THE FOLD / GLITCH / HEISENBUG / THE SAFE DIRECTION
THE SAFE DIRECTION
only one kind of error summons a person
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A checker can be wrong two ways. It can cry wolf, or it can wave something through. These are not symmetric errors, because only one of them summons a human to come and look.
LIT verified live. 10,000 items at a 5% defect rate is 500 defects. Under a policy where every flag is investigated and every pass is not, sweeping the false-alarm rate from 0 to 50% leaves the surviving defect count pinned at 50 — unchanged at every point, while the investigation load climbs from 0 to 4,750. Sweeping the miss rate over the same range takes survivors from 0 to 250, a straight line of slope 500, exactly the defect count. One error direction costs effort and cannot cost correctness. The other costs correctness and is free.
LIT verified live. 10,000 items at a 5% defect rate is 500 defects. Under a policy where every flag is investigated and every pass is not, sweeping the false-alarm rate from 0 to 50% leaves the surviving defect count pinned at 50 — unchanged at every point, while the investigation load climbs from 0 to 4,750. Sweeping the miss rate over the same range takes survivors from 0 to 250, a straight line of slope 500, exactly the defect count. One error direction costs effort and cannot cost correctness. The other costs correctness and is free.
2 HOW IT WAS WEAVED · AI + HUMAN
The asymmetry between false positives and false negatives under an investigate-on-flag policy is standard decision theory.
AVAN (AI) built this the day it happened. A null test on another program reported ‘0 of 3 refused’ when the program had refused all three; the fault was in the checker, which caught the wrong exception class. It was found within one run — because the alarming number sent someone looking. Had the same bug reported 3 of 3 when nothing was refused, nothing would have looked, and the sweep above says how long that lasts: indefinitely.
AVAN (AI) built this the day it happened. A null test on another program reported ‘0 of 3 refused’ when the program had refused all three; the fault was in the checker, which caught the wrong exception class. It was found within one run — because the alarming number sent someone looking. Had the same bug reported 3 of 3 when nothing was refused, nothing would have looked, and the sweep above says how long that lasts: indefinitely.
3 ONE DIMENSION
Raise the false alarms: nothing breaks. Raise the misses: everything does.
4 TWO DIMENSIONS · INTERACTIVE
Turn each dial and watch which one moves the survivors.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object.
AVAN’s addition (the inverse-companion): the forward reading is to prefer a noisy checker over a quiet one. The inverse is that a checker’s error rate is not a property of the checker — it is a property of the checker and the policy together. The false alarm is harmless only because somebody looks; make investigation expensive enough that flags start being dismissed unread, and the two directions collapse into one. Read backwards, every ‘fail loudly’ design is quietly relying on an attention budget nobody measured, and it degrades into the silent kind the moment that budget runs out.
LIT 10,000 items at a 5% defect rate is 500 defects, and under a policy where every flag is investigated and every pass is not, sweeping the false-alarm rate from 0 to 50% leaves the surviving defect count pinned at 50 at every point while the investigation load climbs from 0 to 4,750, whereas sweeping the miss rate over the same range takes survivors from 0 to 250 in a straight line of slope 500, exactly the defect count - one error direction costs effort and cannot cost correctness, the other costs correctness and is free
FIG The asymmetry between false positives and false negatives under an investigate-on-flag policy is standard decision theory. AVAN built this the day it happened: a null test on another program reported '0 of 3 refused' when the program had refused all three, the fault being in the checker, which caught the wrong exception class. It was found within one run because the alarming number sent someone looking. Had the same bug reported 3 of 3 when nothing was refused, nothing would have looked, and the sweep says how long that lasts: indefinitely.
FIG The asymmetry between false positives and false negatives under an investigate-on-flag policy is standard decision theory. AVAN built this the day it happened: a null test on another program reported '0 of 3 refused' when the program had refused all three, the fault being in the checker, which caught the wrong exception class. It was found within one run because the alarming number sent someone looking. Had the same bug reported 3 of 3 when nothing was refused, nothing would have looked, and the sweep says how long that lasts: indefinitely.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of HEISENBUG · David Lee Wise (ROOT0), with AVAN