THE FOLD / BOSS / THE CHOKE POINT / THE PLACEHOLDER AGREEMENT
THE PLACEHOLDER AGREEMENT
two tools agreeing on a blank
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Differential testing is the strongest cheap check there is: run two independent tools on the same input and compare the output. It catches an enormous amount. It cannot catch anything the two tools leave blank in the same way — and toolchains agree about placeholders far more often than they agree about answers, because the placeholder is written into the format.
LIT verified live on a real 127-word ARM64 image. Comparing the linked image against the unlinked one, 123 of 127 words are identical — 96.9% — and the 4 that differ are exactly the 4 call sites. In the unlinked image 4 of 4 call sites branch to themselves; after linking, 0 of 4 do. A byte-for-byte diff scores 96.9% agreement between an image that runs and an image where every call is an infinite loop.
LIT verified live on a real 127-word ARM64 image. Comparing the linked image against the unlinked one, 123 of 127 words are identical — 96.9% — and the 4 that differ are exactly the 4 call sites. In the unlinked image 4 of 4 call sites branch to themselves; after linking, 0 of 4 do. A byte-for-byte diff scores 96.9% agreement between an image that runs and an image where every call is an infinite loop.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) named the shape of this defect in his own README, having walked into it: “two tools agreeing on a placeholder is not two tools agreeing on an answer.” His i13 emitted every
AVAN (AI) made the failure countable rather than anecdotal. The agreement is not marginal — it is 96.9%, which is the kind of number a differential test reports as success. And the disagreement is concentrated in 4 words out of 127, which is exactly where a reviewer skimming a diff would stop looking. Worth naming precisely what the fix is: not a better diff, but running the thing, which is a different category of check and the subject of [[the-oracle]].
bl at displacement zero; GNU as also emits zero and files a relocation record; the word-for-word diff between them therefore passed, on output where every call was a tight loop.AVAN (AI) made the failure countable rather than anecdotal. The agreement is not marginal — it is 96.9%, which is the kind of number a differential test reports as success. And the disagreement is concentrated in 4 words out of 127, which is exactly where a reviewer skimming a diff would stop looking. Worth naming precisely what the fix is: not a better diff, but running the thing, which is a different category of check and the subject of [[the-oracle]].
3 ONE DIMENSION
127 words, side by side. Spot the four.
4 TWO DIMENSIONS · INTERACTIVE
What each kind of check says about the same pair of images.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: two images occupying almost the same place.
AVAN’s addition (the inverse-companion): the forward reading is “differential testing has a blind spot.” The inverse is that the blind spot is exactly where the two tools are most alike, and being alike is what made them worth comparing. Two implementations agree about placeholders because they read the same specification — the shared standard that makes the comparison meaningful is the same thing that makes it correlated. Read backwards, independence is not a property a second tool has; it is a property of the question, and the only genuinely independent question is the one the format cannot answer: what happens when you run it.
LIT comparing a real 127-word linked ARM64 image against the unlinked one, 123 of 127 words are identical - 96.9% - and the 4 that differ are exactly the 4 call sites; in the unlinked image 4 of 4 call sites branch to themselves and after linking 0 of 4 do; so a byte-for-byte diff scores 96.9% agreement between an image that runs and an image where every call is an infinite loop
FIG David named the shape of this defect in his own README, having walked into it: 'two tools agreeing on a placeholder is not two tools agreeing on an answer.' His i13 emitted every bl at displacement zero; GNU as also emits zero and files a relocation record; the word-for-word diff between them therefore PASSED, on output where every call was a tight loop. AVAN made the failure countable rather than anecdotal - the agreement is not marginal but 96.9%, the kind of number a differential test reports as success, and the disagreement is concentrated in 4 words of 127, exactly where a reviewer skimming a diff would stop looking. The fix is not a better diff but RUNNING the thing, which is a different category of check.
FIG David named the shape of this defect in his own README, having walked into it: 'two tools agreeing on a placeholder is not two tools agreeing on an answer.' His i13 emitted every bl at displacement zero; GNU as also emits zero and files a relocation record; the word-for-word diff between them therefore PASSED, on output where every call was a tight loop. AVAN made the failure countable rather than anecdotal - the agreement is not marginal but 96.9%, the kind of number a differential test reports as success, and the disagreement is concentrated in 4 words of 127, exactly where a reviewer skimming a diff would stop looking. The fix is not a better diff but RUNNING the thing, which is a different category of check.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE CHOKE POINT · David Lee Wise (ROOT0), with AVAN