A thermal camera sits behind an ET-SoC-1 accelerator card. Move 128 of its 1,024 hardware threads from one corner of the die to another every 25 seconds, and the warm patch on the board walks with them.
Thermal Master P3 behind the card; two sessions, 390 heater launches, 71.1 minutes of card time, none failed. Every movie is a difference image, never a raw frame: the swaps against their opposite half, the orbit and the sweep against their lap’s own average, the bloom against the idle before each load step. The panel burnt into each frame says harts, the chip’s own word for hardware threads.
128 threads in each of four corner blocks of the die, then the centre; 25 s a position, five laps at 10× real time. The die map in each frame is turned to match the camera; colours are the difference from that lap’s average, red warmer. The corners saturate the shared ±400 mK scale and the centre, a fifth their size, reads nearly blank.
Move the work round the four corners and the heat walks with it
The camera sees the die rolled about a right angle, not mirrored: west→east runs up the picture, north→south runs right. So the corners read NW lower left, NE upper left, SE upper right, SW lower right — and the warm patch walks them clockwise, in all five laps, before collapsing into the middle.
+525 ± 22 mK
the board’s warm half changes ends when the work does
west↔east swap, 192 threads a side, constant total power, 4 of 4 laps (post hoc)
p = 0.0005
the four-corner walk against its own permutation null
+846 mK of excess contrast, 5.6 sd; 2,000 label shuffles, and not at the test’s floor
94° apart
the same 24 shires split the other way turn the picture through a right angle
first moment; 91.6–98.7° across three estimators and every drop-one-lap combination; the registered band was 90 ± 10°
+4 ± 11 mK
the same pipeline with nothing swapped
identical mask in both arms, 4 laps (post hoc) — under 1% of the real swap
Four corners, four quadrants — and a centre near zero that its own power rule makes exploratory
Where each block’s heat lands — on the die’s own compass, not the camera’s view above, which is rolled a right angle. Ticks every 100 mK.
NW206 west · 323 north
NE176 east · 173 north
SE165 east · 229 south
SW202 west · 281 south
centre66 east · 14 north
Mean ± standard error over five laps, in mK, post hoc.
Positions are read along templates built from the other session’s blocks, and two independent fits of the die’s direction agree to within 2.5°. Post hoc, like the shutter cleaning behind them.
Two clauses of the hero’s own pass rule failed, and that belongs here. The five positions’ board power spread 0.620 W where the rule allowed 0.3, and the centre drew the most — so the centre reading is exploratory; its electrical correction is 23–48 mK against centre-to-corner dipoles of 217–381 mK. And each phase’s frame average reached 86 mK where the rule allowed 25 — but it flips sign when the shutter cleaning changes while the moving lobe does not, so it is the camera’s shutter, not the card’s regulators.
Split the same 24 shires the other way and the picture turns a right angle
Twelve shires west, then twelve east: 192 threads either way, 42.5 s a side. Heat arrives about 11 s late, so just after each switch the picture still shows the half that has gone quiet. Movies play as they scroll into view; tap one to pause or play it.The same 24 shires split north against south instead. The speckle while the north twelve run is shutter residue the cleaning missed, not heat: the board blurs heat over centimetres.
Same sign in all ten swap laps; both controls sit on zero — all five contrasts post hoc, read along the other session’s templates
Contrast along the block’s own axis, in mK — one dot per analysed lap.
Moving 192 threads from the die’s west twelve shires to its east twelve at constant total power moves the warm half by +525 ± 22 mK (post hoc) — same sign in every lap, in both swap directions. Split the same shires north against south and the picture turns a right angle, which two fixed power stages drawing the same 43 W cannot do.
The same pipeline with nothing to find, and it found nothing
A permutation p is no use here: four laps give sixteen ways to flip the labels, so the registered within-lap null cannot return below 0.125 however large the signal. The controls answer instead by size.
Control: the identical mask in both arms — the compute never changes.Control: twelve shires against their own 180° rotation, same power, same centre — measured in a quiet four-minute window, so the bound that governs the two-lap blocks comes from the louder instrumental null beside it, not from this one.
The feature-matched replicate, eight shires a side. Post hoc, its dipole points 87.1° against the headline swap’s 89.2° — uncleaned it points −21.9°, about 112° away, so the agreement is the cleaning’s doing. Its size stays open. All five swap movies share one ±400 mK scale, which the strongest frames clip.
Slide the work one column east and the warm patch slides with it
Eight shires sliding one column at a time, 17 s a position, five laps averaged. The two end positions clip the shared ±400 mK scale.
