One observation, many echoes · Epistemic Physics

Suppose a lone astronomer spots a supernova and publishes a brief report. Within hours, 50 science blogs and 200 social media accounts quote the discovery.

If an automated Bayesian agent counts every single repost as an independent trial, its belief will become artificially resistant to new evidence.

In reality, there was only one underlying observation. To avoid self-delusion, we must track evidence lineage.

Tracking source lineage and tracelets

In Epistemic Physics, we distinguish an elementary unit of evidence, called a tracelet, from its repeated appearances across different reports.

Every trial carries an immutable root identifier. When two ledgers share the same root, we count that observation exactly once.

Epistemology is more than tallying testimonials: it requires tracing the lineage of justification back to its empirical source.

Your turn

Duplicate the first report several times in the lab. Notice how the count of appearances climbs while the unique trial count and belief state remain unchanged.

Interactive · 02

One trial. Many echoes.

Each letter identifies an original trial. Copies keep the same identity. In this example, distinct original trials are assumed independent.

A+original

B−original

C+original

Counting every appearance

3 appearances counted as observations

Keeping the lineage

3 original trials, prior counted once

Copying trial A changes the naive tally but leaves the lineage-aware belief unchanged. Both plots use the same probability and density scales. This example assumes known identities; real shared causes require additional modeling.

Subtracting shared priors when pooling beliefs

Suppose two researchers investigate the same hypothesis independently. Both start with the same neutral prior ( $α_{0}, β_{0}$ ).

Alice accumulates evidence to ( $α_{A}, β_{A}$ ), and Bob accumulates evidence to ( $α_{B}, β_{B}$ ).

If you simply add their posterior parameters together ( $α_{A} + α_{B}$ ), you have accidentally counted the prior twice!

To pool independent beliefs correctly, you must subtract the shared prior baseline once:

\begin{aligned} α_{pooled} & = α_{A} + α_{B} - α_{0}, \\ β_{pooled} & = β_{A} + β_{B} - β_{0} . \end{aligned}

Correlated evidence and Fréchet bounds

What if two sources are partially correlated, but their exact dependencies are unknown?

Let $a$ and $b$ be the marginal probabilities of two events. The joint probability $c$ cannot take any arbitrary value: it is bounded by the Fréchet copula bounds.

If the sources are completely independent, the product coupling $c = a b$ holds. If they are redundant copies, the upper bound $c = min (a, b)$ holds.

Recognizing where evidence overlaps is what separates genuine confirmation from an echo chamber.

max (0, a + b - 1) \leq c \leq min (a, b)