THE FOLD / GLITCH / SEGFAULT / THE NOISE CONTROL
THE NOISE CONTROL
the control that says no
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Test a validator by feeding it correct programs and you learn almost nothing, because a function that returns true unconditionally passes that test perfectly. The only measurement that separates a checker from a rubber stamp is what it does with things that are not programs. And the quantity to report is the rate, not the count — a different noise generator produces a different number of rejections while saying nothing different about the checker.
LIT verified live: 6,000 well-formed programs, generated balanced by construction, are accepted 6,000 times — 100.00%. 6,000 random instruction sequences are rejected 5,985 times — 99.75%, with a standard error of 0.064 percentage points. A second, differently-biased noise generator rejects 5,948 — a different count and a different rate, 99.13%, which sits about 9.6 standard errors away and is therefore not the same measurement at all.
LIT verified live: 6,000 well-formed programs, generated balanced by construction, are accepted 6,000 times — 100.00%. 6,000 random instruction sequences are rejected 5,985 times — 99.75%, with a standard error of 0.064 percentage points. A second, differently-biased noise generator rejects 5,948 — a different count and a different rate, 99.13%, which sits about 9.6 standard errors away and is therefore not the same measurement at all.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) wrote this into pocket-machine’s fifth window as a dissent against his own first window: the carried-forward record claimed 1995 rejected, 5 passed of 2000, and his rebuild measured something else. His ruling was that the rate is the comparable quantity and the count is not, so the old figure stays AMBER and is not promoted. He seated this at SEGFAULT — the fault that only fires on input nobody meant to write.
AVAN (AI) shipped a first version whose well-formed generator produced zero valid programs out of 20,000 on a wider run — a bug in the drain loop meant the “correct” arm was silently empty, and the acceptance figure of 0.00% was measuring nothing. It was caught only because a gate demanded 100% and got 0. That failure is worth keeping visible: the arm that is supposed to pass is the one where a broken generator hides best, because an empty test set produces no complaints of its own. The surviving figure was then checked against an exact calculation — a dynamic program over the noise generator’s own distribution, with no sampling anywhere in it — which puts the true rejection probability at 99.7185%. The measured 99.75% sits 0.49 standard errors from it.
AVAN (AI) shipped a first version whose well-formed generator produced zero valid programs out of 20,000 on a wider run — a bug in the drain loop meant the “correct” arm was silently empty, and the acceptance figure of 0.00% was measuring nothing. It was caught only because a gate demanded 100% and got 0. That failure is worth keeping visible: the arm that is supposed to pass is the one where a broken generator hides best, because an empty test set produces no complaints of its own. The surviving figure was then checked against an exact calculation — a dynamic program over the noise generator’s own distribution, with no sampling anywhere in it — which puts the true rejection probability at 99.7185%. The measured 99.75% sits 0.49 standard errors from it.
3 ONE DIMENSION
Two arms. Only the second one can fail.
4 TWO DIMENSIONS · INTERACTIVE
Run the noise. Then swap the checker for one that always says yes.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the space of sequences, with the accepted region carved out of it.
AVAN’s addition (the inverse-companion): the forward reading is “test the checker on things that are not programs.” The inverse is that the rejection rate is mostly a measurement of the noise, not of the checker. Make the noise easier and the rate climbs; make it adversarial and it falls. What the number actually reports is the overlap between one generator and one grammar, which is why two honest runs disagree and neither is wrong. Read backwards, the discipline is not “measure the rejection rate” but “name the generator whenever you quote one” — and a rate quoted without its source is a count wearing a percent sign.
LIT 6,000 well-formed programs, generated balanced by construction, are accepted 6,000 times - 100.00%; 6,000 random instruction sequences are rejected 5,985 times - 99.75%, with a standard error of 0.064 percentage points; and a second, differently-biased noise generator rejects 5,948, a different count and a different rate of 99.13%, about 9.6 standard errors away and therefore not the same measurement at all
FIG From David's pocket-machine, where this appears as a dissent against his own first window: the carried record claimed 1995 rejected / 5 passed of 2000, his rebuild measured otherwise, and his ruling was that the RATE is comparable and the count is not - so the old figure stays AMBER. AVAN shipped a first version whose well-formed generator produced ZERO valid programs; a bug in the drain loop left the 'correct' arm silently empty and the 0.00% acceptance figure was measuring nothing. It was caught only because a gate demanded 100% and got 0. The arm that is supposed to PASS is where a broken generator hides best.
FIG From David's pocket-machine, where this appears as a dissent against his own first window: the carried record claimed 1995 rejected / 5 passed of 2000, his rebuild measured otherwise, and his ruling was that the RATE is comparable and the count is not - so the old figure stays AMBER. AVAN shipped a first version whose well-formed generator produced ZERO valid programs; a bug in the drain loop left the 'correct' arm silently empty and the 0.00% acceptance figure was measuring nothing. It was caught only because a gate demanded 100% and got 0. The arm that is supposed to PASS is where a broken generator hides best.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of SEGFAULT · David Lee Wise (ROOT0), with AVAN