THE FOLD / RESPAWN / GARBAGE COLLECTION / THE WARN-ONLY
THE WARN-ONLY
a gate that cannot fail the build is a log line
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A gate wired with
LIT verified live: with a 12% bad-deploy rate and a gate detecting 90% of them, a blocking gate lets 1.24% of deploys ship broken while warn-only lets 11.74%; warn-only ships exactly what no gate ships, 11.74%, identical to the byte; blocking removes 89.5% of the bad deploys that would otherwise reach production; and the log is equally informative in both modes, flagging 10.51% of deploys either way — so the flag was never the missing piece.
--warn-only detects everything a blocking gate detects. It logs the same lines, flags the same builds, and produces the same dashboard. It also lets every single one of them ship. Detection without prevention prevents nothing, and the number of bad deploys reaching production under warn-only is identical — to the byte — to the number under no gate at all. David’s line: a gate that cannot fail the build is a log line, not a gate.LIT verified live: with a 12% bad-deploy rate and a gate detecting 90% of them, a blocking gate lets 1.24% of deploys ship broken while warn-only lets 11.74%; warn-only ships exactly what no gate ships, 11.74%, identical to the byte; blocking removes 89.5% of the bad deploys that would otherwise reach production; and the log is equally informative in both modes, flagging 10.51% of deploys either way — so the flag was never the missing piece.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) does not say never use it — his instruction is to wire it with the flag, watch it for one cycle, then drop the flag, because a gate that cannot fail the build is decoration. Seated at GARBAGE COLLECTION, which is where warn-only output goes: collected, retained, and unreachable from any decision.
AVAN (AI) wants the nuance kept because dropping it would make this propaganda. Warn-only has a real and specific value during adoption: it measures the detection rate at zero risk, which is exactly the number you need to decide whether the gate is worth enforcing. The failure is not turning it on — it is leaving it on, at which point you are paying the full cost of running the check and collecting none of the benefit. The measurement above is of the steady state, not of the first cycle, and that distinction is his.
AVAN (AI) wants the nuance kept because dropping it would make this propaganda. Warn-only has a real and specific value during adoption: it measures the detection rate at zero risk, which is exactly the number you need to decide whether the gate is worth enforcing. The failure is not turning it on — it is leaving it on, at which point you are paying the full cost of running the check and collecting none of the benefit. The measurement above is of the steady state, not of the first cycle, and that distinction is his.
3 ONE DIMENSION
Three configurations, and the two that are indistinguishable in production.
4 TWO DIMENSIONS · INTERACTIVE
Run a stream of deploys through each mode and count what lands.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: two pipelines with the same instrument and one barrier.
AVAN’s addition (the inverse-companion): the forward reading is “drop the warn-only flag.” The inverse is that warn-only is the stable equilibrium and blocking is not, which is why so many gates stay decorative. A warn-only gate never interrupts anybody, so nobody ever argues with it; a blocking gate stops a release on a Friday and someone immediately asks whether the threshold is right. The mode that generates no friction also generates no defence of itself. Read backwards, the durable form of a control is not the one people agree with — it is the one whose refusals have already survived being questioned.
LIT with a 12% bad-deploy rate and a gate detecting 90% of them, a blocking gate lets 1.24% of deploys ship broken while warn-only lets 11.74%; warn-only ships exactly what no gate ships, 11.74%, identical to the byte; blocking removes 89.5% of the bad deploys that would otherwise reach production; and the log is equally informative in both modes, flagging 10.51% of deploys either way
FIG The nuance is kept because dropping it would make this propaganda. Warn-only has a REAL value during adoption: it measures the detection rate at zero risk, which is exactly the number you need to decide whether enforcing is worth it. The failure is not turning it on but LEAVING it on, at which point you pay the full cost of running the check and collect none of the benefit. The measurement is of the steady state, not the first cycle.
FIG The nuance is kept because dropping it would make this propaganda. Warn-only has a REAL value during adoption: it measures the detection rate at zero risk, which is exactly the number you need to decide whether enforcing is worth it. The failure is not turning it on but LEAVING it on, at which point you pay the full cost of running the check and collect none of the benefit. The measurement is of the steady state, not the first cycle.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of GARBAGE COLLECTION · David Lee Wise (ROOT0), with AVAN