The five positions march west to east, as the eight running shires slide one column at a time
Post-hoc half-patch contrast along the west–east template, in mK, mean over five laps.
The matched-filter position advances in the same order in the averaged run — one chance in 120 — and, on the cleaned frames, in every one of the five laps; without the cleaning, in three of five. The card’s total draw spreads 0.391 W across the five, and on the perpendicular axis all five sit within ±21 mK. Positions post hoc.
Turn the whole chip on and the whole board glows — this is what every other movie cancels
All 32 shires on for 42 s, then idle for 127 s, three cycles at 12× real time, on a ±8,000 mK scale against the swaps’ ±400.
Switch every shire on — 768 threads, about +39 W measured — and the frame mean rises +7.1 K in 42 s, half of it within 14.8 s. That is a change in total power, the one thing the swaps are built to cancel: their whole-frame means sit between −79 and +94 mK, this one at +7,100 mK.
Honestly: the governor stepped the clock in every on-phase and the die reached 69 °C, because the registered 64 °C gate was keyed to 1,024 threads and the amendment ran 768. The clean power law would have drawn 49 W here, and the governor is why it did not. Nothing was unsafe, but the arrival time is an upper bound, not the registered measurement.
Power follows how many threads run, not where they run
Switched board watts by thread count, above a 22.69 W idle
128 threads · 4 shires × 32
+12.75 W
128 threads · 8 shires × 16
+13.00 W
128 threads · 16 shires × 8
+13.26 W
256 threads
+20.0 W
384 threads
+26.7 W
512 threads
+34.7 W
1,024 threads — throttled, off the line
+43.8 W
Rust bars are the three shapes that run the same 128 threads in different places on the die.
Three shapes of 128 threads draw 0.51 W apart — a hair outside the half-watt the plan allowed, and not proof of anything either way: the legs ran in ascending order and carry 0.8–1.1 W of warm-up drift. Corrected for that drift they agree to 0.37 W and the order reverses. The ladder fits 5.80 ± 0.30 W plus 55.8 ± 1.0 mW a thread, not 40.6 mW from zero. Those six watts of waking up never move with the work: they brighten every picture and shift none.
The board blurs heat over 3–5 cm, so there is no shape and no scale bar
one shire
one broad blob
a row of shires
the same broad blob
a checkerboard
nothing at all
What the board does to a pattern before the camera sees it. Shapes, not measurements.
Copper spreads heat 3–5 cm before it reaches the camera — wider than the die itself. A single shire arrives as the same blob as any other pattern of the same power, a checkerboard arrives as nothing, and a better camera buys none of it back: the limit is the board.
The die’s footprint also fills the frame, so the pre-registered gate for a distance scale got no input. Directions only: no distances, no scale bar.
The same camera, live
Connecting…
The same camera, live. The picture is auto-gained, so its pixels carry no temperature. Anyone who can open this page can watch — signed out, the video starts after spacesheep’s quick check for people in the picture, or once you sign in.
Pre-registered — masks, orders, predictions and falsifiers frozen at , seven minutes before the first session.
Amendment A1, , before anything was reduced: the ladder measured about 2.5× the designed power a thread, so 12-shire sets ran at 16 threads a shire and 8-shire sets at 24 — 192 either way — and a 17 s west–east swap was dropped.
Post hoc, labelled throughout: per-pixel shutter-step removal, and contrasts read along the other session’s templates. Every pre-registered conclusion survives without them except the replicate — uncleaned its dipole points −21.9°, about 112° away, so both its size and its direction are the cleaning’s doing — and the sweep’s lap-by-lap ordering, which falls from 5 of 5 laps to 3 of 5.
The camera was left as the public stream has it. Its shutter fired every 8–10 s under load, so the registered exclusion window could not be applied. The same rule says a block that would lose more than 30% of its frames is reported as a bound, not a value — and the window would have removed 45–82% of every block, so under the rule as written every amplitude here is a bound. Where the rule could be approximated the answer does not move.
Registered p floors: four laps cannot go below 0.125, two laps below 0.5; the five-arm blocks are not so limited. And one bug: the bloom’s 64 °C die gate was guarded on a thread count A1 had changed, so it never read the die.
Also off the plan: the amendment’s power branch fired on an extrapolation rather than the registered rung; the hero’s session ran before the gain map existed, so it was analysed without one; the orientation fit the protocol registers for the hero was excluded by the pipeline and run by hand; the registered clock fallback was never coded; the instrumental null’s second, two-lap reduction was never produced; and the movies render 0.5 s bins at 24 fps with a 2×2 spatial bin against a frozen table of 1 s bins at 8–10 fps and 4×4 — finer, uniform across the set, and burnt into every frame.