{
  "claim": "Adding paraphrases of the right answer to the option list collapses the probability the model reports for it \u2014 from 0.619 to 0.215 \u2014 while the belief itself does not move. A 0.5 act/escalate threshold acts at five options and abstains at six, on identical evidence.",
  "method": "DeBERTa-v3-xsmall NLI q8 through the shipped path (transformers.js 4.3.0, ONNX Runtime Web, WASM). Start with the widget's five default options. Add, one at a time, three paraphrases of the correct one: 'a double charge that should be reversed', 'a payment taken twice that needs returning', 'a repeated charge awaiting a refund'. Context, model and hypothesis template fixed throughout. 'logit' is the raw entailment logit of the ORIGINAL correct option, before any softmax; 'p(concept)' sums the probabilities of every duplicate-charge phrasing present.",
  "source": "https://huggingface.co/Xenova/nli-deberta-v3-xsmall",
  "captured": "2026-09-19",
  "note": "The logit of the original option should be exactly invariant \u2014 each option is encoded and scored in its own row, so nothing another option does can reach it. It drifts by at most 0.26 across these four runs because the padded batch width changes as options of different lengths join it, and DeBERTa-v3's disentangled attention is a function of the sequence length. That drift is two orders of magnitude too small to explain the probability collapse, which is entirely the softmax denominator growing.",
  "columns": [
    {
      "key": "added",
      "label": "paraphrases added",
      "align": "right",
      "mono": true
    },
    {
      "key": "n",
      "label": "options",
      "align": "right",
      "mono": true
    },
    {
      "key": "logit",
      "label": "logit of original",
      "align": "right",
      "mono": true
    },
    {
      "key": "p",
      "label": "p(original)",
      "align": "right",
      "mono": true
    },
    {
      "key": "concept",
      "label": "p(concept)",
      "align": "right",
      "mono": true
    },
    {
      "key": "winner",
      "label": "argmax",
      "align": "left"
    }
  ],
  "rows": [
    {
      "added": 0,
      "n": 5,
      "logit": 2.2932,
      "p": 0.6193,
      "concept": 0.6193,
      "winner": "a duplicate charge that needs refunding"
    },
    {
      "added": 1,
      "n": 6,
      "logit": 2.1202,
      "p": 0.4824,
      "concept": 0.5889,
      "winner": "a duplicate charge that needs refunding"
    },
    {
      "added": 2,
      "n": 7,
      "logit": 2.031,
      "p": 0.2647,
      "concept": 0.7619,
      "winner": "a payment taken twice that needs returning"
    },
    {
      "added": 3,
      "n": 8,
      "logit": 2.1844,
      "p": 0.2151,
      "concept": 0.8211,
      "winner": "a repeated charge awaiting a refund"
    }
  ]
}