GameBoyGhost

What happened

Most project pages show you what worked. This one shows you what broke — and why that matters more.

Every story here is a real failure, caught by the project's own checks before it could corrupt a sealed result, and written up the way an incident-response team writes a postmortem:what happened, how it was noticed, how it was fixed, what was learned. No one gets blamed, no failure is quietly dropped, and where the records don't say something, this page says so instead of guessing.

These aren't footnotes. They're the product. If you read one page, read this one.

The stories

1The receipt-clamp saga

A safety timeout put in the wrong layer refused the expert's own correct route. It took four phases to move it where it belonged.

2A replay that would have passed for the wrong reason

Proving where code came from is not the same as proving what it read. A second check caught what the first couldn't.

3When the design said no

A task prompt contained a decision that contradicted the written design. The agent stopped and said so, and the call was withdrawn.

4The files that weren't on the list

A hand-kept commit list missed three changed files. A two-way check against git statusnow makes that impossible.

5Stuck at the cave door

The oracle's first real run pressed against a cave door for 1,500 decisions. The cause was a 4-pixel alignment tolerance.

6The file git archive leaves behind

"Everything is in git" is never quite true. The replay tooling now refuses when a frozen file is missing.

7The S2 gate re-lock: 99% → 97%

The first lock set the wrong bar and pinned exact commits. It was replaced before it ever ran.

8The crumbling-floor fall

The shield stayed up too long, a bat pushed Link back, the floor gave way. One change towhen the shield raises fixed it. See it before and after.

Not yet written up

Three more real incidents, from the operating side, still need their postmortems:

·The Studio wipe and scripted rebuild

Deliberately destroying and rebuilding the primary machine — and the discipline that made it survivable.

·The SMB NUL-padding corruption

A file silently padded with NUL bytes, traced to a defect in writing over a network share. Changed the practice: write natively, never over SMB.

·The v4 contamination catch

862 rows of the "held-out" set were actually in the training split. Found it, counted it, published it.

The journal

Below the stories, the living record: status (what's sealed, certified, and planned right now), results (each experiment's honest outcome), and daily logs (the automated updates). Every entry carries a source note naming the exact files its facts came from — one click from any claim to its proof.