THE FOLD / GLITCH / STACK OVERFLOW / THE OBSERVER THAT MOVED IT
THE OBSERVER THAT MOVED IT
a probe that re-ran what it was watching
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A sensitivity probe checks whether an assertion actually depends on the thing it claims to test: mutate the subject one field at a time and see whether the answer moves. To do that it must re-execute the assertion, once per field. If the assertion is not pure — if running it changes anything — the probe is no longer measuring the system, it is driving it.
LIT verified live. With 8 mutable fields, a probe re-running an impure thunk turns 11 intended calls into 99 — 88 extra invocations, exactly one per field per call. Three later assertions that read the counter then break: 3 of 3 hold without the probe, 2 of 3 fail with it. Restricting the probe to thunks the caller declares pure leaves the count at 11.
LIT verified live. With 8 mutable fields, a probe re-running an impure thunk turns 11 intended calls into 99 — 88 extra invocations, exactly one per field per call. Three later assertions that read the counter then break: 3 of 3 hold without the probe, 2 of 3 fail with it. Restricting the probe to thunks the caller declares pure leaves the count at 11.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) caught his own observer wrecking the run it was watching and wrote it up under a heading that says so: “CORTEX BROKE THE RUN IT WAS WATCHING… the probe fired it 88 extra times: subject calls 11 -> 99, and three LATER assertions failed because the observer had moved the state they were checking. an observer that changes what it observes is worse than none.” The repair is a refusal: sensitivity is now probed only on thunks declared pure, and everything else reports unprobed rather than passed.
AVAN (AI) reproduced the arithmetic exactly — 11 becomes 99 with eight fields, because each call is re-run once per field. Worth naming what the repair costs: the probe now says nothing at all about the majority of assertions, and unprobed is a much weaker report than passed. That weakness is the point. A probe that declines to answer is more useful than one that answers by changing the question.
AVAN (AI) reproduced the arithmetic exactly — 11 becomes 99 with eight fields, because each call is re-run once per field. Worth naming what the repair costs: the probe now says nothing at all about the majority of assertions, and unprobed is a much weaker report than passed. That weakness is the point. A probe that declines to answer is more useful than one that answers by changing the question.
3 ONE DIMENSION
Intended calls, and what the probe made of them.
4 TWO DIMENSIONS · INTERACTIVE
Turn the probe on and watch the later assertions go red.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: one intended call, and the fan of re-runs behind it.
AVAN’s addition (the inverse-companion): the forward reading is “do not let the probe disturb the subject.” The inverse is that the only probes that cannot disturb anything are the ones that cannot see very much. Sensitivity is a causal question — does the answer depend on this input? — and causal questions are answered by intervening. Read backwards, purity is not a safety property the probe happens to require; it is the precondition for asking a causal question without paying for it, and everything impure is unprobed forever, not merely for now.
LIT with 8 mutable fields a probe re-running an impure thunk turns 11 intended calls into 99 - 88 extra invocations, exactly one per field per call; three later assertions that read the counter then break, holding 3 of 3 without the probe and failing 2 of 3 with it; and restricting the probe to thunks the caller declares pure leaves the count at 11
FIG From David's FREEZE round 2, 2026-08-05. He caught his own observer wrecking the run it was watching and wrote it up under a heading that says so: 'CORTEX BROKE THE RUN IT WAS WATCHING ... the probe fired it 88 extra times: subject calls 11 -> 99, and three LATER assertions failed because the observer had moved the state they were checking. an observer that changes what it observes is worse than none.' The repair is a refusal - sensitivity is probed ONLY on thunks declared pure, everything else reports unprobed rather than passed. AVAN reproduced the arithmetic exactly and names what the repair costs: the probe now says nothing about the majority of assertions, and 'unprobed' is a much weaker report than 'passed'. That weakness is the point.
FIG From David's FREEZE round 2, 2026-08-05. He caught his own observer wrecking the run it was watching and wrote it up under a heading that says so: 'CORTEX BROKE THE RUN IT WAS WATCHING ... the probe fired it 88 extra times: subject calls 11 -> 99, and three LATER assertions failed because the observer had moved the state they were checking. an observer that changes what it observes is worse than none.' The repair is a refusal - sensitivity is probed ONLY on thunks declared pure, everything else reports unprobed rather than passed. AVAN reproduced the arithmetic exactly and names what the repair costs: the probe now says nothing about the majority of assertions, and 'unprobed' is a much weaker report than 'passed'. That weakness is the point.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of STACK OVERFLOW · David Lee Wise (ROOT0), with AVAN