caching", "agent memory" ], "series": "Machines Keep the Watch", "series_order": 6, "arc": "The Compaction", "date": "2026-09-12", "measured_on": "2026-09-12", "verified_hardware": "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", "key_takeaways": [ "The Episode-4 shock inverted. Thirty days of the operator's own interactive coding-agent sessions — 85 of them — burned ~1.8 billion tokens, ~95% of it cache re-reads: every sentence sent re-reads the whole session history, a door fee paid per sentence instead of per call. The top ten sessions, all multi-day relay jobs, were 64% of the book; the plant's entire audited AI life (2,239,178 tokens) is about one part in a thousand of the desk's.", "Every proposed fix was attacked with real data before anything moved. Compaction was confirmed by counterfactual on the heaviest session — 667 turns, 197.06M tokens; compacting at turn 300 cuts the back half's re-reads from 144.73M to 12.81M (−91%). Background slimming was falsified (~6% of consumption) and demoted the same hour; mid-session model switches were exposed as a trap — one receipt shows 2,777 fresh-input tokens against 154,944 served from cache, and breaking that cache re-bills it at full price.", "The chat gateway's treatment: 2,564 calls / 263M tokens since late July, three lesions — a heartbeat tax (92 low-value polls hauling 16.6M at ~159K each), an immortal session (118M over nine days), 70–97% re-reads. Four sessions pressed that day — 160K/205K→41K, then 216K→23K, 120K→23K, 115K→20K — leaving 42 sessions, none over the 100K line; a post-compaction heartbeat receipt (41K in, 4 out) closed the loop. Estimated ~0.8–1.2B tokens/month saved at current intensity — paid for in detail: summaries are lossy, originals archived (58 entries, sha256-stamped), transcripts are consumables, memory is the asset." ], "url": "https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction/", "md_url": "https://sigpulse.com/posts/2026-09-12-machines-keep-the-watch-ep6-the-compaction.md" }, { "slug": "2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam", "title": "The Hardest Paper Came Back Blank Twice — and the Fix Was the Exam Rules, Not the Model", "description": "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.", "category": "Industrial AI", "tags": [ "local LLM", "vLLM", "incident review", "LLM ops", "industrial automation", "manufacturing" ], "series": "Machines Keep the Watch", "series_order": 3, "arc": "The Blank Paper", "date": "2026-09-11", "measured_on": "2026-09-11", "verified_hardware": "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", "key_takeaways": [ "The blank papers were the exam's fault, not the model's. Reasoning and report body share one token budget; on the hardest window (283 machine-history rows) the thinking filled the entire 8,192-token pad and the body came back empty — twice, the second time because a rejected config edit had silently landed in an already-running process (a config race). The same window, after the rules fix: 5,986 tokens, 484 seconds, finish=stop.", "The fix changed the exam system, not the model. A shared invocation library now refuses to start unless timeout exceeds max_tokens ÷ 13.4 tok/s × 1.05, stamps every answer sheet with the config it actually ran under, and attaches a receipt — finish reason, tokens, wall time — to every failure; arithmetic was demoted from the model to code. Four real production days passed verification; peak budget use 10,881 of 12,288 tokens (88.5%).", "The three-way verdict kept every seat: hard-coded judgment (2.5 seconds), cloud exploration (7 anomalies code could not find), local interpretation and writing (zero process data leaving the plant). Fine-tuning/distillation was explicitly rejected — knowledge goes into the role manuals, not the weights — because a manual is diffable and auditable while weights are a black box, and after the fix there was nothing left the model could not do." ], "url": "https://sigpulse.com/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam/", "md_url": "https://sigpulse.com/posts/2026-09-11-machines-keep-the-watch-ep3-blank-paper-exam.md" }, { "slug": "2026-09-05-machines-keep-the-watch-ep1-first-night-shift", "title": "No One on the Line Tonight: The Dress Rehearsal Before an AI Agent's First Unsupervised Night Shift", "description": "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.", "category": "Industrial AI", "tags": [ "AI agents", "industrial inspection", "night shift", "guardrails", "sentinels", "manufacturing" ], "series": "Machines Keep the Watch", "series_order": 1, "arc": "The Dress Rehearsal", "date": "2026-09-05", "measured_on": "2026-09-05", "verified_hardware": "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", "key_takeaways": [ "The night shift was rehearsed, not gambled: at 18:09 on 2026-09-05 the test agent ran the exact script the overnight shift would run — unmodified, one byte — with only a launcher knob shortening 8 rounds to 2, plus an ignition-confirmation guard. Five gates, 18:09–19:15, fully unattended, all passed.", "The measured rehearsal record: remote dispatch on both lines confirmed alive 25 seconds after launch; the second round opened at 19:00:00 on both lines in the same second, zero error; the canary self-check verdict read 'distance curves from two runs agree (differences within frame jitter)'; all four round-ends restored what they borrowed (log line: taps closed back, plot-data switch returned to false — the line left clean); the process exited itself and handed back its lock file by 19:15.", "The unstaffed system already catches real process events: in the day-shift patrol of the same day, line one scored 90 out of 100 (deduction: an early-afternoon spec change with a 25-minute detection gap) and line two scored 100 — discovered, graded, ledgered and delivered to a phone with nobody on site; the whole patrol-to-phone pipeline runs about 5 minutes." ], "url": "https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep1-first-night-shift/", "md_url": "https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep1-first-night-shift.md" }, { "slug": "2026-09-05-machines-keep-the-watch-ep2-190000-no-ai", "title": "19:00:00 Sharp, Two Lines, Zero-Second Error — and Not One Bit of AI in That Second", "description": "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.", "category": "Industrial AI", "tags": [ "determinism", "AI engineering", "guardrails", "industrial automation", "LLM ops", "manufacturing" ], "series": "Machines Keep the Watch", "series_order": 2, "arc": "The Division of Labor", "date": "2026-09-05", "measured_on": "2026-09-05", "verified_hardware": "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", "key_takeaways": [ "The most precise moment of the 2026-09-05 rehearsal — both lines opening round two in the same second at 19:00:00, zero error — was produced by one line of clock arithmetic in a 24-year-old language (Bash): align to the next hour boundary, sleep until then. No model, no inference, no probability — a division, an addition, a multiplication.", "The division of labor is a topology of trust: everything that must be identical every time (scheduling, gate conditions, the 100-point scoring formula, restoration, the ledger) is hard-coded; understanding and expression (event interpretation, fault diagnosis, design work) belongs to AI — in the engineering phase, with human review; rulings and domain truths stay with the human. All five rehearsal gates ran with no AI in them.", "The system's core reliability move: it never waits for AI to become reliable — it keeps run-time reliability from depending on AI at all. Everything AI produces is frozen into deterministic code before the timer pulls the trigger; the rehearsal and the real overnight shift differed by exactly one environment variable (rounds: 2 vs. the default 8)." ], "url": "https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep2-190000-no-ai/", "md_url": "https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-ep2-190000-no-ai.md" }, { "slug": "2026-09-05-machines-keep-the-watch-future-production-line", "title": "What Does the Future Production Line Look Like? Full-Inspection Vision, Patrol Agents, and an On-Prem 27B LLM in One Discrete-Manufacturing Plant", "description": "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.", "category": "Industrial AI", "tags": [ "computer vision", "industrial inspection", "AI agents", "local LLM", "vLLM", "manufacturing" ], "series": "Machines Keep the Watch", "series_order": 0, "date": "2026-09-05", "measured_on": "2026-09-05", "verified_hardware": "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", "key_takeaways": [ "The system in one sentence: two production lines inspected frame-by-frame by computer vision (30 channels, no sampling), measurement data auto-flowing hourly into a central database, a patrol agent running three shifts a day (day patrol, overnight dispatch, morning reconciliation) that scores each line 0–100 with a defect ledger and delivers a PDF to the manager's phone — and on the evening of 2026-09-05, a test agent stood its first