The loop amplifies whatever bias arrived first

What it is

Once the loop is turning, a small advantage that appears first is written by the next round of clicks into a larger advantage. A new user’s first screen happens to lean toward one class; a new item is pushed once from an editorial slot; a time-of-day hit is treated as a stable taste — these are early-bias amplification: a small, early deviation becomes, through display–behaviour coupling, a rank gap that is hard to reverse.

What is amplified is rank and exposure, not necessarily “the user liked it more and more.” Liking is a story added when the log is read later.

Why it happens

Ranking is relative: one extra click raises the next score a little, which raises position a little, which buys another click. Advantage is convex — items already in front collect clicks more cheaply. Noise at t=0 (a mis-tap, one editorial push, a popularity prior at cold start) becomes structure at t=n. Path dependence here is mechanical: the same user, seeded with a different first screen, can have an entirely different stable list weeks later.

The bias is not only on the user side. Items too: a new title that gets a few interactions gets more; a near substitute that missed the first exposure stays cold. The loop amplifies who was seen first, not who most deserved to be seen.

Studying it

Fork the same users: change only the first-screen seed, or only the first-week push of new items, then let the loop run. After several sessions, compare overlap of the two stable lists, category mix, head concentration. Independent variables: direction and size of the initial deviation, whether click write-back is cut mid-way. Dependent variables: how much the later list still remembers the first-day seed (can you recover day one from the later list), share of head items.

A write-back-off control is required, or you cannot separate “the user really turned” from “the loop is remembering day one.” How accidental early clicks freeze into profile features is a finer layer; here the question is only whether rank gaps grow with time.

Where it stops holding

Where items live for hours (alerts, timed inventory), early bias has no time to amplify before the item is gone. A human-updated front page that resets rank daily truncates amplification. When users change interests on purpose and often (a trip ends, a course ends), later behaviour overwrites the early stretch and amplification does not show. This entry argues that a small, first deviation is written into structure by the loop. It does not treat the need for a signal from outside the loop, and it does not restate the display–behaviour coupling itself.

Applying it

• Record, per user, the classes treated as positive in week one, and decay them: down-weight early clicks outside the window so first-week noise is not a lifetime label.
• Give new items an exposure window that does not depend on existing clicks; only then enter the closed loop, so “who was seen first” does not decide who survives.
• Check: randomly assign new users to two first screens that differ only in seed; compare list overlap after four weeks. If overlap stays well below what a random fork should regress to, early bias has been amplified and locked.

Related

• Same group: L6.04.1 What is shown shapes behaviour, which then reshapes what is shown · L6.04.3 Breaking self-reinforcement needs a signal from outside the loop
• Nearby: L6.09 Feedback Loops and Preference Entrenchment · L6.03 Cold Start · L6.08 Filter Bubbles and Diversity
• Search terms: early-bias amplification · path dependence · Matthew effect in ranking