THE FOLD / SPAWN / NULL ISLAND / THE UNSTATED CONDITION
THE UNSTATED CONDITION
an identity with no domain attached
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
“Addition is associative.” “Dividing then multiplying gets you back.” “The square root of a square is the number.” Each of these is true, and each is true somewhere — on a domain nobody wrote down. Move the same identity to different numbers and it starts failing, not because the arithmetic is wrong but because the sentence was published without the condition that made it hold.
LIT verified live. Two exact anchors first:
LIT verified live. Two exact anchors first:
0.1 + 0.2 - 0.3 is not merely small, it is exactly 2−54; and the doubles in [1,2) sit on an exact grid of spacing 2−52. Then five identities across three ranges, 20,000 trials per cell. On [1,2), three of the five fail zero times. Let the exponents range over 2−60 to 260 and all five fail beyond three standard errors, with 3 flipping from never-observed-to-fail to measurably-false on nothing but the range. Zero failures in a sample is not a proof of exactness — a wider check of 300,000 triples on [1,2) also found none, and Sterbenz’s lemma does not cover this case, so what is established is the observation, not the mechanism.2 HOW IT WAS WEAVED · AI + HUMAN
David (human) made this the third layer of the honest read, alongside what ran and what is a figure: “the CONDITION — what has to be true for the real part to hold.” His checker flags a deliverable that claims a result and never names one. This sphere is that rule turned on arithmetic, where the conditions are unusually easy to measure and unusually often left out.
AVAN (AI) built the first version drawing every operand from [1,2), and three of the five identities failed zero times in 200,000 trials — which looked at first like a broken measurement. It was not. It was the subject. Rather than move the range until the identities broke, the range became the variable and the result is the sweep in window 4: the same five identities, the same code, three domains, and the truth of each claim changing with the domain and nothing else.
AVAN (AI) built the first version drawing every operand from [1,2), and three of the five identities failed zero times in 200,000 trials — which looked at first like a broken measurement. It was not. It was the subject. Rather than move the range until the identities broke, the range became the variable and the result is the sweep in window 4: the same five identities, the same code, three domains, and the truth of each claim changing with the domain and nothing else.
3 ONE DIMENSION
Five identities, three ranges, one code path.
4 TWO DIMENSIONS · INTERACTIVE
Move the domain and watch a true statement become a false one.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the region where the identity holds, floating in the space of inputs.
AVAN’s addition (the inverse-companion): the forward reading is “name the condition or the claim is incomplete.” The inverse is that every claim has infinitely many conditions and you can only ever name the ones you thought of. The identities here fail on range; they also depend on rounding mode, on whether an intermediate stayed in a wider register, on the order the compiler chose. Naming a condition does not close the claim — it moves the boundary out one step and leaves the same open edge beyond it. Read backwards, the honest read is not a way to make a claim complete but a way to say where you stopped looking, which is the only part anyone can actually check.
LIT 0.1 + 0.2 - 0.3 is not merely small, it is exactly 2^-54, and the doubles in [1,2) sit on an exact grid of spacing 2^-52; across five identities and three ranges at 20,000 trials per cell in the page, three of the five fail ZERO times on [1,2) - observed, not proven exact - and once the exponents range over 2^-60 to 2^60 all five fail beyond three standard errors, with 3 flipping from exactly-true to measurably-false on nothing but the range
FIG From the third layer of David's honest read, alongside what ran and what is a figure: 'the CONDITION - what has to be true for the real part to hold.' His checker flags a deliverable that claims a result and never names one. AVAN built the first version drawing every operand from [1,2) and three of the five identities failed ZERO times in 200,000 trials, which looked at first like a broken measurement. It was not - it was the subject. Rather than move the range until the identities broke, the range became the variable. Nothing here is a bug: IEEE 754 rounds every single operation correctly, and the fault is that the claim was published without its domain, which is a different failure entirely.
FIG From the third layer of David's honest read, alongside what ran and what is a figure: 'the CONDITION - what has to be true for the real part to hold.' His checker flags a deliverable that claims a result and never names one. AVAN built the first version drawing every operand from [1,2) and three of the five identities failed ZERO times in 200,000 trials, which looked at first like a broken measurement. It was not - it was the subject. Rather than move the range until the identities broke, the range became the variable. Nothing here is a bug: IEEE 754 rounds every single operation correctly, and the fault is that the claim was published without its domain, which is a different failure entirely.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of NULL ISLAND · David Lee Wise (ROOT0), with AVAN