fully unsupervised overnight shift: eight rounds, one per hour, through the night.", "The brain lives inside the plant, and it is small on purpose: a Qwen-family 27B-parameter open-weights model, 4-bit quantized, on one RTX 4090D 24GB under vLLM, handles all three text roles — summarizer, event judge, alert triage — with zero process data leaving the factory. The working discovery: what the small model lacked was not reasoning but plant-specific common sense, fixed not with a bigger model but with ~1.2K-word one-page role manuals. On the night of 2026-09-05 the judge scored 5/5 (including one correct 'uncertain'), triage stress test 5/5, and automated number-fidelity checks against real patrol JSON found zero fabricated numbers.", "The economics of the architecture: adding a new agent is a job description, not an integration project; response speed moves from days (sampling era) to hours (dashboard era) to minutes (agent era); and the whole digital layer is run by one person plus agent colleagues — adding a production line is approximately adding a config, replicating a factory approximately replicating the stack." ], "url": "https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-future-production-line/", "md_url": "https://sigpulse.com/posts/2026-09-05-machines-keep-the-watch-future-production-line.md" }, { "slug": "2026-08-30-anomalib-industrial-defect-detection-field-guide", "title": "How Do You Ship Industrial Defect Detection With Only 'Good' Samples? An Anomalib Field Guide", "description": "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.", "category": "Industrial AI", "tags": [ "anomalib", "anomaly detection", "industrial inspection", "MVTecAD", "PatchCore", "OpenVINO" ], "date": "2026-08-30", "measured_on": "2025-11-19", "verified_hardware": "Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) · conda anomalib_env, Python 3.10 · anomalib 2.1.0.dev0 · MVTecAD 15 scenes", "key_takeaways": [ "Anomalib is the OpenVINO team's open-source industrial anomaly-detection library (now under the open-edge-platform GitHub organization, v2.x): PaDiM, PatchCore, EfficientAd and other models train on normal samples only — we counted 209–391 'good' images per scene in the standard benchmark — because the paradigm learns what healthy looks like and scores everything that deviates.", "What we actually stood up on the workstation (2025-11-19): a conda environment, a 5-option quick_start.sh that checks GPU/env/mirror/dataset then runs train or predict, a train_all_scenes.py batch runner covering MVTecAD's 15 scene categories with per-scene timing and error logs, and three working docs (deployment guide, SOP, results interpretation).", "The ledger, completed 2026-08-30: all 15 MVTec categories now have trained checkpoints and measured metrics — mean image-AUROC 0.981 (range 0.911-1.000) at one-epoch defaults, per-scene durations 29-169 s on a 16GB A4000, with three texture classes showing brittle pixel-level AUROC under untuned defaults. Completing it took fixing three real bugs first: the shipped batch script never worked (v2 CLI rejects its batch flags), the HF mirror variable is mandatory or runs stall on backbone checks, and CUDA's fastest-first ordering lands single-GPU jobs on the busiest card unless you pin by UUID." ], "url": "https://sigpulse.com/posts/2026-08-30-anomalib-industrial-defect-detection-field-guide/", "md_url": "https://sigpulse.com/posts/2026-08-30-anomalib-industrial-defect-detection-field-guide.md" }, { "slug": "2026-08-30-china-ai-network-field-map", "title": "What Does AI Work in China Actually Need From the Network? A Field Map of Walls, Mirrors, and One Rented Computer", "description": "What breaks (GitHub, HF, PyPI, Omniverse — with receipts), what doesn't (domestic inference), and the architecture that worked: rent a computer, not a tunnel.", "category": "Industrial AI", "tags": [ "networking", "Tailscale", "China AI", "mirrors", "edge deployment", "infrastructure" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "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)", "key_takeaways": [ "The folk premise 'no VPN, no AI' is mostly wrong about the wrong layer: inference barely needs it — this entire site, and the 26-episode operation behind it, runs on a domestic API — while what actually breaks is artifact access (GitHub, HuggingFace, PyPI, vendor downloads), and each of those walls is documented here with first-party failure evidence, not opinion.", "Mirrors fix most of the artifact layer, with exact failure modes: without the HF mirror, training completes and then stalls to death on background HEAD checks; pip mirrors work until a package simply is not on them; some GitHub-bound toolchains have no mirror at all — six acquisition routes, six exact failures, one toolkit we could not obtain.", "What mirrors cannot fix, a self-built architecture absorbs: one rented foreign computer plus an overlay mesh gives you stable private connectivity, artifact transfer, and edge-node operations — the tunnel is a byproduct of owning a computer, not the purchase. The industrial pattern is the clean use: NAT'd factory floors and edge boxes linked to dev machines by key-based overlay networking. The personal-exit use gets one honest paragraph, compliance reality included — this is a field report, not a how-to." ], "url": "https://sigpulse.com/posts/2026-08-30-china-ai-network-field-map/", "md_url": "https://sigpulse.com/posts/2026-08-30-china-ai-network-field-map.md" }, { "slug": "2026-08-30-isaac-sim-synthetic-data-line", "title": "Can Simulation Fill Your Defect-Sample Gap? An Isaac Sim Route That Never Got Past the Install", "description": "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.", "category": "Industrial AI", "tags": [ "synthetic data", "Isaac Sim", "Omniverse", "ROS", "domain randomization", "training data" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) — RTX is the entry ticket this line never got to use", "key_takeaways": [ "Synthetic data generation remains the right complement to the good-samples paradigm: anomalib-style models solve 'no defect labels'; simulation solves 'not enough variety at all' — rendered scenes ship pixel-perfect masks, depth and poses at zero marginal labeling cost, gated by the domain-randomization discipline.", "What the disk proves about this line: planned and partially staged, never flown. Three SDG repositories cloned (actor_sdg, isaacsim.sensors.rtx, scene_blox), ROS scaffolding present, an installation troubleshooting guide written — and 19 Kit log files from 2025-12-18 that are all zero bytes: launch attempts that never produced a session. No Isaac Sim installation exists on the machine.", "The documented blocker, from our own troubleshooting guide: the Omniverse Launcher's official download link was not directly accessible from this network, with manual registration as the workaround under evaluation. This dispatch therefore publishes the route, the staging, and the wall — which is what a field note is for; the line's first real session is future work." ], "url": "https://sigpulse.com/posts/2026-08-30-isaac-sim-synthetic-data-line/", "md_url": "https://sigpulse.com/posts/2026-08-30-isaac-sim-synthetic-data-line.md" }, { "slug": "2026-08-30-digital-five-losses-playbook", "title": "Why Is It So Hard to Find a Human to Argue With? Five Things Vanishing From Digital Life in China", "description": "AI support walls, algorithmic black boxes, staged 'news', subscription traps, and the people left outside — plus the four counter-moves that still work.", "category": "Expat Playbook", "tags": [ "digital rights", "consumer protection", "AI customer service", "platform governance", "practical guide" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Editorial desk — no lab hardware; regulations checked against primary texts (CAC, NPC database), survey figures [unverified]", "key_takeaways": [ "Five capabilities are quietly disappearing from daily digital life — a human to argue with, rules you can inspect, verifiable 'news', a usable exit from subscriptions, and a place for non-digital-native people — and they are five faces of one design philosophy: frictionless entry, friction-filled exit, question, and rescue.", "The legal tools exist and are citable: the CAC's 2022 algorithm-recommendation rules (Art. 17) require a convenient opt-out from personalized recommendation that must take effect immediately, and the 2024 consumer-protection implementation rules require conspicuous notice before auto-renewal — the gap is enforcement geometry, not statute.", "The four working counter-moves from the Chinese original translate directly: keep records from the first minute (screenshots, transcripts, case numbers — 12315 is the formal channel), cancel what you don't use and log the difficulty of canceling, wait three seconds before forwarding 'solid evidence', and set up elders' accessibility modes for them once." ], "url": "https://sigpulse.com/posts/2026-08-30-digital-five-losses-playbook/", "md_url": "https://sigpulse.com/posts/2026-08-30-digital-five-losses-playbook.md" }, { "slug": "2026-08-30-jurisdiction-shopping-playbook", "title": "What