Contextual irregularity, not a bot badge.

The original thought

Add a heartbeat to ranking so the network can tell who is a real human and who is not — not from a single CAPTCHA, but from how living, irregular interaction differs from mechanical reaction. Measure humans reacting to the algorithm and to other humans, and bots reacting to bots.

That split is better than the usual binary. Humans fight the feed, misunderstand it, over-post after a hit, disappear after a flop. Bots treat the ranking function as an objective and stay on-policy. Human–human edges are messy: delayed, incomplete, unequal, sometimes repaired later. Bot–bot graphs are dense, reciprocal, and low-entropy. The conversation is not a conversation. It is two policies exchanging tokens.

Where the idea tightened

Modern agents can type slowly, sleep on a schedule, imitate slang, and maintain a thread. You cannot perfectly determine human / not-human from any single interaction. You can measure something harder to fake continuously: the relationships among thousands of tiny actions over time.

Humanity is not noise. A random delay generator samples

A human delay is conditional:

Those distributions are not the same. Randomness is not humanity. The irregularity has context, and the context has to cohere across timescales. Unconditional jitter is cheap. Context-conditioned irregularity is not.

Independence, then the counterfactual

Ten thousand individually gorgeous mimics can still form a dead graph. Each account may score

Expose all ten thousand to the same event. Humans ignore, laugh, argue, misunderstand, wait, bookmark, reply, delete, return later, change subject, send it to a friend. An artificial population, even with varied sentences, converges on classify → optimize → publish.

Independence is not raw correlation with the crowd. Humans couple on a concert, a crash, a joke. The signal is residual coupling after you condition on the public stimulus, language, timezone, topic, and the feed actually shown. Coordinated humans exist too — newsrooms, fan accounts, activists. Coordinated is not the same as automated. The target is inauthentic coupling.

Sharper still: after you know what account u was shown, does knowing the rest of a suspected cluster still predict u’s next act? That predictive lift is counterfactual independence. High CI is interesting. It is not a verdict. Formal objects are in Falsify.

Four reactions

• Human → algorithm. Clumsy, off-policy, sometimes spiteful. That clumsiness is signal.
• Human → human. Asymmetric, decaying, loyal to people instead of metrics.
• Bot → algorithm. Gradient ascent on whatever is currently rewarded.
• Bot → bot. Cluster pulse. Reciprocal amplification without conversational debt.

What this is not

It is not proof of personhood. An iris scan answers “is there one living body.” A heartbeat answers “is this action independent and alive.” You may want both. Collapsing them is how you get a papers-please network. See Privacy.

It is not goodness. A living network can be cruel. Optimizing only for human–human volume rewards pile-ons. The heartbeat tells the ranker what kind of engagement it is looking at. It should not become the objective. The architecture is in Architecture. The ranking plug-in is in X Algorithm.