05 — Privacy
A heartbeat over weeks can become a behavioral biometric.
Longitudinal behavioral traces can be identifying. Saying “we don’t know your name” does not save you if the vector re-identifies you across devices. Purpose-bind them or do not collect them.
First principle
Network pulse over account EKG. Use the least individual information necessary to solve the network problem. Event-level aggregates catch coordinated injection with less identity than a weeks-long personal trace. Account-level state exists because ranking still has to score authors. It should be the narrower tool.
Two different products
Proof of personhood is a rare check that a unique living body is attached to an account. Useful against mass creation. Silent about whether this reply was written by an agent.
Liveness of interaction is the heartbeat: continuous, probabilistic, private, used to weight the feed rather than to ban. X’s crude version is “if a human is not tapping the screen.” The richer version is ranking, not a CAPTCHA.
Keep them beside each other. Do not collapse them into one passport.
Rules
- No user-visible score. No “liveness 0.91.” Operators will train on it. Users will be shamed by it.
- Store factors, not raw motor traces. Timing histograms and relationship sketches. Not keystroke streams, if you can avoid them.
- Network pulse is the privileged object. Event-level aggregates need less identity than account-level EKGs.
- Local / noisy where possible. Client dwell and scroll can become on-device sketches. Account-level differential privacy will blunt detection; event-level aggregates tolerate it better.
- Purpose bind. Reweight distribution. Catch coordinated inauthentic behavior. Do not leak into ads, dating, insurance, or employability.
- Appeal without a passport. Low Lu is a prior, not a verdict. Night shift, disability, scheduled posts, shared accounts, newsrooms, customer service, and translators all look “wrong.”
Transparency without leaking the detector
A person whose reach is affected still needs to know something happened.
Do not reveal: exact L, factor weights, the coupling estimator, thresholds, the feature recipe, CI, or σ_L.
Can reveal: “Some engagement associated with this post was discounted because it appeared coordinated.” “Automated activity was detected around this post.” “This did not penalize your account.” “Request review.”
That protects the detector without making ranking occult. Details of the appeal path belong next to the experiment contract, not in a public formula.