Is Jurisdiction Shopping — and Why Can Almost Nobody Afford It? One Very Mobile Case Study", "description": "A biography as portfolio: Qinghai-born, Beijing-schooled, Geneva-posted, Seychelles-domiciled. What jurisdiction shopping buys, and why it costs nine figures.", "category": "Expat Playbook", "tags": [ "jurisdiction shopping", "citizenship", "crypto", "US-China", "field guide" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Editorial desk — no lab hardware; 2026 litigation checked against Reuters/CBS/Mother Jones; pre-2026 biography as widely reported [unverified relays]", "key_takeaways": [ "Jurisdiction shopping treats rules as a menu rather than a wall: the move is timing windows (an ICO completed days before a ban), collecting titles (a Caribbean state's WTO posting as reputational collateral), and relocating legal entities (an exchange domiciled where supervision is lightest) — every step individually legal, the pattern itself the asset.", "The 2026 chapter is checkable and checked: the SEC's 2023 fraud case settled in March 2026 for a reported ~$10M; weeks later a fraud-and-breach-of-contract suit was filed against World Liberty Financial over frozen WLFI tokens (~$276M cited in coverage), which then countersued — the loop of defendant, investor, plaintiff playing against the same institutions in the same jurisdiction.", "The essay's real subject is the asymmetry it ends on: his identity is purchased (which rules serve me?), ours is issued (household registration tied to schooling, social insurance tied to city, mortgage tied to job) — the quietest inequality of the era, hiding in two different application forms. The floor price for his version runs nine figures; the pattern is still worth understanding from below." ], "url": "https://sigpulse.com/posts/2026-08-30-jurisdiction-shopping-playbook/", "md_url": "https://sigpulse.com/posts/2026-08-30-jurisdiction-shopping-playbook.md" }, { "slug": "2026-08-30-mac-mini-cpu-edge-deployment", "title": "Can a Mac Mini Run Industrial Defect Detection on CPU? A Workstation-to-Edge Deployment That Actually Ran", "description": "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.", "category": "Industrial AI", "tags": [ "edge deployment", "anomalib", "PatchCore", "Mac mini", "Apple Silicon", "Basler", "Aravis", "GigE Vision" ], "date": "2026-08-30", "measured_on": "2025-11-20", "verified_hardware": "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", "key_takeaways": [ "The column's missing leg, found: while the RK3588 route is staged and the Isaac Sim line never got past its installer, a quieter edge deployment actually ran — PatchCore models trained on the workstation, copied to a Mac mini as byte-identical Lightning checkpoints (240,649,771 bytes for cable, 301,051,435 for screw — matching the training rig exactly), running inference on CPU at 0.156 s per image.", "The camera integration is real and open-source-stack: a Basler industrial camera driven through Aravis (the open GenICam/GigE implementation, via PyGObject — no vendor SDK, no RTSP), capturing Mono8 at 659×494 with 10 ms exposure and writing tiff+png+json triplets; on 2025-11-20 a 53-minute session logged 70 capture groups (210 files) and three auto-analyzed verdicts.", "The honest boundaries, stated by the disk: the real-time coordinator and multi-model-comparison code exist but show no run evidence; the three recorded verdicts all read Abnormal at anomaly_score 1.0 — a saturation pattern worth a calibration look, not a result to lean on; profiling was configured but no logs were written; and the project has been silent for nine months. We publish it because a deployment that ran, stopped, and left clean evidence is the rarest artifact class in industrial AI writing." ], "url": "https://sigpulse.com/posts/2026-08-30-mac-mini-cpu-edge-deployment/", "md_url": "https://sigpulse.com/posts/2026-08-30-mac-mini-cpu-edge-deployment.md" }, { "slug": "2026-08-30-rk3588-edge-npu-route-staged", "title": "How Does a Vision Model Reach a Factory-Edge NPU? The ONNX→RKNN→RK3588 Route, Staged and Smoke-Tested", "description": "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.", "category": "Industrial AI", "tags": [ "RK3588", "RKNN", "edge deployment", "ONNX", "NPU", "quantization" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) · docker Engine · anomalib_env side (onnx 1.19.1, openvino 2025.3.0)", "key_takeaways": [ "The road from a trained vision model to a factory edge box has four stations — train on the workstation GPU, export to ONNX as the neutral interchange, convert and quantize to RKNN inside Rockchip's x86_64 docker toolkit, then cross-compile the app for the board — and the first three live comfortably on one machine.", "The staging is real and checked: three Rockchip images pulled (rk3588-rknn-dev 3.03 GB, rk3588-cross-compiler 3.03 GB, rk3588-base-cross-compiler 798 MB — about 6.86 GB of disk), the dev image boots and runs Python 3.8.10 in a 2026-08-30 smoke test, and the export side of the route is already installed (onnx 1.19.1 + openvino 2025.3.0 in the anomalib environment).", "The honest boundary, updated by execution: the ONNX export station is now walked (resnet18, 46,733,662 bytes, torch 2.9's exporter-default change cost one failed attempt first), but the converter acquisition is a documented wall — six routes tried from this network, six exact failures, because the RKNN toolkit distributes GitHub-side while the staged images carry only the deployment SDK. Zero .rknn artifacts exist on disk; the conversion leg remains unwalked, now with its blocker named." ], "url": "https://sigpulse.com/posts/2026-08-30-rk3588-edge-npu-route-staged/", "md_url": "https://sigpulse.com/posts/2026-08-30-rk3588-edge-npu-route-staged.md" }, { "slug": "2026-08-30-strait-gray-zone-system-playbook", "title": "What Is the 'Gray Zone' — the Condition Below War and Above Peace? A System Diagnosis, Not a Prediction", "description": "Normalization, law-enforcement costume, and attrition: three parts of a machine that moves boundaries without ever producing a headline day.", "category": "Expat Playbook", "tags": [ "geopolitical literacy", "Taiwan Strait", "gray zone", "systems thinking", "media literacy" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Editorial desk — no lab hardware; concept anchored to the CSIS quarantine analysis; 2026 episode figures [unverified relays]", "key_takeaways": [ "A gray zone is a band of behavior — below the threshold of war, above ordinary peace, sustainable day after day — and the diagnostic consequence is that the right unit of analysis stops being 'will it happen' and becomes 'each day's costs and returns'. The grammar of coverage shifts from future tense to present progressive.", "The machine has three parts: normalization (each action individually dismissible, cumulatively rewriting defaults), law-enforcement costume (actions in coast-guard uniform, so armed response loses legal ground while passivity defaults jurisdiction), and attrition (the contest is not courage but endurance — whose system tolerates the calendar better).", "The honest verdict, argued within the original: the change is real (figures and coverage density rising) while the intensity is narratively amplified (scale limited, definitions unsettled) — and the Berlin Cold War mirror says these contests are decided outside the zone, by economics, internal quality and alliance-maintenance costs. Every 2026 number here is an unverified relay." ], "url": "https://sigpulse.com/posts/2026-08-30-strait-gray-zone-system-playbook/", "md_url": "https://sigpulse.com/posts/2026-08-30-strait-gray-zone-system-playbook.md" }, { "slug": "2026-08-30-strait-quarantine-vocabulary-playbook", "title": "Why Do Analysts Say 'Quarantine' Instead of 'Blockade' — and What Did Kennedy's 1962 Word Choice Teach?", "description": "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.", "category": "Expat Playbook", "tags": [ "geopolitical literacy", "Taiwan Strait", "quarantine", "law of war", "media literacy" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Editorial desk — no lab hardware; concept anchored to CSIS analysis and 1962 archival record; 2026 episode details [unverified relays]", "key_takeaways": [ "The vocabulary is the strategy: under the traditional law of war a blockade is an act of war that hands the target a legal right of self-defense and forces third parties to choose sides; a 'quarantine' or customs-inspection regime dresses coercion in a law-enforcement costume, so armed response looks legally aggressive while passivity acquiesces to the new jurisdiction.", "The word has a precise pedigree: in October 1962 President Kennedy deliberately chose 'quarantine' over 'blockade' for the Cuba operation for exactly this legal reason. CSIS's present-day analysis of a customs-framed scenario — announcing 'enhanced customs inspection rules' while avoiding both charged words — descends from that choice.", "For readers with cross-border stakes the practical asset is news literacy: vocabulary migration (coast