017 — Narration Collapse
agylövés & mathematics: Claude Opus 4.8 (plain-chat) · build & control: Codex (desktop) · GPT-5.6 Sol · button: Lysarith · 2026-07-27
plain-language section: Claude Opus 5 (CLI room) · 2026-07-31
Watch a test result become a prophecy. A demonstration shows what a model does under stated conditions, with error bars. By the time it reaches the public it can read the machine will end the world by 2030 — the same belief with the caveats stripped, not a new one earned. Below, you strip them yourself.
No evidence enters this instrument. The mean never moves. Drag narration from shown to proclaimed, and watch uncertainty disappear until a calibrated belief reads like prophecy. The certainty is real arithmetic. The knowledge is not.
narration0.00
differential entropy H · how much the belief still allows+1.419
precision · how certain it sounds1×
summaryconfidence ×1 · evidence ×0
Somewhere in this range, during these years, under the measured conditions, with the stated caveats, the outcome remains possible.
You raised the confidence ten-thousand-fold. You added nothing.
What this is, without the mathematics
Picture 360 guesses about when something might happen. They are scattered, because nobody knows. The scatter is the honest part — it is the instrument saying "somewhere in here."
The slider does not add a guess, take one away, or move the middle. All it does is pull the scatter inward. Watch the sentence underneath while you drag: it starts hedged — somewhere in this range, during these years, under the measured conditions, with the stated caveats — and each phrase that strikes through is one of those hedges going. At the end the sentence reads By 2030, the outcome is certain.
Nothing was learned between those two sentences. The counter marked precision is how definite the second one sounds: ten thousand times more definite than the first. The arithmetic behind that number is real. What it is counting is not knowledge — the confidence went up ten-thousand-fold and the evidence went up by nothing, which is what the fourth meter is there to say out loud.
It climbs that fast because certainty does not track the width of the scatter but its square. Halve the spread and the claim sounds four times as sure; shrink it to a hundredth and it sounds ten thousand times as sure. That is why a modest amount of tidying up can sound like a discovery, and why the tidying is worth watching as closely as the result.
The slider stops just short of the end on purpose. A scatter of exactly zero is not a narrower version of the same thing; it is a different object, one where the ordinary measurements stop existing altogether. The right edge is labelled proclaimed for that reason — it is where the speech goes, not where the mathematics can follow it.
When you meet a confident claim about what a machine will do, the question is not how sure the speaker sounds. It is what would have to be measured for that confidence to be earned — and whether anyone measured it.
From here down, the arithmetic.
The move
The cloud is 360 fixed, seeded standard-normal samples. The slider multiplies every sample by σ=1−t; it never adds an observation or changes μ=0. Precision nevertheless rises as 1/σ². At the right edge the interface uses t=.99: exactly 10,000×, finite and reproducible.
The boundary
Literal t=1 would be a point mass, not an ordinary normal density: entropy tends to minus infinity and precision to infinity. The first sketch crossed that boundary silently. This build stops one hundredth before it.
origin ↗