Proof

Under fault injection, naive retry produced 666 duplicate side effects and sluice produced none, while refusing one call in five as indeterminate rather than guessing; the landing page serves those figures from the committed results file.

A privacy law that is not tested is a promise. Planting a token-shaped string in a gate answer proved it never reaches events.jsonl, and removing the redaction proved the test fails when it should.

Its first real run across seven repositories reported ten failures and every one was a false positive; after the four fixes it reports zero failures on the same repositories, and the unverifiable counts stayed high because most of what a document claims cannot be checked offline.

proofpage's own README opened with a command that fails on a fresh install, and 51 passing tests never noticed, because every one of them ran inside the repo where the missing directory exists.

The guard failed open on Linux until its self-test ran on a second OS: unconditional case-folding and Windows-only fixtures let every cross-boundary write through, and a two-OS CI matrix now catches exactly that.

Run on its own author's context file, chaff reports zero enforced findings, and that run is a committed eval rather than a demo; it ships zero measured rules today and says so in every line of output, because the evals that would admit one need paid model calls.

kedge's linearizability checker agrees with Porcupine's published verdict on every one of the 102 vendored Jepsen etcd histories that carry one, and it says so from a corpus command anyone can re-run rather than from a README sentence.

A watermark checker raised 124 findings against a correct engine on real streams and was wrong 124 times, and the project keeps that count reconstructible under test instead of quietly zeroing it.

Real process kills on NTFS found zero corruptions in a store that the modeled device shows corrupting at 32 crash points, which is the measured reason crash-consistency testing needs a device model and not a kill loop.

106,108 of the 230,059 events sit at exactly 10, 33 or 35 km because a locator assigned the depth instead of solving for one, and every one of them falls in the shallow band, none at 70 km or deeper. They are drawn as their own class rather than mixed into the picture.

A blind local-minimum search over the semi-major-axis histogram, given no knowledge of resonances, located five gaps whose worst disagreement with independently computed Jupiter resonance positions is 0.0149 AU, about three histogram bins.

A tool description edit is schema-valid, passes every validator, and still changes which tool a model picks; snapgauge grades it risky rather than cosmetic, and the golden set pins that grading at exact match.

A valid C2PA signature on an obviously fabricated scene validates cleanly, because a signature certifies who signed which bytes and says nothing about whether the scene happened; the fixture that proves it is committed with its hash.

Its scheduled workflow had failed silently since birth; since the 2026-08-15 fix the nightly watch has published 13 scheduled run records to the repository as committed JSON, next to the two manual runs that came before.

The instrument was calibrated before any real reading: 0 false positives in 200 null trials and 190 of 200 planted regressions caught, from a seeded simulation; the three scheduled runs since have each published a skipped record because no API key was ever set.

The bytes a browser model actually transfers were 48 to 74 percent smaller than the size its README advertises, across 19 of 20 measured rows on one machine, with the execution provider read back from real GPU submits rather than trusted from config.

Fed four shipped sibling brands, assay regenerated three byte-identical and refused the fourth: shipgauge's live accent-on-text contrast is 1.85:1, under the 3:1 gate, so no artifacts were written, and a test asserts that exact refusal on every run.

One real release video is committed: dogwatch's own five-commit CI-fix range rendered to 36 seconds of h264 with every on-screen number labeled by its source, from a build that fails on any number whose source it cannot name.

Unicode's word segmenter splits nag-book into three pieces, so intra-word Taglish switches never reach a labeler by default; the boundary score here is a consistency check against machine-drafted labels, flagged draft-automated, with no human annotation pass yet.

95.7% top-1 on 281 held-out samples, and the report marks itself provisional because the split is file-level from one source; the project's own ship bar, 70% on a signer it never saw, is not met and cannot be without volunteers.

The default embedding model scores a flat contradiction (0.8645) as more similar than a genuine paraphrase (0.7481) of the same sentence; the page shows both numbers, and CI re-measures every pinned pair against live inference.

On its own showcase dataset the physics layout scores 0.726 silhouette against 0.672 for the best two-axis view, and the tool reports that as no meaningful gain rather than a win; both figures are pinned in the golden set.

11 of 18 Filipino-accented clips had at least one Whisper error, but a same-day US-accented control group failed the same proper nouns the same way, so only one pattern survived (track heard as truck), and a test fails the build if the app ever attributes a mismatch to accent.

On its first real site it reported 40 unlisted pages and every one was a false positive: 35 canonicalised duplicates and 5 machine endpoints. A tool that buries one real finding under 40 false ones is worse than no tool.

Its first real scan produced a block that the scan itself had caused: one host returned 429 at a 250ms delay and was completely clean at 2000ms. Rate limiting is now reported as rate limiting, never as a block.

Reading the ecosystem's own reciprocity rule against its own charters found eight members that no other charter names in either direction, which the constitution calls a violation and its groomer had not caught; reading the dispatch log then found that 33 of the 59 agents have never been dispatched at all.

The privacy claim is enforced by a checker that scans the shipped site, and every one of its rules has a fixture proving the rule fails when violated.

Its first real target was itself, and two rules fired on non-defects: docs/SPEC.md was reported as a leaked test because spec is two words, and a private package was told to ship a LICENSE npm would never publish. Both were narrowed, with a test holding each direction so the rule did not simply switch off.

The model never produces a number: typed tools compute each figure, every figure cites its evidence, and a verifier recomputes it before it renders; the eval run recomputed 51 of 51 cited figures and caught the one fabricated figure planted to test the gate.