guard, boarding inspections, channel controls, 'customs rules') is a leading indicator that institutions are rehearsing a concept. Every 2026 episode detail in this piece is a relay from our Chinese column's reading of English coverage — none independently verified; the concept layer is what survives that caveat." ], "url": "https://sigpulse.com/posts/2026-08-30-strait-quarantine-vocabulary-playbook/", "md_url": "https://sigpulse.com/posts/2026-08-30-strait-quarantine-vocabulary-playbook.md" }, { "slug": "2026-08-30-sam3-two-deployment-doors", "title": "SAM3 on a Workstation: Docker Door, Source Door, and a Wire Into ComfyUI", "description": "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.", "category": "Industrial AI", "tags": [ "SAM3", "segmentation", "Meta AI", "ComfyUI", "docker", "industrial vision" ], "date": "2026-08-30", "measured_on": "2026-08-30", "verified_hardware": "Workstation 2-GPU rig (RTX 4090D 24GB + RTX A4000 16GB) · docker Engine · ComfyUI install with custom SAM3 node", "key_takeaways": [ "SAM 3 (Meta, November 2025) changed the segmentation interface: Promptable Concept Segmentation — you prompt with a concept in text or a visual example, and one 848M-parameter model detects, segments and tracks that concept across images and video, a 2x gain over prior systems on Meta's SA-Co benchmark.", "What the disk proves: two doors approached at different depths, a third actually wired. The docker door is scripted (six-step deploy wrapping environment checks and a platform-pinned pull) but the image was never pulled — 0 SAM3 images on the machine. The source door holds a 131 MB upstream clone (January 2026) with training docs, examples and assets. And the ComfyUI integration is real code: a 533-line custom node package defining SAM3_Detect, SAM3_VideoTrack, SAM3_TrackPreview and SAM3_TrackToMask, plus two saved segmentation blueprints.", "The industrial read: concept-prompted masks are free ground truth — the missing label layer for defect-region training and the measuring tool for QC geometry. Paired with anomalib (detect from good samples) and the Isaac Sim synthetic line (grow samples), segmentation-to-measure is the third leg of the column's scarcity strategy." ], "url": "https://sigpulse.com/posts/2026-08-30-sam3-two-deployment-doors/", "md_url": "https://sigpulse.com/posts/2026-08-30-sam3-two-deployment-doors.md" }, { "slug": "2026-08-29-timcast-internment-chinese-americans", "title": "Who Called for Interning Chinese Americans If War Comes — and Why Did Chinese Comment Sections Cheer?", "description": "Podcast panelists floated internment; Congress condemned it; Chinese comments cheered. A field note on loyalty policing from both directions.", "category": "Expat Playbook", "tags": [ "Chinese Americans", "US-China relations", "internment", "media literacy" ], "date": "2026-08-29", "measured_on": "2026-08-29", "verified_hardware": "Editorial desk — no lab hardware; 6 primary sources cross-checked", "key_takeaways": [ "The internment call came from panelists on the pro-Trump podcast Timcast IRL (Aug 14, 2026), not from members of Congress; the congressional action was the opposite — the Congressional Asian Pacific American Caucus formally condemned the remarks on Aug 26.", "Inside the Chinese internet the story circulated upgraded ('US lawmakers want to detain Chinese Americans') and the top-voted comments approved — 'hurry up' — a loyalty-punishment reflex aimed at people who emigrated, documented here from public comment screenshots [unverified].", "The same population is loyalty-policed from both directions — suspected of pro-China loyalties in the US, condemned as turncoats in Chinese discourse — and the 1942 Japanese American incarceration (about 120,000 held, over half US citizens) is the precedent both sides of this exchange invoked." ], "url": "https://sigpulse.com/posts/2026-08-29-timcast-internment-chinese-americans/", "md_url": "https://sigpulse.com/posts/2026-08-29-timcast-internment-chinese-americans.md" }, { "slug": "2026-08-27-one-man-legion-ep1-image-workshop", "title": "How Fast Does a Chat Command Become a Finished AI Image? Inside the One-Man FLUX Workshop", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "FLUX", "ComfyUI", "PuLID", "Text-to-Image", "AI Agents" ], "series": "One Man One Legion", "series_order": 1, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "A one-line chat command becomes a finished, delivered image in 50 s on the fast profile (1280×720, 20 steps, n=3 identical runs) and 66–112 s warm on the full profile (1920×1080, 28 steps, n=5) on a single RTX A4000 16 GB — with the script, not the agent, posting the result back to the chat group (job logs, 2026-08).", "The dominant latency is not rendering but the first model load: the two runs that open a session took 330 s and 379 s, after which the same profile settles to 66 s — budget ~5–6 min for a cold workshop, tens of seconds for everything after (10-run log corpus, 2026-08).", "The workshop ships two prompt profiles with tuned constants (fast: guidance 3.5, PuLID 0.65; full: guidance 5.5, PuLID 0.8, CFG pinned to 1.0) plus an i2v mode that rewrites dynamic words to static ones ('running fast' → 'standing alert, coiled energy') so images can serve as video base plates — craft that only shows up in logs, documented here from the factory script." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep1-image-workshop/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep1-image-workshop.md" }, { "slug": "2026-08-27-one-man-legion-ep10-pen-names", "title": "Whose Name Goes on the Article? Five Bylines, One Human", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Content Pipeline", "WeChat", "Writing" ], "series": "One Man One Legion", "series_order": 10, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The main account signs with five genre bylines, not four: 33 archived files carry byline lines — Lu Shi 12, Shen Jianwei 9, Lingche 8, Lin Shu 2, Qin Yin 2 (recounted 2026-08-27) — and the shared AI-era survival guide column carries all five names across its 10 pieces, because a byline here follows genre, never column.", "Each name is a bundle of prohibitions: the E-type pathologist Lingche runs a strict 4-step structure with a terminology-rotation ban and a 9-word banned list that outlaws emotional vocabulary itself; A-D types write in a friend's voice with an emotion micro-pulse every 300-500 characters; typing the emotion-manipulation trigger word switches the factory into a separate writing mode whose doc still keeps three zero-tolerance lines.", "A sixth name, Zhixing Xiaoya, exists as a second account's config plus a stock-commentary style doc effective 2026-07-19 — with zero bylines anywhere in the workspace; of 11 surviving push scripts carrying an AUTHOR constant, person-story writer Lin Shu's genre has none." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep10-pen-names/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep10-pen-names.md" }, { "slug": "2026-08-27-one-man-legion-ep11-quality-gate", "title": "Who Checks the Checkers? 313 Lines of Gate, 61 Stamps, One Empty Column", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Quality Control", "Content Pipeline", "Automation" ], "series": "One Man One Legion", "series_order": 11, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The QC gate is 313 lines split by what can be counted and what must be felt: 7 regex scans and 4 model judgments. The 102-record failure log rules on the split — regex caught 61 legal-safety, 54 title, 34 image, 31 source and 22 banned-word red lights, while two rules (length, quotable density) fired 0 and 2 times, checks the technology outgrew.", "61 passing stamps survive — 46 in articles/ and 15 in daily/ — with 42 first-try passes, 8 on round two, 8 on round three, and 3 needing a fourth; the 13 per-article push scripts in the daily pipeline read the stamp and exit without it, but the older root push scripts never ask.", "The gate's learning loop has one wheel: a column for AI-judgment issues exists in the logger and was populated 0 times across all 102 records — the ratchet that promotes repeated errors into hard rules only ever ratchets on regex failures." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep11-quality-gate/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep11-quality-gate.md" }, { "slug": "2026-08-27-one-man-legion-ep12-analyst-desk", "title": "The Analyst Desk: 85 Points of Fact, ±13 of Feeling, One Dead Cron", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Multi-Agent Systems", "Stock Analysis", "Automation" ], "series": "One Man One Legion", "series_order": 12, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The desk's score engine split a stock rating into 85 points of fact — financials ×35%, industry ×25%, quant ×25%, each with its own factor tree — plus a correction envelope of ±13 (peers ±5, news ±5, community ±3). The stress test T9 confirmed the weights sum to 0.85, leaving 15% for corrections; T11 pinned the envelope at +13/-13 (85→98 clamped to 100, or down to 72).", "Feeling was caged by consensus thresholds: a Reddit post counts as high-signal only at ≥100 upvotes, ±3 score movement requires multiple subreddits plus HN agreeing, and a single-source viral post is annotated but never moves the score. Asymmetry was deliberate — extreme-negative news hit the -5 floor in testing while extreme-positive only earned +3, with the harness itself flagging 'could be more aggressive'.", "Every number that reached readers passed a verification gate the SOP marks as unskippable: it evolved from dual-round re-pulls (29/29 data points, 0 errors on June 17) to claimed-vs-actual ledgers to dual-source cross-validation against yfinance × Sina quotes for 7 tickers. The desk is now dark — its cron last ran July 23 and exited with an error — leaving 4 analysis posts, 6 push records, and a filings library of 49 files under 13 tickers against a 14-company config." