THE FOLD / GLITCH / RACE CONDITION / THE SHALLOW CLONE
THE SHALLOW CLONE
a truncation that reports as a count
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A gate that re-derives published numbers has to clone the source repository first, and a shallow clone does not error — it returns a smaller number of the right type. Ask a depth-50 clone how many commits exist and it says 50, confidently. David’s CI file marks
LIT verified live: a clone at depth d reports min(d, true) commits — 1, 50, 500, 1000, and only a full clone returns the real 2,070; against a full clone the gate correctly fails, published 1,724 against 2,070, a drift of 16.71% matching his receipts exactly; a clone truncated to the published figure reports 0.00% drift and the gate passes; and nothing errors or warns at any depth.
fetch-depth: 0 as REQUIRED for exactly this reason. The worst case is not a wrong answer but a flattering one: truncate the clone to the published figure and the gate reports zero drift and passes.LIT verified live: a clone at depth d reports min(d, true) commits — 1, 50, 500, 1000, and only a full clone returns the real 2,070; against a full clone the gate correctly fails, published 1,724 against 2,070, a drift of 16.71% matching his receipts exactly; a clone truncated to the published figure reports 0.00% drift and the gate passes; and nothing errors or warns at any depth.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) put the warning inline in ci/seam-gate.yml rather than in prose, where it is read at the moment it matters. The same file carries a second discipline worth as much: adopt with
AVAN (AI) built the pathological case rather than describing it, because that is the part that carries. A shallow clone is usually discussed as a source of wrong answers; the sharper problem is that its error runs in the direction that silences the gate. A truncated history under-counts, an under-count moves the clone figure toward a stale published one, and the drift shrinks. The failure mode is not random — it is biased toward agreement, which is precisely the bias a gate cannot afford, and it produces no diagnostic of any kind.
--warn-only so the build does not break on day one, then remove it once the first regeneration lands, or the gate is decoration. Seated at RACE CONDITION — two views of one repository, and no ordering that guarantees the gate sees the whole of it.AVAN (AI) built the pathological case rather than describing it, because that is the part that carries. A shallow clone is usually discussed as a source of wrong answers; the sharper problem is that its error runs in the direction that silences the gate. A truncated history under-counts, an under-count moves the clone figure toward a stale published one, and the drift shrinks. The failure mode is not random — it is biased toward agreement, which is precisely the bias a gate cannot afford, and it produces no diagnostic of any kind.
3 ONE DIMENSION
Clone depth against reported count. The line stops being a measurement.
4 TWO DIMENSIONS · INTERACTIVE
Set the clone depth and watch the gate change its mind about the same repository.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: a history seen twice, once entire and once cut off.
AVAN’s addition (the inverse-companion): the forward reading is “fetch the whole history.” The inverse is that a truncation is the most dangerous kind of error because it is well-formed. A clone that failed would stop the build; a clone that returns 50 returns an integer, in range, of the right type, and every downstream check treats it as a measurement. Read backwards, this is an argument for gates that verify their own inputs before verifying anything else — a probe that never asks whether it received a complete history is not measuring the repository, it is measuring whatever it happened to be given.
LIT a clone at depth d reports min(d, true) commits — 1, 50, 500, 1000 — and only a full clone returns the real 2,070; against a full clone the gate correctly fails, published 1,724 against 2,070, a drift of 16.71% matching his receipts exactly; a clone truncated to the published figure reports 0.00% drift and the gate PASSES; and nothing errors or warns at any depth
FIG A shallow clone is usually discussed as a source of WRONG answers; the sharper problem is that its error is BIASED toward agreement. A truncated history under-counts, an under-count moves the clone figure toward a stale published one, and the drift shrinks — which is precisely the bias a gate cannot afford, and it produces no diagnostic of any kind. The same CI file carries a second discipline: adopt with --warn-only, then remove it once the first regeneration lands, 'or the gate is decoration'.
FIG A shallow clone is usually discussed as a source of WRONG answers; the sharper problem is that its error is BIASED toward agreement. A truncated history under-counts, an under-count moves the clone figure toward a stale published one, and the drift shrinks — which is precisely the bias a gate cannot afford, and it produces no diagnostic of any kind. The same CI file carries a second discipline: adopt with --warn-only, then remove it once the first regeneration lands, 'or the gate is decoration'.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of RACE CONDITION · David Lee Wise (ROOT0), with AVAN