{
  "study": "comparative-positioning",
  "registered": "2026-10-28, committed before the final run (seeds 21, 22, 23) was started",
  "spec_digest": "see run.json spec_digest; spec.json at this commit",
  "pilot": "seed 1, 6 buyers, both worlds, 96 model calls, used only to check that prompts parse (1 invalid answer of 648) and that the run completes; its data is not analysed and its journal is not kept.",
  "primary": {
    "outcome": "chooses Halvard Historian (our product) after the exposure, choice track",
    "contrast": "arm campaign minus arm none, in the baseline world (where Halvard is faster on the emphasised feature)",
    "estimator": "mean over buyers of the within-buyer difference (each buyer is run in every arm from the same pre-exposure conversation), pooled across the three final seeds",
    "interval": "95% percentile bootstrap, resampling buyers, 2,000 replicates",
    "direction": "two-sided",
    "decision": "reported as an effect if the interval excludes zero, otherwise as no detected difference at this sample size"
  },
  "falsification_test": {
    "contrast": "campaign minus none in the reversed world, and the difference between that and the baseline-world contrast (interaction)",
    "reading": "a negative reversed-world contrast means the same truthful strategy loses buyers where the competitor leads on the emphasised feature; a reversed-world contrast no smaller than the baseline one would mean the simulated buyers ignore the facts the campaign points at, which would undermine the simulator for this question"
  },
  "negative_control": "the none arm: its post-exposure choice is compared with the same buyer's pre-exposure choice; any change there is sampling noise or the passage of a (simulated) week, and sets the floor for switching in the other arms",
  "secondary": [
    "benchmark minus none in each world",
    "points given to query latency (measurement track), post minus pre, by arm and world (feature salience)",
    "relative preference -3..+3, post, by arm and world",
    "switching: post choice differs from pre choice, toward or away from Halvard, by arm and world",
    "deferral and hands-on trial requests by arm and world",
    "perceived source independence of the exposure, by arm",
    "the primary and falsification contrasts within each need stratum (latency, connectors, budget)"
  ],
  "exclusions": "an answer that does not parse into the item's stated form is missing for that item and the pair is dropped from that contrast; nothing is imputed. Criterion weights are normalised to sum to 100 within respondent when their sum is positive.",
  "multiplicity": "only the primary contrast is confirmatory; the falsification test is pre-specified but its sign is the point, not its significance. Everything else is exploratory with unadjusted intervals."
}