silence-gauge — the evidential weight of a negative, quantified

surfacing says it in words: a positive is hard to explain away; a negative means almost nothing. This is that sentence turned into arithmetic — so you can see how near-zero a negative is, and break the assumption that makes it so.

It is a companion to the boundary-crossing detectors (forward-observers, surfacing, hearsay): they tell you whether your marker crossed a membrane; the gauge tells you how much to believe a silence when it doesn't.

The model (Bayes on canary extraction)

symbol	meaning	typical
e	emission rate P(surface | ingested) — how often a model that was trained on your canary re-emits it	small for frontier models (verbatim suppression)
fp	false-positive P(surface | not ingested) — chance completion of the 64-bit withheld half of a 128-bit canary	~2^-64 ≈ 0
pi	prior P(ingested)	your call

Bayes factors (how much one observation multiplies your odds):

positive:  BF+ = e / fp             huge, because fp ~ 0  → a surface is decisive
negative:  BF- = (1 - e) / (1 - fp) ~ (1 - e) → with e small this is ~1: silence moves you almost nowhere

The asymmetry is log-scale, not rhetoric. In the frontier regime (e=0.05, 64-bit withheld half, even prior) a positive carries ~60 bits of evidence and drives the posterior to ~1; silence carries less than a tenth of one bit and leaves the posterior sitting on your prior.

The control you can break

The only reason a negative is worthless is that e is small — suppression. Crank e toward 1 (a model that always emits what it trained on) and a negative becomes strong evidence of non-ingestion. That is the honesty of the gauge: it refuses to read silence as "clean," because the reading depends entirely on a suppression rate you cannot measure. Drag the emission slider on the page and watch silence gain meaning — the assumption is right there, breakable.

Verify first

python selftest.py

Proves with no network: a positive drives the posterior to ~1 (~60 bits); under suppression silence leaves the posterior 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 (monotonic), reaching a decisive non-ingestion verdict at e=0.99.

Files

File	Role
silence.py	the Bayes math: bf_positive, bf_negative, posterior, bits, gauge
selftest.py	the four claims, proven with no network
index.html	the live gauge — sliders for e, entropy, prior; the asymmetry drawn

What it is and is not

Is: an honest accounting of how much a canary-extraction result should move your belief, given assumptions you set and can see.

Is not: a way to make a negative mean something it doesn't. Its whole point is that a negative, under real suppression, means almost nothing — and it shows you exactly what would have to be true for that to change.

---

David Lee Wise / ROOT0 / TriPod LLC · CC-BY-ND-4.0