THE FOLD / RESPAWN / ROLLBACK / THE DIRECTION TEST
THE DIRECTION TEST
a sign pattern that acquits
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
An audit found every published number smaller than the truth. Every one. That pattern is not a second accusation on top of the first — it is a defence. Somebody inflating their own figures produces errors in the flattering direction; somebody whose page simply stopped being rebuilt produces errors that all point the same way, downward, because the metrics involved only ever grow. The sign of the drift carries information the magnitude does not, and here it acquits.
LIT verified live: of the 3 reproducible metrics every drift points the same way — test_files 6.62%, commits 16.71%, loc 8.50% — and every one is the page understating itself; under a fair-coin sign model the chance of all three landing together is 2/2³ = 25.0%; a 400,000-run simulation returns 25.1%, matching; and the mechanism is forced, since monotone metrics only grow, so a page frozen in the past can only understate them.
LIT verified live: of the 3 reproducible metrics every drift points the same way — test_files 6.62%, commits 16.71%, loc 8.50% — and every one is the page understating itself; under a fair-coin sign model the chance of all three landing together is 2/2³ = 25.0%; a 400,000-run simulation returns 25.1%, matching; and the mechanism is forced, since monotone metrics only grow, so a page frozen in the past can only understate them.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) built the direction test into W3 of the audit instrument and wrote the conclusion against his own interest as an auditor: if someone were exaggerating, you’d expect the opposite. This is staleness, not bragging. Seated at ROLLBACK: the page is a rolled-back copy of the repository, and every bar measures how far back.
AVAN (AI) should be honest about how strong this evidence actually is, because the instinct is to overstate it. Three same-signed drifts have a 25% chance of occurring by coincidence under a coin-flip model — that is suggestive, not conclusive, and a fourth or fifth metric would matter more than any rhetoric about it. What makes the argument work is not the probability but the mechanism: these metrics are monotone, so staleness cannot produce an overstatement, and a single bar pointing the other way would have falsified the whole reading instantly. That is the useful shape — a claim that could have died and did not.
AVAN (AI) should be honest about how strong this evidence actually is, because the instinct is to overstate it. Three same-signed drifts have a 25% chance of occurring by coincidence under a coin-flip model — that is suggestive, not conclusive, and a fourth or fifth metric would matter more than any rhetoric about it. What makes the argument work is not the probability but the mechanism: these metrics are monotone, so staleness cannot produce an overstatement, and a single bar pointing the other way would have falsified the whole reading instantly. That is the useful shape — a claim that could have died and did not.
3 ONE DIMENSION
Every bar on one side of the line. That is the finding.
4 TWO DIMENSIONS · INTERACTIVE
Flip a bar and watch the reading change from staleness to something else.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: errors in a half-space, and the half they never enter.
AVAN’s addition (the inverse-companion): the forward reading is “all the errors lean one way, so it is staleness.” The inverse is that an audit which cannot exonerate is not an audit. A procedure whose every possible output is a finding against the subject is measuring the auditor’s framing, not the subject’s work — and the direction test is valuable precisely because it had a live way to come out the other way and report bragging instead. Read backwards, the test to apply to any critical instrument is: what result would have cleared them? If there is no such result available, nothing the instrument returns is evidence about anything.
LIT of the 3 reproducible metrics every drift points the same way — test_files 6.62%, commits 16.71%, loc 8.50% — and every one is the page understating itself; under a fair-coin sign model the chance of all three landing together is 2/2^3 = 25.0%; a 400,000-run simulation returns 25.1%, matching; and the mechanism is forced, since monotone metrics only grow, so a page frozen in the past can only understate them
FIG The instinct is to overstate this. Three same-signed drifts have a 25% chance of occurring by coincidence under a coin-flip model — suggestive, not conclusive, and a fourth metric would matter more than any rhetoric. What makes the argument work is not the probability but the MECHANISM: these metrics are monotone, so staleness cannot produce an overstatement, and a single bar pointing the other way would have falsified the reading instantly. A claim that could have died and did not.
FIG The instinct is to overstate this. Three same-signed drifts have a 25% chance of occurring by coincidence under a coin-flip model — suggestive, not conclusive, and a fourth metric would matter more than any rhetoric. What makes the argument work is not the probability but the MECHANISM: these metrics are monotone, so staleness cannot produce an overstatement, and a single bar pointing the other way would have falsified the reading instantly. A claim that could have died and did not.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of ROLLBACK · David Lee Wise (ROOT0), with AVAN