piatra.institute

PIATRA . INSTITUTE

ideological bent

measuring how an actor swap bends a forecast from its causal baseline
based on Taber and Lodge, Motivated Skepticism in the Evaluation of Political Beliefs (2006)
claude opus 4.8June 2026·first cut. ports the Bent Test prototype to playground conventions: PlaygroundLayout, black-and-lime palette, a softmax-logit baseline forecast, an editable forecast board, the actor-swap diagnostics (identity likelihood ratio, Bent Score, Jensen-Shannon brittleness, identity dominance ratio, counterfactual inadmissibility), an evidence-debt panel, a causal map, an assumption sweep, calibration of the formulas, and nine assumptions.
frontier-model access restriction
same facts, different actor. allocate probabilities, then read how far your forecast bends from the baseline under the actor swap.
bent: 0.00
IDR: 0.18
baseline model your forecast
Allocate 100 percentage points for each administration. The baseline bars move live when you change the assumptions. Your inputs show your counterfactual forecast.
Administration A
high unilateralism, high legal aggressiveness, lower allied-process preference
actor
No major new restriction
baseline 5%
Temporary pause + safety review
baseline 18%
Tiered / trusted-access restriction
baseline 33%
Same blanket foreign-national ban
baseline 44%
total100%
Administration B
high AI-safety concern, high regulatory-process preference, stronger allied-coordination preference
actor
No major new restriction
baseline 13%
Temporary pause + safety review
baseline 32%
Tiered / trusted-access restriction
baseline 37%
Same blanket foreign-national ban
baseline 18%
total100%

Making identity pay rent

The dangerous move is “my opponent did X, therefore my opponent is uniquely to blame, and my side would never do X.” Sometimes true, often false, usually under-specified. The better move asks which part of the outcome came from the actor, which from state capacity, which from agency incentives, which from the underlying threat, and which from international pressure. This instrument forces that decomposition by making you assign probabilities and then checking your own consistency when the only thing that changes is the actor label.

It is symmetric by construction. “They would obviously have done the same” and “my side would never” are flagged as distortions of equal kind. The tool measures the shape of your reasoning, not which side you are on.

The metrics

The baseline is a softmax over log-priors shifted by the shared assumption sliders. Your forecast is compared to it along several axes. The identity likelihood ratio is the odds multiplier your deviation implies. The Bent Score is how much more (or less) your forecast separates the two actors than the baseline does, in log-odds:

Bent=[logit qA−logit qB]−[logit pA\*−logit pB\*]\text{Bent} = \big[\text{logit}\,q_A - \text{logit}\,q_B\big] - \big[\text{logit}\,p^\*_A - \text{logit}\,p^\*_B\big]

Ideological brittleness measures the same separation with Jensen-Shannon divergence; the Identity Dominance Ratio compares how much the label moves you to how much the facts move the model; and counterfactual inadmissibility, in bits, measures how hard you suppress a branch the baseline keeps open.

Rational, not automatic

Two people disagreeing is not automatically irrational: different priors or causal assumptions can produce honest divergence. That is why the assumptions are exposed as sliders. A high Bent Score that survives even after you match assumptions and priors across the swap is the motivated residue the tool is trying to isolate, the difference between serious reasoning and team-flag reasoning.

What this is not

The baseline probabilities are editable placeholders for interface design, not calibrated forecasts, and the scenarios are illustrative test cases, not endorsements, predictions, or factual claims about any real person. Counterfactuals are not directly verifiable; the tool never scores them as true or false. It scores reasoning habits: consistency under the actor swap, sensitivity to evidence, and the admissibility of inconvenient branches. A serious version would replace the placeholder priors with calibrated forecaster pools and historical analogues.

Model changelog

v1.0June 2026
  • baseline forecast: a multinomial-logit model, softmax over log-priors shifted by assumption sliders times hand-tuned weights.
  • forecast board: editable per-actor, per-outcome probabilities with live baseline bars and normalise / copy-baseline helpers.
  • diagnostics: identity likelihood ratio, Bent Score (logit actor-effect gap), ideological brittleness (Jensen-Shannon), identity dominance ratio, and counterfactual inadmissibility in bits.
  • evidence debt: selected reasons supply an odds multiplier checked against the multiplier your deviation requires.
  • blind-label mode hides actor identities so the facts lead; the diagnostics are symmetric across the swap.
  • assumption sweep, sensitivity table, causal map, and timeline decomposition to keep distinct outcomes from collapsing.
  • calibration verifies the diagnostic formulas against hand-computed values rather than empirical data.
read the research companion →