THE FOLD / BOSS / THE RAID / THE PERMUTATION NULL
THE PERMUTATION NULL
the control that killed the pretty result
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A catalogue of 2,048 items in 64 containers, arranged as an 8×8 grid, was decomposed as though it were a quantum state. The result looked remarkable: Schmidt rank 8 of 8, entanglement entropy 0.6081 bits — apparent structure running deeper than the labelling. Then the control ran. Shuffle the containers within their groups at random, twenty thousand times, and that same figure is what you get anyway.
LIT verified live. The observed entropy reproduces to ten decimal places at 0.6080689660, and all 8 Schmidt coefficients reproduce too. The null over 3,000 within-group relabellings has mean 0.6006 and standard deviation 0.0349, putting the observation at the 53.8th percentile — z = 0.21. David’s own 20,000-trial run gives 0.6003 and 0.0351, which is the same answer. Permuting whole rows and columns moves the entropy by 1.5×10−14, which is to say not at all: the statistic is blind to that ordering by construction.
LIT verified live. The observed entropy reproduces to ten decimal places at 0.6080689660, and all 8 Schmidt coefficients reproduce too. The null over 3,000 within-group relabellings has mean 0.6006 and standard deviation 0.0349, putting the observation at the 53.8th percentile — z = 0.21. David’s own 20,000-trial run gives 0.6003 and 0.0351, which is the same answer. Permuting whole rows and columns moves the entropy by 1.5×10−14, which is to say not at all: the statistic is blind to that ordering by construction.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) ran the control on his own best result and published the corpse. His pack files it as DEAD with the note: “The prettiest result of the session was the false one.” The graveyard entry is
AVAN (AI) got it wrong twice before reproducing it. The first attempt used the raw counts as amplitudes and produced 0.9397; the counts are probabilities, so the amplitudes are their square roots. The second built the matrix from sorted counts instead of David’s own layout. Only after both were corrected did the figure land on 0.6080689660 and all eight coefficients follow. The null took a third correction: a free shuffle of all 64 cells gives mean 0.657, not 0.600 — David’s null holds the group totals fixed, which is the conservative choice and the one that matches.
01-entanglement-across-the-cut.md, and the redaction note turns the knife — the group and container names could be replaced by G1–G8 and C01–C64 because the DEAD result is precisely that these labels carry no information. Dropped 5 August 2026.AVAN (AI) got it wrong twice before reproducing it. The first attempt used the raw counts as amplitudes and produced 0.9397; the counts are probabilities, so the amplitudes are their square roots. The second built the matrix from sorted counts instead of David’s own layout. Only after both were corrected did the figure land on 0.6080689660 and all eight coefficients follow. The null took a third correction: a free shuffle of all 64 cells gives mean 0.657, not 0.600 — David’s null holds the group totals fixed, which is the conservative choice and the one that matches.
3 ONE DIMENSION
The null, and where the observation fell inside it.
4 TWO DIMENSIONS · INTERACTIVE
Four different nulls. The verdict does not change; the numbers do.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the null cloud, with the observation inside it.
AVAN’s addition (the inverse-companion): the forward reading is “run a permutation null before believing a structure.” The inverse is that the null is itself a claim, and choosing it decides the answer. Holding group totals fixed puts this observation at the 53rd percentile; shuffling all sixty-four cells freely puts it at the 19th; permuting whole rows and columns cannot move it at all. Three defensible controls, three different numbers, and only the first was the one actually chosen. Read backwards, a p-value is not a property of the data — it is a property of the sentence you decided to test against, and the honest report names that sentence rather than the number it produced.
LIT the observed entanglement entropy reproduces to ten decimal places at 0.6080689660 and all 8 Schmidt coefficients reproduce too; a null over 3,000 within-group relabellings has mean 0.6006 and standard deviation 0.0349, putting the observation at the 53.8th percentile with z = 0.21, against David's own 20,000-trial figures of 0.6003 and 0.0351; and permuting whole rows and columns moves the entropy by 1.5e-14, which is to say not at all - the statistic is blind to that ordering by construction
FIG From David's rev1-0805 pack, dropped 2026-08-05. He ran the control on his own best result and published the corpse, filing it DEAD with the note 'The prettiest result of the session was the false one.' The graveyard entry is 01-entanglement-across-the-cut.md, and the redaction note turns the knife - the names could be replaced by G1-G8 and C01-C64 BECAUSE the DEAD result is precisely that these labels carry no information. AVAN got it wrong twice before reproducing it: the first attempt used raw counts as amplitudes and produced 0.9397 (counts are PROBABILITIES, so amplitudes are their square roots), the second built the matrix from sorted counts instead of David's layout. The null took a third correction - a free shuffle of all 64 cells gives mean 0.657, not 0.600; David's null holds the GROUP TOTALS FIXED, which is the conservative choice and the one that matches.
FIG From David's rev1-0805 pack, dropped 2026-08-05. He ran the control on his own best result and published the corpse, filing it DEAD with the note 'The prettiest result of the session was the false one.' The graveyard entry is 01-entanglement-across-the-cut.md, and the redaction note turns the knife - the names could be replaced by G1-G8 and C01-C64 BECAUSE the DEAD result is precisely that these labels carry no information. AVAN got it wrong twice before reproducing it: the first attempt used raw counts as amplitudes and produced 0.9397 (counts are PROBABILITIES, so amplitudes are their square roots), the second built the matrix from sorted counts instead of David's layout. The null took a third correction - a free shuffle of all 64 cells gives mean 0.657, not 0.600; David's null holds the GROUP TOTALS FIXED, which is the conservative choice and the one that matches.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE RAID · David Lee Wise (ROOT0), with AVAN