04 — Specification

Human Liveness Signal.

A posterior over how an account participates, with a pulse and a width. Not identity. Not a public number. Consumed by ranking and by event-level monitoring.

Multi-timescale state

H_u(t) = [ H_seconds, H_session, H_day, H_social, H_long-term ]

Scale	Examples
Seconds / minutes	Reading time, click-to-reply, corrections, interrupted scrolls.
Session	Warm-up, session shape, unfinished conversations, density changes.
Day	Sleep windows, work, meals, commuting, weekends, irregular interruptions.
Social	Reply to one person immediately, ignore another for three days.
Long-term	Interests change, ties form, vocabulary shifts, habits appear and die.

Posterior over the taxonomy

pᵤ(t) = P(zᵤ | Hᵤ(t)), zᵤ ∈ {human, human+agent, declared-agent, service, coordinated}

Liveness is derived from that posterior, carried as (μ_L, σ_L). Cold start shrinks toward a weakly informative prior. Evidence insufficient → interval too wide to multiply the score.

Residual independence, then CI

Do not use 1 − Corr(R_u, R_population). Condition first:

I_u = 1 − Coupl(R_u, R_others | stimulus, feed, topic, tz, language)

Alignment after a football match is not a swarm. The sharper object is counterfactual independence — whether the cluster residual still predicts u after the stimulus u actually saw:

CÎᵤ ≈ ℓ(Rᵤ | Sᵤ, Fᵤ, Cᵤ, R_cluster) − ℓ(Rᵤ | Sᵤ, Fᵤ, Cᵤ)

Then inspect the event field P(E, t). Diversity here means unique kinds of act — ignore, linger, like, argue, bookmark, return, delete, forward — not unique strings.

Ranking consumption

Phoenix (or any action-head ranker) stays. Heartbeat is causal q_t on evidence and a prior on an author’s own distribution. Suggested product constraints:

Never show Lᵤ, σ_L, CI, or factor recipes to the user or to other users.
Wide interval on cold start. Do not hang new accounts.
Down-rank swarm engagement on a post more aggressively than the author when the author is independently alive and the audience is not.
Keep a labeled-agent path so legitimate automation is not forced to fake a pulse.
Only information before the scoring event is legal. P(E) in q_t is P(E, [t−W, t)), or a forecast conditioned on history before t.

Evaluation

If you only beat last year’s bots, you have not measured this system.

Holdout	Must catch / must not
Jitter bot	Unconditional delays. Layer 1.
Persona bot	LLM + sleep. Dies when D and G fail to cohere.
Behavioral clone	One scraped human. Dies at residual coupling of many clones.
Independent agent swarm	Separate models, shared objective. Catch in CI and ΔL.
Human farm	Alive pulse, economic regularity. Heartbeat alone is not enough.
False positives	Night-shift, ADHD, journalists, non-native, scheduled posts, accessibility tools.

Privacy rules live in Privacy. Attack models live in Adversaries. The miniature implementation is the lab.