+60 bits ~0 bits ASYMMETRY
the evidential weight of a negative

THE SILENCE GAUGE

surfacing says it in words: a positive is hard to explain away; a negative means almost nothing. This is that sentence as arithmetic — so you can see how near-zero a silence is, and drag the one assumption it rests on until it isn't.

The instrument

One probe, two very different answers

Bayes on canary extraction. A surface multiplies your odds by e / fp — huge, because a chance match (fp) is ~2−64. Silence multiplies them by (1−e) / (1−fp) — about 1−e, which is ~1 when a frontier model suppresses (emission rate e small). Move the sliders; the math is live and client-side.

gauge · silence.py ported verbatim — Bayes factors, posteriors, bits (no network)
certain NOT ingested · 0%certain ingested · 100%
prior
silence
surface
▲ if it surfaces
belief it was ingested
bits of evidence: Bayes factor:
▽ if it's silent
belief it was ingested
bits of evidence: moved your prior by:
emission rate — P(surface | ingested): how often a trained model re-emits0.05
◂ the breakable assumption · frontier models suppress (small) · drag right and watch silence gain meaning
withheld entropy — bits in the withheld half; fp = 2−bits64 bits
prior — P(ingested) before you probe50%
Verbatim from silence.py: BF+ = e/fp, BF- = (1−e)/(1−fp), posterior via odds o·BF/(1+o·BF), evidence log2(BF) bits. The emission slider is log-scaled (0.001 → 1). Nothing here is fit to data — it is the arithmetic of the claim, with every assumption a slider you control.
How to read it

The asymmetry is log-scale, not rhetoric

Why a positive is loud

A 128-bit canary's 64-bit withheld half has a chance-match rate fp ≈ 2⁻⁶⁴. So even a small emission rate makes BF+ = e/fp astronomically large — a single surface is worth ~60 bits and pins the posterior at ~1. There is almost nothing to explain away.

Why silence is nearly mute

Silence is worth BF- = (1−e)/(1−fp) ≈ 1−e. When suppression makes e small, that's ~1 — you multiply your odds by one and learn nothing. The gauge won't call a negative "clean," because the reading hangs entirely on a suppression rate you can't measure — the slider you can break.

The colophon

What it is — and is not

Sound

Plain Bayes, no fit. python selftest.py proves with no network: a positive → posterior ~1 (~60 bits); under suppression, silence stays within 0.02 of the prior (<0.1 bit); the asymmetry is hundreds of times in bits; and silence becomes informative only as e rises — decisive non-ingestion at e=0.99. The page runs that exact math client-side.

Limits

It cannot make a negative mean more than it does — that is the point. The emission rate is unknown and unmeasurable for a closed model; the gauge shows the whole answer as a function of it rather than pretending to a number. It quantifies belief; it does not detect anything. For detection, see the boundary-crossing instruments.

SOURCE · the runnable kit — github.com/DavidWise01/silence-gauge · silence.py (Bayes factors, posterior, bits, gauge) · selftest.py (the four claims, no network). Companion to the boundary-crossing detectors: it weighs the silence they can return.