Witness 0 directly observes a staged "sighting" — a vector of 10 remembered details (shape, colour, motion, altitude, sound, duration, size, count, speed, glow). Perception itself is imperfect, so even the first account starts a little short of 100% true-to-event.
The story then spreads: at each retelling, the listener copies the teller's current account, and every detail in it has an independent chance — the contamination probability — of drifting to a different, incorrect value. This mirrors Elizabeth Loftus's misinformation-effect research: post-event suggestion, leading questions and exposure to other accounts don't just add noise, they can overwrite the original memory trace itself, and the false detail often feels just as vivid as a true one.
With cross-talk on, witnesses don't just retell down a chain — pairs also compare notes directly, so a false detail introduced anywhere can jump sideways into unrelated witnesses' accounts (social contagion of memory), degrading the whole group's accuracy faster and more evenly than a single chain of retellings would.
accuracy(w) = matches(detail(w), truth) / 10
P(detail flips at a retelling) = contamination probability
drift count += 1 each time a detail actually changes
- Witnesses — how many people the account passes through; more witnesses means more retelling hops per generation.
- Retelling rate — how often accounts are exchanged; faster retelling reaches everyone sooner but doesn't by itself change accuracy per hop.
- Contamination probability — chance each remembered detail is altered at a single retelling; this is the real driver of testimony decay.
- Cross-talk — toggles whether witnesses also swap details with a random peer, not just their chain neighbour, spreading contamination sideways.
Real-world relevance: this is why investigators separate witnesses immediately and avoid showing them each other's statements or media coverage before independent accounts are recorded — every additional exposure is another chance for contamination.