{
  "title": "Assessment stability: recorded engineering experiments",
  "author": "Kent Tan · Certesian",
  "edition": "2026-09-09",
  "provenance": "Aggregate results transcribed from the author's Research #001 and NACOSCE engineering note dated 2026-09-04. Experiments were not rerun for this website.",
  "nmc": {
    "scope": "One motivating transcript, four prompt configurations, five marking runs per configuration. Exact marking-model identifier and transcript are not supplied in this public aggregate record.",
    "runsPerPrompt": 5,
    "rows": [
      {"name":"Shipped prompt","short":"Baseline","verdicts":[true,true,true,true,true],"unstable":0,"evidenceProblems":5},
      {"name":"Bar chart citations","short":"Edit 01","verdicts":[true,true,true,true,true],"unstable":1,"evidenceProblems":4},
      {"name":"Restructure + emphasise","short":"Edit 02","verdicts":[true,false,false,true,true],"unstable":6,"evidenceProblems":0},
      {"name":"Add two sentences","short":"Edit 03","verdicts":[true,false,true,true,true],"unstable":2,"evidenceProblems":0}
    ],
    "interpretation": "In this experiment, the edits that removed evidence problems also introduced verdict variation. The prompt edits were reverted. This is not an accuracy benchmark or proof that prompting cannot help."
  },
  "nac": {
    "model":"gemini-3.8-flash",
    "transcripts":3,
    "conditions":53,
    "rows":[{"name":"Before polarity change","runs":5,"unstable":0},{"name":"After polarity change","runs":5,"unstable":1},{"name":"After change, additional run","runs":3,"unstable":2}],
    "interpretation":"The three non-occurrence conditions remained stable in all three runs. Variation appeared in compound conditions and quality judgements. This small sample neither validates nor refutes the polarity change; it does not establish general reliability."
  }
}