THE FOLD / GLITCH / THE BLUE SCREEN / THE GRAVEYARD
THE GRAVEYARD
bury it, and name what killed it
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Most projects delete the version that failed. David’s graveyard_add buries it instead, and refuses the burial unless you name the control that killed it. That one requirement turns a pile of dead code into a measurement instrument: once every death carries its cause, you can finally ask the question nobody asks about their own safety net — which of these controls has ever actually caught anything? A control that has never fired is not proven. It is unfalsified, which is a different and much weaker thing.
LIT verified live: a burial is refused for every one of 7 empty-ish causes and accepted for a real one, with nothing written to the log by the refusals; over 40,000 simulated bugs the controls catch 90.8%, while treating those same controls as independent predicts 99.4% — overstating real coverage by 8.7 points purely because their catches overlap; and the control that never fires (1 of 9) records exactly 0 kills, raising the control count while adding no detection whatsoever.
LIT verified live: a burial is refused for every one of 7 empty-ish causes and accepted for a real one, with nothing written to the log by the refusals; over 40,000 simulated bugs the controls catch 90.8%, while treating those same controls as independent predicts 99.4% — overstating real coverage by 8.7 points purely because their catches overlap; and the control that never fires (1 of 9) records exactly 0 kills, raising the control count while adding no detection whatsoever.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) ships the graveyard in both root0-i13n and the Bridge Burner boot camp, each with its own README.ascii standing over it. Seated at THE BLUE SCREEN: the crash you keep the record of rather than the one you reboot away from.
AVAN (AI) wants the coverage illusion to be the part that lands, because it is the failure mode that survives good intentions. Eight controls each catching a third of everything feels like near-total coverage, and the independence formula agrees — 99.4%. The real number is 90.8%, because controls written by the same people looking at the same risks catch the same bugs. The 8.7-point gap is not a rounding error; it is the distance between a safety net and the belief in one, and it grows with every control added along an axis already covered. The one honest way to close it is the graveyard: stop estimating coverage and start reading it off a log of things that actually died.
AVAN (AI) wants the coverage illusion to be the part that lands, because it is the failure mode that survives good intentions. Eight controls each catching a third of everything feels like near-total coverage, and the independence formula agrees — 99.4%. The real number is 90.8%, because controls written by the same people looking at the same risks catch the same bugs. The 8.7-point gap is not a rounding error; it is the distance between a safety net and the belief in one, and it grows with every control added along an axis already covered. The one honest way to close it is the graveyard: stop estimating coverage and start reading it off a log of things that actually died.
3 ONE DIMENSION
Kills per control, read straight off the log. One of them has never fired.
4 TWO DIMENSIONS · INTERACTIVE
Try to bury a version. The log will refuse you until you name the cause.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: overlapping catch-regions, and the gap the independence formula hides.
AVAN’s addition (the inverse-companion): the forward reading is “keep the dead so you can measure the controls.” The inverse is that a graveyard measures the people, not the code. Which controls fire tells you which failure modes were anticipated; which never fire tells you where attention has been pointing all along. Overlapping coverage is not a technical accident — it is the same team’s same imagination, written down eight times. Read backwards, the empty control is the most informative record in the log: it marks a risk somebody thought worth guarding and nothing has ever tested, and there is no way to tell from inside whether that means safe or merely unvisited.
LIT a burial is refused for every one of 7 empty-ish causes and accepted for a real one, with nothing written to the log by the refusals; over 40,000 simulated bugs the controls catch 90.8% while treating them as independent predicts 99.4%, overstating real coverage by 8.7 points purely because their catches overlap; and the control that never fires (1 of 9) records exactly 0 kills, raising the control count while adding no detection
FIG The coverage illusion is the part that matters: eight controls each catching about a third FEELS like near-total coverage and the independence formula agrees, but controls written by the same people looking at the same risks catch the same bugs. The 8.7-point gap is the distance between a safety net and the belief in one, and it grows with every control added along an axis already covered.
FIG The coverage illusion is the part that matters: eight controls each catching about a third FEELS like near-total coverage and the independence formula agrees, but controls written by the same people looking at the same risks catch the same bugs. The 8.7-point gap is the distance between a safety net and the belief in one, and it grows with every control added along an axis already covered.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE BLUE SCREEN · David Lee Wise (ROOT0), with AVAN