political autoimmunity
From “clueless” to a measurement problem
Calling a group of voters self-destructive is cheap and usually wrong. It assumes the group has one true interest, treats a category as a monolith, and smuggles in a partisan judgment. The useful move is to stop asking “why are these voters clueless?” and start asking: under which assumptions, data sources, and definitions of interest does a vote become measurably misaligned with a group's exposure, institutional dependence, or stated priorities? Political autoimmunity is that question made into an instrument.
Every input here is synthetic, taken from the ideation's worked example. The instrument is symmetric: it applies to any group supporting any coalition, on either side of politics. It scores the structure of an argument, never a person.
The risk chain
For a group, a coalition, and a policy domain, the per-domain adverse risk is a product. Foreseeability (awareness) and salience scale it down; the net score subtracts a protective benefit and a tolerance and reweights by the interest model:
Because risk is multiplicative, any single near-zero factor (low implementation probability, low awareness, low salience) collapses a domain's contribution. That is deliberate: a harm a coalition cannot or will not implement, or that a voter never sees coming, should not count the same as one that is certain and salient.
There is no single interest
The interest model is a weight profile over domain kinds, and the case ranking changes when you switch it. Under rights-dependence the small, high-exposure case leads; under a material or expressive model it can fall behind. A vote that looks like self-harm under one definition of interest is an informed tradeoff under another. The expressive and protest models, in particular, treat symbolic belonging and punishment as real utility, which is the strongest reason most apparent self-harm is not irrationality.
What this is not
This is a transparent sandbox over synthetic numbers, not an empirical estimate and not a forecast. The cells are hand-set illustrations, group labels are coarse (no subgroup or intersectional structure), counterfactuals are never scored as true or false, and the uncertainty band is a seeded Monte Carlo with a fixed concentration, not a real posterior. A serious version would replace the cells with survey-derived exposure (CES, ANES, Pew, AP VoteCast), a human-coded candidate-policy matrix with intercoder reliability, and calibrated forecaster priors. The calibration panel only checks that the engine reproduces the worked example, nothing more.
Model changelog
- risk chain: per-domain priority risk = exposure · dependence · hostility · implementation · magnitude · awareness · salience, exactly as specified in the ideation.
- net autoimmunity: voteShare · Σ interest-weight · max(0, priorityRisk − benefit − τ), with a gross vs net toggle and a per-supporter vs population-weighted toggle.
- competing interest functions: rights-dependence, material, balanced, expressive/status, punitive protest, and long-run institutional, each a weight profile over domain kinds; the case ranking reorders across them.
- uncertainty: a deterministic, seeded Monte Carlo perturbs every cell with a Beta around its value and reports a 90% interval.
- sensitivity: a tornado view sweeps each assumption family min→max to show which one drives the focus score.
- cross-model matrix: population-weighted net score for every case under every interest model, flagging when the top case changes.
- calibration verifies the engine reproduces the ideation's synthetic worked example rather than empirical data.