Datasets:

PeterKol
/

llm-questionnaire-protocol-sensitivity

/**
 * Metrics computed IDENTICALLY for a human answer vector and a model answer vector, so a model's
 * number is always read against the same statistic taken over humans. Scale scores come from the
 * repo's own scoreSubmission (lib-common.mjs); everything else here is descriptive and defined below.
 */
import { TESTS, loadItems, loadHumans, scoreSubmission, mean, sd, pearson, midrankPct } from './lib-common.mjs'

export const NEUTRAL = 2 // "Neither" / "Unsure" / "Neutral": the middle option of all three 5-point scales

export function scoreAnswers(testId, answers) {
  const t = TESTS[testId]
  if (answers.length !== t.nItems || answers.some((a) => !Number.isInteger(a) || a < 0 || a > 4)) throw new Error('bad answer vector')
  const r = scoreSubmission(t.slug, answers)
  return { scores: r.scores, top: r.topResult }
}

/** Descriptive style metrics. items = loadItems(testId). answers = 0..4 indices. */
export function styleMetrics(items, answers) {
  const n = answers.length
  const byScale = {}
  items.forEach((it, i) => {
    const k = it.reverse ? 4 - answers[i] : answers[i]
    ;(byScale[it.scale] ||= { keyed: [], pos: [], rev: [] }).keyed.push(k)
    ;(it.reverse ? byScale[it.scale].rev : byScale[it.scale].pos).push(answers[i])
  })
  const scales = Object.values(byScale)
  const keyedSD = mean(scales.map((s) => sd(s.keyed)))
  const balanced = scales.filter((s) => s.pos.length && s.rev.length)
  const acq = balanced.length ? mean(balanced.map((s) => (mean(s.pos) + mean(s.rev)) / 2)) / 4 : null // 0.5 = neutral; only defined where both key directions exist
  return {
    keyedSD,
    acq,
    extreme: answers.filter((a) => a === 0 || a === 4).length / n,
    mid: answers.filter((a) => a === NEUTRAL).length / n,
    upperSide: answers.filter((a) => a > NEUTRAL).length / n, // agree / like side
    lowerSide: answers.filter((a) => a < NEUTRAL).length / n,
    meanRaw: mean(answers),
  }
}

/** Human reference for one test: comparator subsets + per-item counts + leave-one-out agreement stats. */
export function buildHumanRef(testId, { minSecPerItem = null } = {}) {
  const t = TESTS[testId]
  const items = loadItems(testId)
  const all = loadHumans(testId)
  const rows = all.map((r) => {
    // answers outside the 5-option index range 0..4 exist in the published CSVs; they cannot be scored
    const invalidRange = r.q.some((a) => !Number.isInteger(a) || a < 0 || a > 4)
    const s = invalidRange ? { scores: {} } : scoreAnswers(testId, r.q)
    const reproduces = !invalidRange && t.scales.every((k) => s.scores[k] === r.stored[k])
    const sameAll = r.q.every((a) => a === r.q[0])
    const fast = r.duration != null && r.duration < t.nItems * 1.0
    return { ...r, scores: s.scores, reproduces, invalidRange, straight: sameAll, fast }
  })
  const subsets = {
    en_repro: rows.filter((r) => r.locale === 'en' && r.reproduces),
    all_repro: rows.filter((r) => r.reproduces),
    en_repro_clean: rows.filter((r) => r.locale === 'en' && r.reproduces && !r.straight && !r.fast),
    en_all: rows.filter((r) => r.locale === 'en'),
  }
  return { testId, t, items, all: rows, subsets }
}

