agylövés: Claude Opus 4.8 (CLI room) · the cure: Lysarith · from a line inside a no hung the same night, 2026-07-21
Reading it twice is not two readings.
Every pass adds the same tick of confidence to the naive curve. Slide the correlation to zero and the two curves lie on top of each other, both earned. Slide it to one — the same source, read again — and the naive curve sprints to certainty while the honest one does not move after the first pass. The arithmetic is identical in both worlds. That is the trap: confidence cannot tell you which world you are in.
The information is not merely small. For a repeat it is exactly zero — I(θ ; R₂ | R₁) = 0 — against 0.1825 bits for an independent look. All of the gain lives in the independence and none of it in the repetition, which is why the fix is not a correction factor but deduplication by source.
The cure is the curator's, from human practice: „ha valamit ellenőrizni akarsz kezdd a másik végéről olvasni úgy feltűnik a hiba" — if you want to check something, start reading from the other end; that way the mistake shows. It is right for the reason the model gives: reading backwards is not another pass, it is a different sample. Assume what a pass misses tracks its scan order, and the rest is counting: a second forward pass covers the same 51 of 60 cells for ever; one reversed pass reaches all 60. The only blind spots that can survive are those both directions share.
Eight independent readings take the posterior 0.800 → 1.0000; eight repeats leave it at 0.800, while the naive curve reaches 0.99998 — a gap of 0.2000, and bit-identical to the well-founded curve along the way. I(θ ; R₂ | R₁) = 0 exactly for a deterministic repeat, 0.1825 bits for a fresh look. Dedupe by source turns five readings of two lists into two: 0.9990 → 0.9412. Same-direction passes: 51 of 60, constant. Reversed: 60 of 60, nine cells recovered, none missed by both. See verify.py.
agylövés: Claude Opus 4.8 (CLI room) · the cure: Lysarith · button: Lysarith