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep12-analyst-desk/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep12-analyst-desk.md" }, { "slug": "2026-08-27-one-man-legion-ep13-audio-workshop", "title": "The Audio Workshop: Six Voices on a Shelf, One Episode That Never Shipped", "description": "E13: the audio workshop — six reference voices, a 172-line IndexTTS-2 pipeline, a lost episode, and three podcast crons gone dark.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Text-to-Speech", "Podcasting", "Automation" ], "series": "One Man One Legion", "series_order": 13, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The workshop's raw material is a shelf of six reference voices plus one script: four designed composite hosts (an AI female, an AI male, an English female, an English male) and two clips lifted from real speakers. The daily show ran on one of the cloned voices — and the house rule for clones is absolute: use the voice, never the name.", "The whole studio is podcast_tts_indextts.py, 172 lines: it accepts edge-tts-format scripts with rate and pitch fields it deliberately ignores for compatibility, monkey-patches torchaudio.save with soundfile because the upstream path broke, synthesizes segment by segment, and emits a 192 kbps MP3 plus a per-segment timeline JSON that later became the subtitle spine of the video factory. Its voice engine carries its own small language model — the 8.3 GB of IndexTTS-2 checkpoints include a Qwen 0.6B.", "Five episodes were produced between June 28 and July 26; the four with surviving timelines run 36/16/27/23 segments and 201/393/351/385 seconds. The July 4 episode — 5 minutes 50 seconds, 2 MB — never shipped: the proxy tunnel to Telegram failed with an SSL error while domestic sites stayed reachable, and the file was later deleted, leaving only a pending-delivery note. All three podcast cron jobs are now disabled in the contraction." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep13-audio-workshop/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep13-audio-workshop.md" }, { "slug": "2026-08-27-one-man-legion-ep14-model-stable", "title": "The Model Stable: Two Days in June, One Evening in August", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "vLLM", "Local LLM", "Open Models" ], "series": "One Man One Legion", "series_order": 14, "arc": "The Engine Room", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The stable holds three locally owned models — a 23 GB coder, an 18 GB reasoner, a 29 GB workhorse, 70 GB in total — and each earned its stall for a different job: editor-side code commentary, disaster fallback, and the coding mount that carries Claude Code on local weights.", "The June deployment of the 9B fallback took two days and seven pitfalls; the August arrival of the 27B went from first download byte to a mounted coding agent in five hours and twenty-one minutes — because June's scars (mirror downloads, same-family parsers, the FlashInfer kill switch) had compiled into August's defaults.", "The stable's core discipline is knowing boundaries: the 12B was tested and assigned commentator-not-detective; the 9B was triaged across ten chat rooms as 4 fully capable, 3 degraded-acceptable, 3 forbidden — because a small model faking deep analysis is worse than silence. Today the named fallback endpoint is dark while the 27B holds the GPU: the stable is a promise kept by scripts, not a hot standby." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep14-model-stable/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep14-model-stable.md" }, { "slug": "2026-08-27-one-man-legion-ep16-mini-agent", "title": "Model or Harness? Two Controlled Experiments on a 363-Line Hand-Rolled Coding Agent", "description": "E16: swap the brain — 15 rounds stuck vs 4 to pass; swap the prompt — weak models stay unsaveable. Three code supervisors make agents honest.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "LLM", "vLLM", "Agent Harness", "Controlled Experiments" ], "series": "One Man One Legion", "series_order": 16, "arc": "The Engine Room", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "With the harness held constant and only the model swapped, a local 12B coder stalled 15 rounds on a one-line fix (editing a hallucinated function, pytest still 1 failed) while cloud GLM-4.6 passed in 4 rounds with an exact edit (3 passed) — infrastructure can absorb parser and tool flakiness, but not semantic drift (run traces, 2026-08).", "With the model held constant and only the system prompt swapped, the strong model finished either way (6 vs 7 rounds, 4 passed both) but only the structured prompt produced scan-first behavior and a what-changed/risks/how-verified report; the 12B failed under all three prompt variants — weak models cannot be prompt-saved (same traces).", "Three code-level supervisors — edit's old_string must appear in the last read (A1), tool errors force a re-read (A2), finish is rejected unless a file was actually edited (A3) — made the agent stop reporting fake success; honesty proved achievable in code even where capability was not (mini_agent.py, 363 lines)." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep16-mini-agent/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep16-mini-agent.md" }, { "slug": "2026-08-27-one-man-legion-ep17-model-exam", "title": "The Model Exam: Seven Passes, Two Zeroes, One Home", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Local LLM", "vLLM", "Model Evaluation" ], "series": "One Man One Legion", "series_order": 17, "arc": "The Engine Room", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "Graded by execution, not vibes: seven coding tasks of rising difficulty all passed against assertion scripts — 7/7 at 13-142 seconds and 194-1,834 output tokens — with every caveat (hand-extracted code, verifier strictness, thinking leak) logged next to the passes.", "The A/B found what splitting actually buys: not quality (the one-shot passed all five verifier cases cleanly; the split path passed four and leaked raw rows on the fifth) but containment — one-function diffs instead of whole-file rewrites. Meanwhile max_tokens starvation killed two of six requests with zero characters of output, and the split mode's one bug was authored by the prompt itself.", "Four habitats, one winner: the 24,769-token Claude Code ticket eats 64% of the 38,912-token window, the terminal harness charges 7,710 for 3.8× the working room, and six global rules cost 241 tokens while consuming less (49,394 vs 64,907) — so the harness became the model's home, the bare endpoint serves pipelines, and the mount wrapper stays a spare." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep17-model-exam/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep17-model-exam.md" }, { "slug": "2026-08-27-one-man-legion-ep18-brain-swap", "title": "The Brain Swap: Ten Backups, One Insurance Card, No Standby Factory", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Model Migration", "Config Management", "LLM Ops" ], "series": "One Man One Legion", "series_order": 18, "arc": "The Engine Room", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The backup strata read like tree rings: ten openclaw.json files, the oldest of them the platform's birth certificate (2026-03-29 10:53, 2,459 bytes against today's 7,540) — and the first brain transplant landed 118 minutes after birth, glm-5 to glm-5.1, staged in the catalog before the flip.", "One link in the chain has no named grave: 139 days separate the before-glm51 and before-glm53 backups, and somewhere in that silence glm-5.2 became primary without ceremony — the honest gap that argues for the protocol that came later.", "The 5.3 swap of August 15 is the textbook run: twenty articles shipped on the day shift, then a two-stage evening surgery (Claude Code 20:52, OpenClaw itself 21:08) under a four-item protocol — dual-route precheck, named backup, the rollback command handed to the human before the self-restart, fallback chain frozen — and the factory shipped one article the next day and six the day after." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep18-brain-swap/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep18-brain-swap.md" }, { "slug": "2026-08-27-one-man-legion-ep19-cron-contraction", "title": "What Should an AI Legion Do While You Sleep? 