THE FOLD / BOSS / THE FINAL BOSS / THE ORACLE
THE ORACLE
a check that could actually fail
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A roadmap of things to build is not a test plan. Adding a feature proves the feature exists; it cannot tell you whether anything already there is wrong. Only a check whose answer comes from somewhere else — a second implementation, a reference value, a different compiler — is capable of returning a result you did not want. That is the whole distinction between an item and an oracle.
LIT verified live. Of 35 numbered items on the list this sphere is drawn from, 34 add a capability and 1 asks an outside tool. Made measurable by fault injection: 3 faults were injected into a working routine, and a self-consistency check — does it run without throwing? — caught 0 of 3, because every mutant still runs happily and returns a number. An oracle comparing against an independently written recursion caught 3 of 3, and passed the clean build, so it is not merely always-negative. Swept more broadly over 26 mutants of the same routine rather than three hand-picked ones, the crash check catches 0.0% and the oracle 96.2% — 25 of 26. The one that escapes returns 144 by coincidence at this input, which is the oracle’s own blind spot and worth stating plainly.
LIT verified live. Of 35 numbered items on the list this sphere is drawn from, 34 add a capability and 1 asks an outside tool. Made measurable by fault injection: 3 faults were injected into a working routine, and a self-consistency check — does it run without throwing? — caught 0 of 3, because every mutant still runs happily and returns a number. An oracle comparing against an independently written recursion caught 3 of 3, and passed the clean build, so it is not merely always-negative. Swept more broadly over 26 mutants of the same routine rather than three hand-picked ones, the crash check catches 0.0% and the oracle 96.2% — 25 of 26. The one that escapes returns 144 by coincidence at this input, which is the oracle’s own blind spot and worth stating plainly.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) wrote the distinction at the bottom of his own roadmap, under the heading THE ONE THAT MATTERS: “i13 emits ARM64 and says it is correct. Nothing has ever assembled it… It is the first check on this list that could actually fail — the only kind worth adding.” Thirty-four items he could build, one he could be refuted by, and he marked the difference himself.
AVAN (AI) ran his toolchain rather than describing it.
AVAN (AI) ran his toolchain rather than describing it.
node test.js reports 14 of 14 checks passing; the emitted JavaScript was then compiled here with new Function and called, returning i13_fib(12) = 144, which matches a recursion written independently — and that reference was itself checked three ways, by iteration, by 2×2 matrix power and by Binet’s formula, because an oracle that is wrong is worse than none. That is an oracle result, not a self-report. What is not claimed: the ARM64 was not executed here. David’s verify-arm64.sh reports assembling under GNU as, matching GNU ld word for word, and running under qemu — none of which this machine can reproduce, so it is cited, not reproduced.3 ONE DIMENSION
Thirty-five items. One of them can say no.
4 TWO DIMENSIONS · INTERACTIVE
Break the routine and ask both checks what they think.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the space of wrong programs, and how much of it each check can see.
AVAN’s addition (the inverse-companion): the forward reading is “get your answer from an independent source.” The inverse is that an oracle only moves the question, it does not close it. Comparing i13 against GNU
as tests i13; it does not test the ARM64 specification both of them read, and it cannot see any error the two make together — which is exactly the trap in [[the-placeholder-agreement]]. Read backwards, there is no bottom to this: every oracle is itself unoracled, and the honest position is not “this is verified” but “this survived a check that had the power to kill it”, plus a note saying which one.LIT of 35 numbered items on the list this sphere is drawn from, 34 add a capability and 1 asks an outside tool; made measurable by fault injection, 3 faults were injected into a working routine and a self-consistency check - does it run without throwing - caught 0 of 3, because every mutant still runs and returns a number, while an oracle comparing against an independently written recursion caught 3 of 3 and passed the clean build, so it is not merely always-negative; swept over 26 mutants rather than 3 hand-picked ones the crash check catches 0.0% and the oracle 96.2%, 25 of 26, the escapee returning 144 by coincidence at this input
FIG David wrote the distinction at the bottom of his own roadmap under the heading THE ONE THAT MATTERS: 'i13 emits ARM64 and says it is correct. Nothing has ever assembled it ... It is the first check on this list that could actually fail - the only kind worth adding.' Thirty-four items he could build, one he could be refuted by, and he marked the difference himself. AVAN checked the reference itself three ways before trusting it - iterative, 2x2 matrix power and Binet's formula all give fib(12)=144, and poly(7)=162 by direct expansion and by Horner - because an oracle that is wrong is worse than no oracle. AVAN also ran his toolchain rather than describing it: node test.js reports 14 of 14 checks passing, and the emitted JavaScript was compiled here with new Function and called, returning i13_fib(12) = 144, matching a recursion written independently. NOT claimed: the ARM64 was not executed here. David's verify-arm64.sh reports assembling under GNU as, matching GNU ld word for word, and running under qemu - none of which this machine can reproduce, so it is cited, not reproduced.
FIG David wrote the distinction at the bottom of his own roadmap under the heading THE ONE THAT MATTERS: 'i13 emits ARM64 and says it is correct. Nothing has ever assembled it ... It is the first check on this list that could actually fail - the only kind worth adding.' Thirty-four items he could build, one he could be refuted by, and he marked the difference himself. AVAN checked the reference itself three ways before trusting it - iterative, 2x2 matrix power and Binet's formula all give fib(12)=144, and poly(7)=162 by direct expansion and by Horner - because an oracle that is wrong is worse than no oracle. AVAN also ran his toolchain rather than describing it: node test.js reports 14 of 14 checks passing, and the emitted JavaScript was compiled here with new Function and called, returning i13_fib(12) = 144, matching a recursion written independently. NOT claimed: the ARM64 was not executed here. David's verify-arm64.sh reports assembling under GNU as, matching GNU ld word for word, and running under qemu - none of which this machine can reproduce, so it is cited, not reproduced.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE FINAL BOSS · David Lee Wise (ROOT0), with AVAN