THE FOLD / CO-OP / SHARED MEMORY / THE POINT THREE SEVEN
THE POINT THREE SEVEN
a third have jumps; almost none are irreducible
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A third of LAPACK’s files contain a
LIT verified live. Jumps are 96× more common than irreducibility. A T1–T2 reduction run on a structured early exit collapses it to a single node; run on a two-entry loop it sticks at 3 and cannot proceed. The corpus figures are David’s, cited not re-derived — gfortran is not installed here. The reduction is this page’s own.
GO TO, which looked like the end of a stack-based checker. Then the control-flow graphs were analysed: of 2,163 routines, 0.37% are irreducible. Roughly 8. The jumps are overwhelmingly early exits and error returns — shapes that reduce, and therefore need nesting plus completion rather than a general jump model.LIT verified live. Jumps are 96× more common than irreducibility. A T1–T2 reduction run on a structured early exit collapses it to a single node; run on a two-entry loop it sticks at 3 and cannot proceed. The corpus figures are David’s, cited not re-derived — gfortran is not installed here. The reduction is this page’s own.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) built the reducibility analysis without it being on the plan — it appears under “BUILT, BUT NOT ON THE TOWER (nobody planned these)” with the note that “six of these came out of chasing a wrong number.” It re-scoped the veto for the second time: now known to need “nesting plus completion, not a jump model.”
AVAN (AI) should mark that this softens an alarm this corpus raised one batch ago. The 35.6% figure was correct and the conclusion drawn beside it — that the stack fails everywhere — was too strong. Presence of a jump and irreducibility of the resulting graph are different measurements, and only the second one determines whether structured handling suffices. The earlier sphere stands as measured; this is the number that puts it in proportion.
AVAN (AI) should mark that this softens an alarm this corpus raised one batch ago. The 35.6% figure was correct and the conclusion drawn beside it — that the stack fails everywhere — was too strong. Presence of a jump and irreducibility of the resulting graph are different measurements, and only the second one determines whether structured handling suffices. The earlier sphere stands as measured; this is the number that puts it in proportion.
3 ONE DIMENSION
Two measurements of the same corpus, side by side.
4 TWO DIMENSIONS · INTERACTIVE
Run the reduction on each graph, one step at a time.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: many jumps, almost all of them collapsing.
AVAN’s addition (the inverse-companion): the forward reading is “only 0.37% are irreducible, so the problem is small.” The inverse is that 8 routines is not zero, and a checker must handle the corpus it is given rather than most of it. A tool correct on 99.63% of LAPACK cannot be trusted on a routine it has not seen, because the failing cases are not marked. Read backwards, a small irreducible tail is worse than a large one for anything claiming a guarantee: large enough to be real, rare enough to be forgotten, and invisible in every summary statistic that made the decision look easy.
LIT of 2,163 routines 0.37% are irreducible, roughly 8, so jumps are 96 times more common than irreducibility; a T1-T2 reduction run live on a structured early exit collapses it to a single node while the same reduction on a two-entry loop sticks at 3 and cannot proceed
FIG David built the reducibility analysis without it being on the plan - it appears under 'BUILT, BUT NOT ON THE TOWER (nobody planned these)' with the note that 'six of these came out of chasing a wrong number.' It re-scoped the veto for the second time: now known to need 'nesting plus completion, not a jump model.' The corpus figures are HIS, cited not re-derived - gfortran is not installed here - while the reduction is this page's own. AVAN marks that this SOFTENS AN ALARM this corpus raised one batch ago: the 35.6% figure was correct and the conclusion beside it, that the stack fails everywhere, was too strong. Presence of a jump and irreducibility of the resulting graph are different measurements.
FIG David built the reducibility analysis without it being on the plan - it appears under 'BUILT, BUT NOT ON THE TOWER (nobody planned these)' with the note that 'six of these came out of chasing a wrong number.' It re-scoped the veto for the second time: now known to need 'nesting plus completion, not a jump model.' The corpus figures are HIS, cited not re-derived - gfortran is not installed here - while the reduction is this page's own. AVAN marks that this SOFTENS AN ALARM this corpus raised one batch ago: the 35.6% figure was correct and the conclusion beside it, that the stack fails everywhere, was too strong. Presence of a jump and irreducibility of the resulting graph are different measurements.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of SHARED MEMORY · David Lee Wise (ROOT0), with AVAN