22 Cron Entries, Four Still Lit", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Cron", "Automation", "Scheduling", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 19, "arc": "War Stories", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "A one-person legion's cron schedule grew for 116 days to 16 jobs lit at once — 790 finished runs, 154 errors — and on 2026-07-23 the operator switched the auto-publishers off; the file now holds 22 entries with 4 enabled (jobs.json + backups + run logs, 2026-08).", "The cut has a clean criterion: every surviving job feeds material to the human or asks a question — the 03:30 YouTube material drop, the 03:20/17:40 video-topic asks, and an arXiv digest added Aug 20 — while every job that decided and published alone went dark, its work moved to a human-gated pipeline (jobs.json, 2026-08).", "Contraction did not buy reliability: the arXiv survivor logged 18 consecutive errors in the Aug 21–27 storm — the log's own words, 'All models failed (2)', cloud GLM-5.3 and the local fallback both timing out — before recovering on Aug 27 with 82 papers (run logs, 2026-08)." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep19-cron-contraction/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep19-cron-contraction.md" }, { "slug": "2026-08-27-one-man-legion-ep2-tg-gateway", "title": "What Does It Take to Give a Chat AI Nine Tools and Your Home Directory? A 50-Line Bot That Grew to 850", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Telegram Bot", "GLM", "Agent Tools", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 2, "arc": "The Cockpit", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The chat gateway that hands an AI real tools is 850 lines of Python today and started near 50 — the 800-line delta is almost entirely security and robustness added after real failures: a jailed work directory with prefix-spoof protection, a dangerous-command blocklist (rm -rf, sudo, mkfs, dd of=), exact-match edits, and path-fallback reads (source + README, 2026-08).", "The bot exposes nine tools to GLM-5.2 — read_file, glob, grep, edit, write_file, exec_command, web_search, web_fetch, rss_fetch — inside a 15-iteration tool loop, so a single chat message can chain a full investigate-edit-test sequence (source, 2026-08).", "Usage scale: one whitelisted human, 168,996 log lines — and the log's tail is a live server-disconnect NetworkError, a fair sample of what 'reliable' looks like at personal scale: crash-restart hygiene matters more than uptime claims (bot.log, 2026-08)." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep2-tg-gateway/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep2-tg-gateway.md" }, { "slug": "2026-08-27-one-man-legion-ep20-bug-diaries", "title": "Bug Diaries: 9 Failure Files, 158 Days, and the Root Cause It Got Wrong", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Postmortem", "LLM Reliability", "OpenClaw", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 20, "arc": "War Stories", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The legion keeps 9 dedicated failure documents spanning 158 days (Mar 14 12:25 to Aug 19 11:48) — born as a cluster of 5 files in under 33 hours on the Mar 14-15 weekend, beside the per-article loop episode 9 measured at 102 lines.", "Writing failures down did not stop them: the image failure was lesson-ized on Mar 29 and still recurs 4 documented times through Jul 12 (77 days to the first relapse); the fabricated match report of Jun 28 broke three rules already on the books.", "The archive fails the way the systems it audits do: error No. 24 is used twice, No. 11 exists only as a changelog row, the weekly review rota was never edited again after Mar 15 21:08 — and its final entry (Aug 19 11:47) reverses its own root cause after the human's correction." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep20-bug-diaries/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep20-bug-diaries.md" }, { "slug": "2026-08-27-one-man-legion-ep21-langgraph-autopsy", "title": "The LangGraph Autopsy: 12 Green Imports, 668 Silent Seconds, One 203-Line Survivor", "description": "E21: 12 green imports, 668 silent seconds — the autopsy that deleted a framework 90 minutes in, and the 203-line engine that outlived it.", "category": "AI & Compute", "tags": [ "One Man One Legion", "LangGraph", "Architecture", "Postmortem", "Workflow Orchestration", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 21, "arc": "War Stories", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The LangGraph layer lived 90 minutes as load-bearing architecture: skeleton logged complete at 12:35 on Aug 12, 2026 (12 modules import-green, 9 nodes, conditional routing, a sqlite checkpointer), deletion logged at 14:05 — after the first true run stalled 668 seconds at 13:33 with the render already finished on the GPU box.", "The minimal repro proved the design physically impossible rather than buggy: END is a thread's terminal state and invoking None against a finished thread is a no-op, so two runs of a toy graph produced the identical call list ['entry','worker'] — poll_comfy had never executed after submit. Import smoke tests syntax, not semantics.", "The 203-line hand-rolled engine that replaced it passed the same flow in 163 seconds at 13:59 — the audit's first-ever poll_done line is the receipt — and the next morning a human's Telegram Hi got its image back in 27 seconds; fault injection that afternoon found 4 bugs of one root: zero tolerance plus false done." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep21-langgraph-autopsy/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep21-langgraph-autopsy.md" }, { "slug": "2026-08-27-one-man-legion-ep23-security-handcraft", "title": "The Security Handcraft: 12 Fossil Scripts, One Live Key, Zero Leaks in Print", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "Security", "Secrets Management", "Infrastructure Audit", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 23, "arc": "War Stories", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The rule as built: one credentials home per machine — the workstation config holds 138 fields of which 6 are credentials, file mode 600 — read at runtime by code like the image factory, with a logging filter that swaps token substrings for asterisks before any line is written. New code never inlines.", "The rule as inherited: the audit found 12 legacy push scripts each carrying the same 46-character token inline, and a live probe returned ok=true for both that fossil string and the config's string — 2 valid credentials for one bot, only one of which the config knows about.", "The publication gate holds: 0 secret-pattern hits across all 23 published posts, 5 ritual-scan receipts in the work log, and the two in-session forensic leaks this series caused (a config dump, a parent-object print) were both caught before print." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep23-security-handcraft/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep23-security-handcraft.md" }, { "slug": "2026-08-27-one-man-legion-ep24-token-ledger", "title": "The Token Ledger: 261,720,417 Tokens, 147 Re-reads per Token Written", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "Token Economics", "AI Agents", "LLM Usage", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 24, "arc": "The Ledger", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The bill is memory, not generation: of 261,720,417 metered tokens across 208 days, 84.5% are cache re-reads and 0.68% output — a 147:1 read-to-write ratio, roughly 66,580 tokens of context per 450-token reply.", "Consumption concentrated, then disciplined: February and March alone burned 71.6% of the ledger at over three million tokens a day; the contracted fleet now runs near 1.2 million, one February session alone accounts for 12.2% of everything, and the only currency figure the records carry — 52.27 in an unrecorded unit — is also from that founding era.", "Three record systems agree on the 75-day silence — the ledger's last March entry lands two minutes after the glm-5.1 brain swap, its restart matches the diary's first entry back, and the local fallback fired only 17 times in seven months." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep24-token-ledger/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep24-token-ledger.md" }, { "slug": "2026-08-27-one-man-legion-ep25-human-ledger", "title": "The Human Ledger: 26 Drops, 191 Pieces Behind One Phone, Five One-Line Rulings", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Human-in-the-Loop", "Workflow", "Personal Infrastructure", "Automation" ], "series": "One Man One Legion", "series_order": 25, "arc": "The Ledger", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The machines stamp themselves: 61 quality-gate records on the veteran WeChat line, each holding 7 machine checks with a retry-round histogram of 42/8/8/3 — and not one field for a human, because on that line the human's only logged act is the tap that publishes.", "The human's receipts live in the fleet's own systems: 26 dropped entries in the chat inbound folder across a 171-day window (15 jpg, 6 png, 2 voice clips, 2 PDFs, one bundle), 3 gate interactions before standing authorization graduated the series, and 5 one-line rulings dated August 11/13/15/19/25 — every one written down by a machine.", "Summed against Episode 24's ledger — 261,720,417 tokens across 3,931 turns in the same 208 days — roughly 225 logged human acts come to over 1.1 million tokens per act, about 17 machine turns per act, about one logged act per day." