Notes from the workbench.

The infrared had to go

A 74LVC2G17 where the infrared transceiver was, and a ⌀70 steel ball to put it in. The shell weighs 86.1 g, the Poké Ball look turns out to be paint, and the USB-C plug cannot reach its socket.

The leaderboard that does not exist

Three of our own published speed numbers did not survive being measured properly. What replaced them puts the fastest emulator on this machine at zero on the hardest accuracy suite in the field.

The trainer who saw him black out

He fainted next to a Bug Catcher, blacked out, and woke up in Viridian City. Forty-five tiles and 385 decisions later he stepped back into the forest and the Bug Catcher fought him where he stood — on the doorway tile, nine decisions after crossing the boundary.

The trap was always on the map

A compiled copy of the student finished 0 of 10 cold boots where the original finished 8. Nothing was retrained and the two paths pick the identical button a thousand times running. The obvious conclusion — the new thing is broken — was wrong.

A perfect score and 361,567 false alarms

A recogniser scored a held-out balanced accuracy of 1.000 and then said yes to 361,567 of 1,383,924 frames of real play. Five rounds, one of them poisoned by the tool's own refusal message — and the bug that was actually ending runs turned out to have no weights in it at all.

Teaching a network to notice

Five triggers that read the cartridge were replaced with networks that look at the screen. Held-out accuracy said 1.0000 and the probe said zero false alarms — and the labels themselves were poisoned, in a way only a person looking at the pile could see.

Exponential decay of belief

A confidence gate locked onto one button, fourteen times out of fourteen the same way. Letting the belief in it decay instead of banning it took the chain from 594 to 600 of 600 — and taught it to heal itself mid-run for free.

One character becomes three thousand

A single character stands in a bedroom, then three thousand of it walk themselves into a Pokéball, in one unbroken shot. Getting there took a stage with no map and six real bugs, most of them nowhere near the camera.

Three bytes a cell, doing nothing

A reference clip filled the cell behind every glyph instead of drawing bright characters on black. The plumbing to do the same here turned out to already be wired in, end to end, and doing nothing.

The chiptune that outgrew its blog post

A demo video needed music with no licence to worry about, so a 114-line Python script wrote its own chiptune. The captain asked why a generator that times itself to a video lived beside a blog asset instead of inside the renderer — and eighteen minutes later, it didn't.

The number that could have sunk it

Could Remotion open on Pokémon Blue's title screen, pull back to the whole world map, and burst out real battle footage from where it happened? Two of the three ingredients already existed. The one real unknown came back at 4.3 milliseconds a frame.

The one clean miss

The captain wanted the display style of a specific YouTube editor. Six of the seven techniques it needed already existed. The seventh was small, and it still isn't built.

Every failure was the same starter

A twenty-six agent swarm scored 88.5% and looked healthy. Watching the film instead of reading the table showed all three failures were the same starter Pokémon — one chance in 2,600.

The wall you hit walking home

An agent learned the walk to the shop perfectly and could not learn the walk back. Two measured rates that sum to a hundred are the signature of an impossible task, not a hard one.

The day that added nothing

Five identical training runs scored 100, 92, 92, 85 and 77 per cent. Every ablation published before that lacked the control.

293 runs stopped on the same text box

A nine-link chain collapsed from 300 of 300 to 7. The student was doing exactly what its supervisor had told it to do.

The one failure in twelve thousand

Four certification cells of 3,000 attempts produced exactly one failure. Somebody was standing on the corner tile, and stayed there for 4.7 minutes of game time.

Reading the screen without reading the letters

At 40×36 no letterform survives. Six in-game messages are still 97.2% separable — and the first two experiments that said otherwise were both too small to say anything.

271 colours in a four-shade picture

Half-block terminal output was blurry on GNOME terminals and crisp on WezTerm. The cause was a hard-coded smoothing filter four lines down in a dependency — and doing the sampling ourselves is cheaper.

The transfer controller owns the bus

Object DMA is usually modelled as a copy. It is also an address-bus conflict, and the sprite scan is on the other end of those lines.

The red we refused to fix

One sweep would have turned a failing test green. Taking it would have moved every Color cartridge's divider on the evidence of a single row.

The double-speed bug that cost 2,304 pixels

A register-commit trick worth exactly one dot of time vanished when the CPU clock doubled — and a Color game drew a wrong stripe across the screen.

TACKLE is 95 per cent and we said 94

Nothing failed. A number had been computed from a belief rather than read from the cartridge, and it sat in four files for as long as they existed.

The ledge in Viridian City

Seventeen runs of six hundred died on the same row of the same town. A flood fill said the tile was reachable, and the flood fill was right.

Let go of the D-pad

One line took a cartridge map capture from 113 maps to 216. A door is a warp on both sides, and a player still holding the direction walks straight back out.

The door rule

The cartridge says where outdoor maps go and says nothing at all about interiors. Placing 190 rooms is a decision, and it was settled by measurement.

A hundred checks in one session

A new check found a real bug on the day it was written — one that thousands of rows across five public suites could not see, because of the shape they have.

Three ways this cartridge was wrong about itself

A verdict decided by which checks failed above it, a subroutine that ate a measurement, and two checks that skipped on every console, forever.

Four bugs, one line each

A dropped field, a RETURNING clause that returns the wrong thing, a bigint that arrived as a string, and a grid that could not get narrower than 16rem.