we present tribe-v2 — trait-based reader inference & behavioral emulation — a method for forecasting the downstream consequence of a piece of content before it is published. a population of synthetic readers, each parameterised by a modeled prior, an attention budget, and a position in a social graph, simulates response to a draft. we report calibration against a held-out corpus of 2.1m posts and characterise the regimes in which the forecast remains reliable and those in which it breaks down.

01introduction

platforms measure what lights up in the first hour. that signal is cheap and abundant, and it is not the thing most writers actually care about: what a piece of content does — to whom, for how long, and with what second-order effect. we treat this as a forecasting problem. given a draft and an audience, what response will it generate, and what is our uncertainty about that estimate?

prediction from text features alone is brittle. the same sentence lands differently across communities, and aggregate metrics hide the segment that reads against the grain. we model readers, not keywords. a forecast is produced by simulating a population and aggregating individual responses, which lets us report distributional outcomes — including the tail risks that a single point estimate would erase.

02method

tribe-v2 constructs a synthetic population of n = 8,412 agents calibrated to a target audience. each agent carries three parameters that govern how it reads:

  • prior — the agent's existing stance and familiarity with the topic, sampled from the calibrated population.
  • attention budget — how much of the content the agent actually consumes before forming a response; most agents do not finish.
  • graph position — the agent's connectivity, which governs whether its response propagates or terminates.

agents read the draft, form a response, and — conditioned on graph position — propagate it. we run the simulation forward to estimate projected reach, the response half-life, the probability of breakout into adjacent communities, and any latent backlash risk.

fig 1. simulated response field across the calibrated population, t+12h. darker regions denote concentrated propagation.

a forecast you cannot audit is an opinion wearing a number.

03results

against the 2.1m-post held-out corpus, tribe-v2 forecasts of projected reach correlate at r = 0.79 with measured outcomes, and backlash-risk flags precede measured adverse response in 83% of flagged cases. every forecast ships with error bars derived from the population variance, not a single deterministic score.

calibration degrades predictably as the draft moves further from the calibration distribution. we report this explicitly rather than extrapolating — a forecast outside the validated regime is returned with a widened interval and a coverage warning.

04limitations

synthetic agents are calibrated approximations, not people. they drift toward agreement over long simulation horizons (we quantify this sycophancy bias in a companion note) and they cannot anticipate exogenous events that reshape a community between draft and publication. the method forecasts the response a draft is likely to generate under stable conditions — it does not predict the world.

cite this work
@techreport{afterwave2026tribe,
  title  = {TRIBE-v2: Agent-Based Forecasting of Content Consequence},
  author = {Afterwave Research},
  year   = {2026},
  institution = {Afterwave Research},
  note   = {v2.0, preprint}
}