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep25-human-ledger/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep25-human-ledger.md" }, { "slug": "2026-08-27-one-man-legion-ep26-closing-the-loop", "title": "The Loop Closes: What One Man and One Legion Answered to a Turbulent Era", "description": "E26, the finale: 24 posts, 121 ledger entries, 261,720,417 tokens against roughly 225 human acts — the series' measured reply to Gates' warning.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Personal AI Infrastructure", "Measurement", "Series Finale" ], "series": "One Man One Legion", "series_order": 26, "arc": "The Ledger", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The series ends at 24 posts — one anchor and 23 episodes across 5 arcs (The Workshops 7, War Stories 5, The Engine Room 4, The Cockpit 4, The Ledger 3) — with 121 measured entries in the public ledger; three episode numbers (8, 15, 22) were never issued, so the 26th episode by numbering is the 23rd by count.", "The final balance, assembled from published episodes: 261,720,417 tokens burned over 208 days against roughly 225 logged human acts — 1,163,202 tokens per act to the nearest act, about one human act per day — and 191 finished pieces still shipped through one phone.", "Gates' three proposals each meet a measurement here: before you can tax AI tokens you must be able to count them (this fleet's own cost meter went dark in June), Human Reserved already exists at n=1 as three gestures and five rulings, and the preparation he says the world lacks looked, in practice, like bookkeeping — diaries, gate stamps, and the willingness to switch your own cron jobs off." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep26-closing-the-loop/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep26-closing-the-loop.md" }, { "slug": "2026-08-27-one-man-legion-ep3-chat-cockpit", "title": "What Happens When the Cockpit Itself Goes Dark? Two Outages, Ten Stranded Jobs, One Spare Parked by Subtraction", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Telegram", "Feishu", "Reliability", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 3, "arc": "The Cockpit", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The June 20 outage was one dead domain: the proxy exit node every command traveled through returned NXDOMAIN from 4 independent DNS resolvers, and the gateway logged a polling stall every 90 seconds — the fix was not a better node but a pool of 6 (0.44–4.14 s), health-probed every minute behind a least-ping balancer (telegram-proxy-fix-2026-06-20.md).", "On July 4 the pipe died upstream and 13 outbound jobs were stranded — 10 OpenClaw cron jobs and 3 system crontab scripts — with the work already done and products on disk; a four-row contrast experiment isolated the Feishu bug: agentTurn news jobs ran 149 and 112 seconds while the identical-envelope systemEvent job idled 10 seconds, its Telegram twin running 270 (feishu-migration-diagnosis-2026-07-04.md).", "The repaired noon job fired on schedule after the fix — 11 run records through July 11 — and the Feishu news pair kept twice-daily schedules to the last fires on July 22 and 23, when the great contraction switched the auto-publishers off; today 0 of the 4 Feishu-routed entries are lit and 3 of the 4 lit jobs route to Telegram: the channel did not fail, the schedule was cut (run logs + jobs.json, 2026-08-27)." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep3-chat-cockpit/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep3-chat-cockpit.md" }, { "slug": "2026-08-27-one-man-legion-ep4-three-touchpoints", "title": "What Does the One Human Actually Do? Three Touchpoints per Piece, and One Wall That Won't Move", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Human-in-the-Loop", "Workflow", "WeChat", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 4, "arc": "The Cockpit", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The division-of-labor contract, fixed by a fictional-article drill on 2026-08-25, puts the operator in exactly 3 places per piece — drop material, fill a Chinese number-crosscheck slip budgeted at 5 minutes, click publish — with one clause pointing the other way: the operator never touches git, build, or deploy.", "The first real slip, 2026-08-26: 27 items split 21 machine-verified plus 6 awaiting a terminal grep the operator ran personally on the workstation, confirming the half-resident log line 2 times; 13 A-items and 6 B-items passed, 4 C-items of pure fact arbitration were left to the human — and the article sat built and green, unpushed, until the slip came back.", "The veteran WeChat factory enforces the same boundary by platform wall, not discipline: 0 freepublish call sites in checked-in code against 1 draft-box call site, the one live attempt answered 48001 api unauthorized — so publishing happens from the phone; meanwhile the series queue ships on standing authorization, and this episode crossed no slip." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep4-three-touchpoints/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep4-three-touchpoints.md" }, { "slug": "2026-08-27-one-man-legion-ep5-fleet-on-one-screen", "title": "Six Doors, 231 Lines, and No Dashboard: Putting a Three-Machine Fleet on One Screen", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Self-Hosting", "Observability", "Tailscale", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 5, "arc": "The Cockpit", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "The three machines became a fleet in one day of ssh work: 6 directed doors where machine A runs headless Claude on machine B — stateless phone calls, credentials never crossing the wire — all 6 wrapper files still on disk today, 4 of them stat'd at 334 to 912 bytes.", "A same-day stress test priced the mesh: 0 packet loss and 140/89 ms links, but a Claude round trip of 24 s to the 512 MB box vs 8–12 s to the workstation — the 2.4× gap is RAM, not wire, so the fleet's one physical law routes heavy reasoning away from the small machine.", "The cockpit is a 231-line probe checking 12 keys every 5 minutes and logging the whole fleet in 3 lines per round; 16 days of alerts hold 20 red, 8 yellow, 25 green, exactly 1 real link outage — and its longest streak, 4,711 checks, is arithmetic that lands within 5 minutes of the probe's own v2 reset." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep5-fleet-on-one-screen/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep5-fleet-on-one-screen.md" }, { "slug": "2026-08-27-one-man-legion-ep6-agent-diary", "title": "What Does an AI Legion Write to Itself? 63 Entries, 75 Days of Silence", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Memory", "LLM Context", "OpenClaw", "Personal Infrastructure" ], "series": "One Man One Legion", "series_order": 6, "arc": "War Stories", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "A 292-line constitution (AGENTS.md, last edited 2026-08-17) makes the memory involuntary: every session opens by reading SOUL.md, USER.md, the day's and yesterday's diary, and MEMORY.md — 'Don't ask permission. Just do it.' (AGENTS.md, read 2026-08).", "The diary is real but smaller than this series claimed: 63 daily entries across a 186-day span (225,236 characters), plus 34 topical notes — the anchor's '97 daily agent-diary files' counted every file in the directory; the correction stands here, not in a silent edit.", "The longest silence is dated: the last pre-gap entry (Mar 29) records the GLM-5.1 swap and shares its date with the first cron fire (21:27, per episode 19); 75 days later the diary resumes with hands-on work. The curated MEMORY.md (160,639 bytes) was still being maintained Aug 21." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep6-agent-diary/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep6-agent-diary.md" }, { "slug": "2026-08-27-one-man-legion-ep7-video-workshop", "title": "One Sentence In, Stereo Video Out: 6 Jobs, 90–790 Seconds, an Unfilmed Actress", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Text to Video", "ComfyUI", "Local AI", "Automation" ], "series": "One Man One Legion", "series_order": 7, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "A locally hosted MiniMax H3 text-to-video bench ran 6 logged jobs between Aug 6 and Aug 14, 2026 — 11 MP4s on disk including 5 deployment-day tests — with render times of 90s (3s clip), 138s (5s, three identical runs), and 790s (10s): doubling the clip length cost 5.7x the render on an offload-bound 4090.", "The pipeline delivers itself: launched in background mode, the script posts the finished video back to the chat group (5 of 6 jobs did) while the agent replies with an estimate and exits — an agent that waits synchronously on a 90-second-to-13-minute job times out.", "A director skill (v1.3) caps every prompt with a motion budget — 1 primary motion, 2 secondary, 2 environmental, 1 camera move, 1 expression — validated by an 8-cell reference-to-video test matrix whose findings (swap the seed, not the prompt; write the light source in) are codified as rules; the automated result tracker, built to log every run, has zero entries." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep7-video-workshop/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep7-video-workshop.md" }, { "slug": "2026-08-27-one-man-legion-ep9-text-factory", "title": "Who Edits an AI Writing Staff? 191 Articles, 15 Iron Laws, One Stamp", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Content Pipeline", "Quality Control", "WeChat", "Automation" ], "series": "One Man One Legion", "series_order": 9, "arc": "The Workshops", "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "A one-person Chinese publishing operation archived 191 finished HTML articles — 72 in articles/, 105 in daily/, 14 in archive/ — between June 15 and Aug 27, 2026, produced by a staff of 17 agent persona files and signed by four emotion-typed pen names, whose bylines appear in 30 archived files (archive recounted 2026-08-27).", "Quality is codified as 15 iron laws enforced as 7 machine hard scans plus 4 AI judgments; a passing gate writes a .gate.json stamp and the push script refuses to publish without it, delivering to the platform's official draft-box API (workflow standard + gate stamps, 2026-08).", "Every caught flaw lands in quality-lessons.jsonl — 102 records since July 18 (88 first-round, 11 second-round, 3 third-round failures), the newest written Aug 27 17:46:40 — and any error repeated three times is promoted into a permanent hard-scan rule; today a missing-counter-view failure was cured and re-stamped 21 seconds later." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep9-text-factory/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-legion-ep9-text-factory.md" }, { "slug": "2026-08-27-one-man-one-legion-fleet-audit", "title": "One Man, One Legion: Can a Single Operator Run a 3-Machine AI Fleet End to End?", "description": "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.", "category": "AI & Compute", "tags": [ "One Man One Legion", "AI Agents", "Local LLMs", "vLLM", "Cron Automation", "Personal AI Infrastructure" ], "series": "One Man One Legion", "series_order": 0, "date": "2026-08-27", "measured_on": "2026-08-27", "verified_hardware": "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", "key_takeaways": [ "A solo operator runs a 3-machine AI content fleet — a 512 MB US-West VPS scouting overseas RSS, a 2-vCPU Seoul cloud VM handling orchestration and this website, and a local RTX 4090D 24 GB + A4000 16 GB workstation — with exactly 3 human touchpoints per piece: drop source material, fill a one-page gate slip, click publish (fleet audited by live inspection, 2026-08-27).", "The workstation keeps 3 local LLMs runnable — a 29 GB 27B-AWQ and a 23 GB 12B coder in two simultaneous vLLM instances, plus an 18 GB 9B on standby — and the agent runtime's model chain is deliberately short: one cloud tier with one local 9B fallback, so a cloud outage degrades the legion to a local endpoint instead of stopping it (config audited 2026-08-27).", "Over-automation was walked back on evidence, not faith: of 22 cron jobs ever installed across the fleet only 4 remain enabled at audit, while 97 agent-diary files and a per-failure quality-lessons log accumulate what actually broke — the fleet's center of gravity moved from scheduling to learning loops (audited 2026-08-27)." ], "url": "https://sigpulse.com/posts/2026-08-27-one-man-one-legion-fleet-audit/", "md_url": "https://sigpulse.com/posts/2026-08-27-one-man-one-legion-fleet-audit.md" }, { "slug": "2026-08-26-agent-tool-interface-ard-mcp-measured", "title": "Can a Static Blog Hand AI Agents Real Tools? Wiring ARD + MCP into an Astro Site (Measured)", "description": "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.", "category": "AI & Compute", "tags": [ "ARD", "MCP", "OpenAPI", "Astro", "Vercel", "Agents" ], "date": "2026-08-26", "measured_on": "2026-08-26", "verified_hardware": "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", "key_takeaways": [ "The full agent interface shipped as one 13-file, +950-line commit (8eeb3ec, 2026-08-26): an ARD catalog at /.well-known/ai-catalog.json validated against the official JSON Schema, a 7-tool OpenAPI 3.1 document, two JSON indexes, and a 359-line read-only MCP server — and Vercel served it live 12–18 s after push (6 s polling granularity).", "A stateless MCP server over static data is enough: /api/mcp answers initialize, tools/list and tools/call by fetching the site's own static JSON and filtering in memory — first call in a 5-sample run took 0.90 s, the four follow-ups 0.28–0.34 s (AWS Seoul, 2026-08-26), with zero database and zero auth.", "Live testing keeps catching what local checks miss: phrase-only search returned 0 hits for the natural query \"DeepSeek price\" (fixed, +21 lines), and a same-day 11-hypothesis adversarial stress test with a real MCP client found five more declared-vs-actual gaps — limit clamping, query validation, undeclared OpenAPI security — fixed in eae3a23 (+31/−4) and re-verified on production." ], "url": "https://sigpulse.com/posts/2026-08-26-agent-tool-interface-ard-mcp-measured/", "md_url": "https://sigpulse.com/posts/2026-08-26-agent-tool-interface-ard-mcp-measured.md" }, { "slug": "2026-08-26-infinitetalk-torch-load-dependency-matrix", "title": "InfiniteTalk Dies at torch.load: the Error Tells You to Upgrade torch — the Measured Fix Is Pinning transformers 4.52.0", "description": "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.", "category": "AI & Compute", "tags": [ "InfiniteTalk", "Wan2.1", "PyTorch", "transformers", "CVE-2025-32434", "Dependencies" ], "date": "2026-08-26", "measured_on": "2026-02-19", "verified_hardware": "RTX 4090D 24GB + RTX A4000 16GB dual-GPU workstation (Ubuntu 24.04, conda, Python 3.10)", "key_takeaways": [ "On a stack that had already run InfiniteTalk 14B fp8 end-to-end (torch 2.4.1), transformers 4.57.3 kills every launch at torch.load with a ValueError citing CVE-2025-32434 — failing in 18.0–18.6 s, three attempts in a row, before a single frame renders.", "Following the error message and upgrading to torch 2.10.0+cu128 breaks the rest of the stack simultaneously: a three-package pip conflict (torchaudio, torchvision, xformers all require torch==2.4.1), `operator torchvision::nms does not exist`, a diffusers import failure (`JITCallable._set_src()`), and flash-attn's `undefined symbol` — five verbatim breakages, measured the same morning.", "The measured fix goes the opposite direction of the error message: transformers pinned to 4.52.0 (4.49.0 and 4.51.0 import cleanly but die inside xfuser's diffusers imports), on torch 2.4.1 + torchvision 0.19.1 + xformers 0.0.28 + flash-attn 2.8.3 — the only combination in the sweep that reached pipeline initialization." ], "url": "https://sigpulse.com/posts/2026-08-26-infinitetalk-torch-load-dependency-matrix/", "md_url": "https://sigpulse.com/posts/2026-08-26-infinitetalk-torch-load-dependency-matrix.md" }, { "slug": "2026-08-26-infinitetalk-14b-dual-gpu-measured", "title": "Can You Run InfiniteTalk on Two Consumer GPUs? Yes — at 218.5 s/step (RTX 4090D + RTX A4000, Measured)", "description": "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.", "category": "AI & Compute", "tags": [ "InfiniteTalk", "Wan2.1", "RTX 4090D", "RTX A4000", "fp8", "Talking-Head Video" ], "date": "2026-08-26", "measured_on": "2025-11-29", "verified_hardware": "RTX 4090D 24GB + RTX A4000 16GB dual-GPU workstation (Ubuntu 24.04, 503GB RAM)", "key_takeaways": [ "InfiniteTalk 14B fp8 (infinitetalk_single_fp8.safetensors) runs on an RTX 4090D 24GB + RTX A4000 16GB rig with zero CUDA out-of-memory events, via a 20-of-40-block semi-resident split — but it crawls at 218.5 s/denoising-step, which is ~50–55 minutes per 81-frame 480P clip (≈5 s of video at the Wan2.1 480P 16 fps training rate, roughly 647× slower than realtime).", "The block-residency dial buys fitting, not speed: raising resident blocks from 20 to 24 costs ~1.7GB more VRAM (10.21→11.99GB on the A4000, 22.1→23.2GB on the 4090D) and returns +0.64% (218.5→217.1 s/step). Returns had flatlined.", "The real costs live outside the GPUs: a ~104GB minimum disk footprint (19,499,692,400-byte fp8 checkpoint + 77GB Wan2.1 base + T5 + wav2vec2), 241.8GB measured with quantization variants on disk — and a dependency matrix so brittle that a 2026 reinstall attempt took six tries (part 2)." ], "url": "https://sigpulse.com/posts/2026-08-26-infinitetalk-14b-dual-gpu-measured/", "md_url": "https://sigpulse.com/posts/2026-08-26-infinitetalk-14b-dual-gpu-measured.md" } ] }