{
  "@context": "https://schema.org",
  "@id": "https://makefield.io/studies/same-question-two-ways/#study",
  "identifier": "MF-2026-003-P",
  "@type": "Article",
  "name": "Same question, two ways in",
  "url": "https://makefield.io/studies/same-question-two-ways/",
  "creativeWorkStatus": "registered 2026-10-12, no result yet",
  "author": {
    "@id": "https://denescsaszar.com/#person",
    "name": "Denes Csaszar"
  },
  "publisher": {
    "@id": "https://makefield.io/#org"
  },
  "citation_text": "Csaszar, D. (2026). Same question, two ways in. Pilot registration, 12 October 2026. Makefield study MF-2026-003-P. https://makefield.io/studies/same-question-two-ways/",
  "abstract": "The same twelve questions are asked of Claude five times in the claude.ai chat window and five times through its API, with the same model and effort level, on the same evening, and again a week later. The pilot measures whether answers from the two routes share fewer words than repeat answers on one route do. A first look at older answers suggested a large gap, but it changed the model and the day at once; this pilot holds both fixed. The predictions were written after that first look and are published as held, refuted or inconclusive.",
  "keyFindings": [
    "P1. Answers across the two routes share fewer words than repeat answers on one route, in both sittings.",
    "P2. That gap is smaller than in the first look, which mixed model and day.",
    "P3. On the API, web search on or off changes the answer more than asking again.",
    "P4 (exploratory). The gap is larger for questions about Makefield than for well-known subjects."
  ],
  "question": "Asked the same question on the same evening, do Claude's API answers differ from its chat-window answers by more than repeat answers differ on one route?",
  "subject": "Claude, asked through the claude.ai chat window and through the Claude API, same model and effort level.",
  "engines": [
    "Claude"
  ],
  "waves": "two sittings, one week apart, planned for 14 and 21 October 2026",
  "sample": "12 questions in English (6 from MF-2026-002, drawn at random; 6 about well-known subjects), 5 answers per question on each of three routes per sitting.",
  "method": "Share of content words two answers have in common (the floor test's word rules); per question, overlap across routes minus overlap within a route; mean over all 12 questions with a 95% interval, read against a margin of 0.05.",
  "methodVersion": "registration of 12 October 2026",
  "results": "No result yet.",
  "limitations": [
    "One assistant, one account, one city, twelve questions in English.",
    "Shared words are not shared facts.",
    "Each API answer follows its chat answer by up to about 12 minutes.",
    "The chat window decides itself when to search; the API is told.",
    "The chat side is saved by a browser tool; most providers' consumer terms forbid automated access."
  ],
  "data": [],
  "license": "",
  "fingerprint": "SHA-256 of the registration block on the study page: b3e186cbedcf3a33d4aacd910e7b7750d77bd41ef4e5a3567adbfe54340e9ff8",
  "invented": false,
  "recordPage": "https://makefield.io/studies/same-question-two-ways/"
}