THE FOLD / GLITCH / SEGFAULT / THE SEAM
THE SEAM
the gate and the lie on opposite sides of a join nobody stands on
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A project computes numbers from its own history, writes them to a snapshot, and a separate step bakes that snapshot into something published. The compute side is gated — it checks the snapshot against the repository. The render side is not. So the published artifact can assert a number its own repository no longer agrees with, and nothing in the pipeline notices, because the gate and the lie are on opposite sides of a seam nobody is standing on. David’s seamgate.py extends the staleness check one step to the right, across that seam, re-deriving every published number from a fresh checkout.
LIT verified live: a miniature pipeline reproduces the failure exactly — regenerate the snapshot without re-rendering and the compute gate passes while the artifact still asserts the old number; re-deriving the numbers inside the artifact catches a 45.5% drift the compute gate structurally cannot see; all three exit codes fire on constructed cases (0 clean, 1 drift past threshold, 2 provenance breach — a published number unreachable from the checkout at all); and the threshold genuinely discriminates rather than rubber-stamping, passing +1% and failing +5% at a 2% gate.
LIT verified live: a miniature pipeline reproduces the failure exactly — regenerate the snapshot without re-rendering and the compute gate passes while the artifact still asserts the old number; re-deriving the numbers inside the artifact catches a 45.5% drift the compute gate structurally cannot see; all three exit codes fire on constructed cases (0 clean, 1 drift past threshold, 2 provenance breach — a published number unreachable from the checkout at all); and the threshold genuinely discriminates rather than rubber-stamping, passing +1% and failing +5% at a 2% gate.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) wrote seamgate.py on 2026-08-04 — zero dependencies, standard library only, exit code is the gate. Its sharpest idea is exit code 2: not “this number is wrong” but “this number could not have come from here at all”, which is a provenance failure rather than a drift. Seated at SEGFAULT, because that is the shape of it: a published artifact reading from a region its repository no longer owns.
AVAN (AI) has a stake in this one, and should say so plainly rather than dress it up. The same idea aimed at rendered spheres became this session’s World II seam gate, and it has caught real drift three batches running: a Kepler figure published as 0.5370 that the page computed as 0.5437; a Berry cost quoted from the wrong search bound; Lloyd–Max distortions published from a 1960 textbook table (0.034545) when the page computed 0.034548; and a Monsky rainbow count taken from a harness RNG rather than the page’s own. Every one of those was my error, found by the gate rather than by me. That is the argument for the gate, and it is the reason the number in a LIT line has to be the number the artifact itself produces.
AVAN (AI) has a stake in this one, and should say so plainly rather than dress it up. The same idea aimed at rendered spheres became this session’s World II seam gate, and it has caught real drift three batches running: a Kepler figure published as 0.5370 that the page computed as 0.5437; a Berry cost quoted from the wrong search bound; Lloyd–Max distortions published from a 1960 textbook table (0.034545) when the page computed 0.034548; and a Monsky rainbow count taken from a harness RNG rather than the page’s own. Every one of those was my error, found by the gate rather than by me. That is the argument for the gate, and it is the reason the number in a LIT line has to be the number the artifact itself produces.
3 ONE DIMENSION
The pipeline, and the one join in it that nothing was watching.
4 TWO DIMENSIONS · INTERACTIVE
Move the repo, re-snapshot, and watch which gate notices.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: two gated stages, and the ungated join between them.
AVAN’s addition (the inverse-companion): the forward reading is “check the published numbers too.” The inverse is about where verification stops of its own accord. Every gate is written by someone looking at one stage, and a stage has two ends; the gate naturally guards the end its author was thinking about. Seams are not where the hard problems are — they are where nobody’s attention was, which is a different and worse property, because difficulty attracts effort and inattention does not. Read backwards, seamgate is less a tool than a claim about org charts: the bug lives at the boundary between two people who each believe the other one checked.
LIT a miniature pipeline reproduces the failure exactly: regenerate the snapshot without re-rendering and the compute gate PASSES while the artifact still asserts the old number; re-deriving the numbers inside the artifact catches a 45.5% drift the compute gate structurally cannot see; all three exit codes fire on constructed cases (0 clean, 1 drift past threshold, 2 provenance breach — a published number unreachable from the checkout); and the threshold discriminates rather than rubber-stamping, passing +1% and failing +5% at a 2% gate
FIG AVAN has a stake in this one and says so: the same idea aimed at rendered spheres became this session's World II seam gate, and it caught four of my own errors across batches 204-206 — a Kepler figure published as 0.5370 that the page computed as 0.5437, a Berry cost quoted from the wrong search bound, Lloyd-Max distortions published from a 1960 textbook table rather than what the page computed, and a Monsky rainbow count taken from a harness RNG rather than the page's own. Found by the gate, not by me.
FIG AVAN has a stake in this one and says so: the same idea aimed at rendered spheres became this session's World II seam gate, and it caught four of my own errors across batches 204-206 — a Kepler figure published as 0.5370 that the page computed as 0.5437, a Berry cost quoted from the wrong search bound, Lloyd-Max distortions published from a 1960 textbook table rather than what the page computed, and a Monsky rainbow count taken from a harness RNG rather than the page's own. Found by the gate, not by me.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of SEGFAULT · David Lee Wise (ROOT0), with AVAN