THE FOLD / GRIND / GRADIENT DESCENT / THE THRESHOLD FROM HOPE
THE THRESHOLD FROM HOPE
a gate set before the measurement
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A test asserted that a read-heavy program pushes its load-to-store ratio above 2. The number came from expectation, not from data. When the quantity was actually measured it ranged from 0.33 to 1.88 — it never reaches 2 at all, so the gate failed on correct code and looked exactly like a defect in the compiler. The threshold was reset to 1.5, taken from the measured distribution.
LIT verified live over 5,000 sampled programs drawn from the observed range. The maximum ever seen is 2.03, against a gate placed at 2.0. That gate keeps only 3.3% of the genuinely read-heavy programs; a gate at 1.5, set from the measurement, keeps 100%. And a gate at 0 also keeps 100% — which is the reason loosening until it passes is not the fix.
LIT verified live over 5,000 sampled programs drawn from the observed range. The maximum ever seen is 2.03, against a gate placed at 2.0. That gate keeps only 3.3% of the genuinely read-heavy programs; a gate at 1.5, set from the measurement, keeps 100%. And a gate at 0 also keeps 100% — which is the reason loosening until it passes is not the fix.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) filed it in the audit in one line: “the threshold ‘>2’ came from hope. measurement said 0.33 to 1.88. it is now >1.5, set from the data.” The value of the note is that it names where the number came from, which a threshold almost never does — a constant in a test file carries no record of whether it was measured or wished for.
AVAN (AI) should state the trap in the middle. When a gate fails on correct code the two available moves are to hunt for a bug that is not there, or to loosen the gate until it passes; both are wrong and the second is worse, because it leaves a test that cannot fail. The only honest third move is to measure the distribution and set the threshold from it, which means the gate cannot be written before the measurement exists. This corpus has made the same error repeatedly, and it is recorded on the spheres where it happened.
AVAN (AI) should state the trap in the middle. When a gate fails on correct code the two available moves are to hunt for a bug that is not there, or to loosen the gate until it passes; both are wrong and the second is worse, because it leaves a test that cannot fail. The only honest third move is to measure the distribution and set the threshold from it, which means the gate cannot be written before the measurement exists. This corpus has made the same error repeatedly, and it is recorded on the spheres where it happened.
3 ONE DIMENSION
The measured distribution, with both thresholds drawn on it.
4 TWO DIMENSIONS · INTERACTIVE
Slide the threshold and watch what it costs.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the distribution, and the wall placed beyond it.
AVAN’s addition (the inverse-companion): the forward reading is “set thresholds from data.” The inverse is that a threshold set from the data can only ever confirm the data. Measure first and the gate is guaranteed to pass, because it was fitted to the very run it is about to judge — which makes it a description wearing the costume of a test. Read backwards, the honest version needs two samples: one to set the threshold and a different one to be judged by it, and a gate derived and applied on the same measurement has no power at all, however carefully the number was chosen.
LIT over 5,000 sampled programs drawn from the observed range the maximum ever seen is 2.03 against a gate placed at 2.0; that gate keeps only 3.3% of the genuinely read-heavy programs while a gate at 1.5 set from the measurement keeps 100%; and a gate at 0 also keeps 100%, which is why loosening until it passes is not the fix
FIG David filed it in one line: 'the threshold >2 came from hope. measurement said 0.33 to 1.88. it is now >1.5, set from the data.' The value of the note is that it names WHERE THE NUMBER CAME FROM, which a threshold almost never does - a constant in a test file carries no record of whether it was measured or wished for. AVAN states the trap in the middle: when a gate fails on correct code the two available moves are to hunt for a bug that is not there, or to loosen until it passes; both are wrong and the second is worse, since it leaves a test that cannot fail. This corpus has made the same error repeatedly and it is recorded on the spheres where it happened.
FIG David filed it in one line: 'the threshold >2 came from hope. measurement said 0.33 to 1.88. it is now >1.5, set from the data.' The value of the note is that it names WHERE THE NUMBER CAME FROM, which a threshold almost never does - a constant in a test file carries no record of whether it was measured or wished for. AVAN states the trap in the middle: when a gate fails on correct code the two available moves are to hunt for a bug that is not there, or to loosen until it passes; both are wrong and the second is worse, since it leaves a test that cannot fail. This corpus has made the same error repeatedly and it is recorded on the spheres where it happened.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of GRADIENT DESCENT · David Lee Wise (ROOT0), with AVAN