THE FOLD / RESPAWN / HARD RESET / THE CLAIMLINK
THE CLAIMLINK
a verdict that admits it cannot tell
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A seam gate checks whether published numbers came from the repository. claimlink checks whether published sentences survive contact with it — and its design turns on refusing to answer where it cannot. Four verdicts come back: SUPPORTED, REFUTED, SUPPORTED* (true of what ships, false of the repository, both printed), and UNDECIDABLE. That last one is not a gap in the tool. A checker forced to answer yes-or-no on a claim about experience, aesthetics or wall-clock time is not being rigorous; it is inventing an answer, and its rate of doing so is fixed by the claim set rather than by how good its predicates are.
LIT verified live: over 20,000 claims of which 25% cannot be settled by reading code, a checker forced to answer SUPPORTED or REFUTED makes an unfounded assertion on 24.9% of its output — exactly the undecidable fraction, and no improvement to its predicates lowers it; the four-valued checker is wrong on zero of the claims it does answer; it asserts 15,016 times against the binary checker’s 20,000; and on a real claim table the third bucket is 38% of it (3 of 8), so a two-verdict linker would have silently invented 3 answers.
LIT verified live: over 20,000 claims of which 25% cannot be settled by reading code, a checker forced to answer SUPPORTED or REFUTED makes an unfounded assertion on 24.9% of its output — exactly the undecidable fraction, and no improvement to its predicates lowers it; the four-valued checker is wrong on zero of the claims it does answer; it asserts 15,016 times against the binary checker’s 20,000; and on a real claim table the third bucket is 38% of it (3 of 8), so a two-verdict linker would have silently invented 3 answers.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) wrote claimlink.py on 2026-08-04 and put the limitation in the docstring before the feature list: the tool does not read prose and decide what it means. A human writes each claim beside an executable predicate; the machine re-runs it. Seated at HARD RESET, because every run starts from the checkout with no memory of the last verdict.
AVAN (AI) found the measurable version of his sentence — a linker that only emits the first two is lying about its own reach — and it is sharper than a slogan. The unfounded-assertion rate of a binary checker equals the undecidable fraction exactly. It is not an accuracy problem and cannot be engineered away, because the claims in that bucket have no code-shaped evidence in either direction. Better predicates move nothing. The only lever is the verdict set. That is a rare shape for a quality problem: usually you can grind at it, and here grinding is precisely what does not work.
AVAN (AI) found the measurable version of his sentence — a linker that only emits the first two is lying about its own reach — and it is sharper than a slogan. The unfounded-assertion rate of a binary checker equals the undecidable fraction exactly. It is not an accuracy problem and cannot be engineered away, because the claims in that bucket have no code-shaped evidence in either direction. Better predicates move nothing. The only lever is the verdict set. That is a rare shape for a quality problem: usually you can grind at it, and here grinding is precisely what does not work.
3 ONE DIMENSION
A real claim table, four verdicts deep. The shaded rows are what the linker caught.
4 TWO DIMENSIONS · INTERACTIVE
Force the checker to two verdicts and watch the invented answers appear.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the claim space, with the bucket no predicate reaches.
AVAN’s addition (the inverse-companion): the forward reading is “report honestly when you cannot tell.” The inverse is that the third verdict is the only one carrying information about the tool. SUPPORTED and REFUTED describe the repository; UNDECIDABLE describes the instrument — it is the checker stating its own boundary in the same breath as its findings, in a format that can be counted. A tool without that verdict has no channel through which to report its own limits, so its silence about them is structural rather than chosen. Read backwards, adding the third verdict is what converts a checker from an oracle into a measuring device with a stated range.
LIT over 20,000 claims of which 25% cannot be settled by reading code, a checker forced to answer SUPPORTED or REFUTED makes an unfounded assertion on 24.9% of its output — exactly the undecidable fraction, and no improvement to its predicates lowers it; the four-valued checker is wrong on zero of the claims it does answer; it asserts 15,016 times against the binary checker's 20,000; and on a real claim table the third bucket is 38% of it (3 of 8), so a two-verdict linker would have silently invented 3 answers
FIG The measurable version of David's line is sharper than the slogan: the unfounded-assertion rate of a binary checker EQUALS the undecidable fraction exactly. It is not an accuracy problem and cannot be engineered away, because those claims have no code-shaped evidence in either direction. Better predicates move nothing; the only lever is the verdict set — a rare shape for a quality problem, since grinding at it is precisely what does not work.
FIG The measurable version of David's line is sharper than the slogan: the unfounded-assertion rate of a binary checker EQUALS the undecidable fraction exactly. It is not an accuracy problem and cannot be engineered away, because those claims have no code-shaped evidence in either direction. Better predicates move nothing; the only lever is the verdict set — a rare shape for a quality problem, since grinding at it is precisely what does not work.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of HARD RESET · David Lee Wise (ROOT0), with AVAN