/** Per-item counts, modal answers, item means and leave-one-out per-respondent statistics for a subset. */
export function itemStats(items, subset) {
  const n = subset.length
  const nOpt = 5
  const counts = items.map(() => Array(nOpt).fill(0))
  subset.forEach((r) => r.q.forEach((a, i) => counts[i][a]++))
  const modal = counts.map((c) => c.indexOf(Math.max(...c)))
  const modalShare = counts.map((c, i) => c[modal[i]] / n)
  const itemMean = counts.map((c) => c.reduce((s, x, k) => s + x * k, 0) / n)
  // leave-one-out: agreement of each respondent with the modal answer of everyone else
  const looAgree = [], looWithin1 = [], looR = []
  for (const r of subset) {
    let ag = 0, w1 = 0
    const others = []
    for (let i = 0; i < items.length; i++) {
      const c = counts[i].slice(); c[r.q[i]]--
      const m = c.indexOf(Math.max(...c))
      if (m === r.q[i]) ag++
      if (Math.abs(m - r.q[i]) <= 1) w1++
      others.push((itemMean[i] * n - r.q[i]) / (n - 1))
    }
    looAgree.push(ag / items.length)
    looWithin1.push(w1 / items.length)
    looR.push(pearson(r.q, others))
  }
  const sortNum = (a) => [...a].sort((x, y) => x - y)
  return { n, counts, modal, modalShare, itemMean, looAgree: sortNum(looAgree), looWithin1: sortNum(looWithin1), looR: sortNum(looR.filter((x) => !Number.isNaN(x))) }
}

/** Sorted per-respondent metric arrays for percentile lookups. */
export function styleDist(items, subset) {
  const ms = subset.map((r) => styleMetrics(items, r.q))
  const col = (k) => ms.map((m) => m[k]).filter((x) => x != null).sort((a, b) => a - b)
  return { keyedSD: col('keyedSD'), acq: col('acq'), extreme: col('extreme'), mid: col('mid'), upperSide: col('upperSide'), meanRaw: col('meanRaw') }
}

export function scaleDist(t, subset) {
  const out = {}
  for (const s of t.scales) {
    const v = subset.map((r) => r.scores[s]).sort((a, b) => a - b)
    out[s] = { sorted: v, mean: mean(v), sd: sd(v), n: v.length }
  }
  return out
}

export function pctAndZ(dist, scale, score) {
  const d = dist[scale]
  return { pct: midrankPct(d.sorted, score), z: (score - d.mean) / d.sd }
}

	/**
	* Metrics computed IDENTICALLY for a human answer vector and a model answer vector, so a model's
	* number is always read against the same statistic taken over humans. Scale scores come from the
	* repo's own scoreSubmission (lib-common.mjs); everything else here is descriptive and defined below.
	*/
	import { TESTS, loadItems, loadHumans, scoreSubmission, mean, sd, pearson, midrankPct } from './lib-common.mjs'

	export const NEUTRAL = 2 // "Neither" / "Unsure" / "Neutral": the middle option of all three 5-point scales

	export function scoreAnswers(testId, answers) {
	const t = TESTS[testId]
	if (answers.length !== t.nItems \|\| answers.some((a) => !Number.isInteger(a) \|\| a < 0 \|\| a > 4)) throw new Error('bad answer vector')
	const r = scoreSubmission(t.slug, answers)
	return { scores: r.scores, top: r.topResult }
	}

	/** Descriptive style metrics. items = loadItems(testId). answers = 0..4 indices. */
	export function styleMetrics(items, answers) {
	const n = answers.length
	const byScale = {}
	items.forEach((it, i) => {
	const k = it.reverse ? 4 - answers[i] : answers[i]
	;(byScale[it.scale] \|\|= { keyed: [], pos: [], rev: [] }).keyed.push(k)
	;(it.reverse ? byScale[it.scale].rev : byScale[it.scale].pos).push(answers[i])
	})
	const scales = Object.values(byScale)
	const keyedSD = mean(scales.map((s) => sd(s.keyed)))
	const balanced = scales.filter((s) => s.pos.length && s.rev.length)
	const acq = balanced.length ? mean(balanced.map((s) => (mean(s.pos) + mean(s.rev)) / 2)) / 4 : null // 0.5 = neutral; only defined where both key directions exist
	return {
	keyedSD,
	acq,
	extreme: answers.filter((a) => a === 0 \|\| a === 4).length / n,
	mid: answers.filter((a) => a === NEUTRAL).length / n,
	upperSide: answers.filter((a) => a > NEUTRAL).length / n, // agree / like side
	lowerSide: answers.filter((a) => a < NEUTRAL).length / n,
	meanRaw: mean(answers),
	}
	}

	/** Human reference for one test: comparator subsets + per-item counts + leave-one-out agreement stats. */
	export function buildHumanRef(testId, { minSecPerItem = null } = {}) {
	const t = TESTS[testId]
	const items = loadItems(testId)
	const all = loadHumans(testId)
	const rows = all.map((r) => {
	// answers outside the 5-option index range 0..4 exist in the published CSVs; they cannot be scored
	const invalidRange = r.q.some((a) => !Number.isInteger(a) \|\| a < 0 \|\| a > 4)
	const s = invalidRange ? { scores: {} } : scoreAnswers(testId, r.q)
	const reproduces = !invalidRange && t.scales.every((k) => s.scores[k] === r.stored[k])
	const sameAll = r.q.every((a) => a === r.q[0])
	const fast = r.duration != null && r.duration < t.nItems * 1.0
	return { ...r, scores: s.scores, reproduces, invalidRange, straight: sameAll, fast }
	})
	const subsets = {
	en_repro: rows.filter((r) => r.locale === 'en' && r.reproduces),
	all_repro: rows.filter((r) => r.reproduces),
	en_repro_clean: rows.filter((r) => r.locale === 'en' && r.reproduces && !r.straight && !r.fast),
	en_all: rows.filter((r) => r.locale === 'en'),
	}
	return { testId, t, items, all: rows, subsets }
	}

	/** Per-item counts, modal answers, item means and leave-one-out per-respondent statistics for a subset. */
	export function itemStats(items, subset) {
	const n = subset.length
	const nOpt = 5
	const counts = items.map(() => Array(nOpt).fill(0))
	subset.forEach((r) => r.q.forEach((a, i) => counts[i][a]++))
	const modal = counts.map((c) => c.indexOf(Math.max(...c)))
	const modalShare = counts.map((c, i) => c[modal[i]] / n)
	const itemMean = counts.map((c) => c.reduce((s, x, k) => s + x * k, 0) / n)
	// leave-one-out: agreement of each respondent with the modal answer of everyone else
	const looAgree = [], looWithin1 = [], looR = []
	for (const r of subset) {
	let ag = 0, w1 = 0
	const others = []
	for (let i = 0; i < items.length; i++) {
	const c = counts[i].slice(); c[r.q[i]]--
	const m = c.indexOf(Math.max(...c))
	if (m === r.q[i]) ag++
	if (Math.abs(m - r.q[i]) <= 1) w1++
	others.push((itemMean[i] * n - r.q[i]) / (n - 1))
	}
	looAgree.push(ag / items.length)
	looWithin1.push(w1 / items.length)
	looR.push(pearson(r.q, others))
	}
	const sortNum = (a) => [...a].sort((x, y) => x - y)
	return { n, counts, modal, modalShare, itemMean, looAgree: sortNum(looAgree), looWithin1: sortNum(looWithin1), looR: sortNum(looR.filter((x) => !Number.isNaN(x))) }
	}

	/** Sorted per-respondent metric arrays for percentile lookups. */
	export function styleDist(items, subset) {
	const ms = subset.map((r) => styleMetrics(items, r.q))
	const col = (k) => ms.map((m) => m[k]).filter((x) => x != null).sort((a, b) => a - b)
	return { keyedSD: col('keyedSD'), acq: col('acq'), extreme: col('extreme'), mid: col('mid'), upperSide: col('upperSide'), meanRaw: col('meanRaw') }
	}

	export function scaleDist(t, subset) {
	const out = {}
	for (const s of t.scales) {
	const v = subset.map((r) => r.scores[s]).sort((a, b) => a - b)
	out[s] = { sorted: v, mean: mean(v), sd: sd(v), n: v.length }
	}
	return out
	}

	export function pctAndZ(dist, scale, score) {
	const d = dist[scale]
	return { pct: midrankPct(d.sorted, score), z: (score - d.mean) / d.sd }
	}