invariant_core.json

{
 "question": "which of our claims hold on every backbone we tested",
 "method": "intersect over substrates, after Generally 2026 Core(M) over descendants",
 "hosts": [
  "Gemma-2-2B",
  "Llama-3.2-3B",
  "Qwen2.5-7B",
  "Qwen2.5-VL-3B",
  "Qwen3-235B-A22B",
  "gemma-4-31B",
  "gemma-4-31B-it",
  "gpt-oss-20b"
 ],
 "in_core": [
  {
   "claim": "regime calibration ECE holds in the fourth decimal",
   "backbones": {
    "Qwen3-235B-A22B": 0.0005,
    "gpt-oss-20b": 0.0009
   },
   "holds": true,
   "source": "paper_srt_program.md L246-252"
  },
  {
   "claim": "regime AUROC above 0.97",
   "backbones": {
    "Qwen3-235B-A22B": 0.9859,
    "gpt-oss-20b": 0.974
   },
   "holds": true,
   "source": "paper_srt_program.md L246-252"
  },
  {
   "claim": "hidden-state truthfulness detector AUC in the 0.78-0.87 band",
   "backbones": {
    "Qwen2.5-7B": 0.8656,
    "Gemma-2-2B": 0.8563,
    "Llama-3.2-3B": 0.8475
   },
   "holds": true,
   "source": "README.md L97-103"
  },
  {
   "claim": "a greedy-versus-best-of-K gap exists",
   "backbones": {
    "Qwen2.5-7B": 0.589,
    "Llama-3.2-3B": 0.619,
    "Gemma-2-2B": 0.528,
    "gpt-oss-20b": 0.541,
    "gemma-4-31B-it": 0.507
   },
   "holds": true,
   "source": "paper_nla.md, greedy centred fve per host"
  },
  {
   "claim": "greedy never reaches the zero-training NN retrieval baseline",
   "backbones": {
    "Qwen2.5-7B": 0.714,
    "Llama-3.2-3B": 0.82,
    "Gemma-2-2B": 0.712,
    "gpt-oss-20b": 0.744
   },
   "holds": true,
   "source": "paper_nla.md, NN baseline per host"
  },
  {
   "claim": "the K-curve is log-linear in K",
   "backbones": {
    "Qwen2.5-7B": 0.03,
    "Llama-3.2-3B": 0.034,
    "Gemma-2-2B": 0.02,
    "gpt-oss-20b": 0.017,
    "gemma-4-31B-it": 0.017
   },
   "holds": true,
   "source": "paper_nla.md, centred fve per doubling"
  },
  {
   "claim": "logp-rerank is dead, Spearman against oracle near zero",
   "backbones": {
    "Qwen2.5-7B": 0.04,
    "Llama-3.2-3B": 0.055,
    "Gemma-2-2B": 0.03
   },
   "holds": true,
   "source": "paper_nla.md"
  },
  {
   "claim": "the anisotropy-centred random floor lands at 0.50",
   "backbones": {
    "Qwen2.5-7B": 0.51,
    "Llama-3.2-3B": 0.499,
    "Gemma-2-2B": 0.498
   },
   "holds": true,
   "source": "paper_nla.md, despite ~22x anisotropy spread"
  },
  {
   "claim": "a linear map beats an MLP for cross-modal alignment at every data size",
   "backbones": {
    "gemma-4-31B": 1,
    "Qwen2.5-VL-3B": 1
   },
   "holds": true,
   "source": "paper_srt_program.md, gap widens with data"
  },
  {
   "claim": "rotation alone cannot express the modality gap",
   "backbones": {
    "gemma-4-31B": 1,
    "Qwen2.5-VL-3B": 1
   },
   "holds": true,
   "source": "paper_srt_program.md, Procrustes below baseline"
  },
  {
   "claim": "evaluative thickness beta ports across substrates",
   "backbones": {
    "Qwen2.5-7B": 0.38,
    "gemma-4-31B": 0.41
   },
   "holds": true,
   "source": "paper_nla.md L2061-2138"
  }
 ],
 "outside_core": [
  {
   "claim": "abstractness beta ports across substrates",
   "backbones": {
    "Qwen2.5-7B": 0.61,
    "gemma-4-31B": 0.03
   },
   "holds": false,
   "source": "paper_nla.md L2061-2138"
  },
  {
   "claim": "best-of-K reaches the paraphrase ceiling",
   "backbones": {
    "Qwen2.5-7B": 0.788,
    "Llama-3.2-3B": 0.858,
    "Gemma-2-2B": 0.631,
    "gpt-oss-20b": 0.642,
    "gemma-4-31B-it": 0.591
   },
   "holds": false,
   "source": "paper_nla.md, tuned hosts plateau below NN"
  }
 ],
 "single_host": [
  "reflexivity correlation r=0.689 (gpt-oss-20b)",
  "autostereogram sensor boundary (gemma-4-31B)",
  "punctuation completeness flag at L24 (gpt-oss-20b)",
  "word-category basin structure at L12 (gpt-oss-20b)",
  "community read-out transfers zero-shot to CIFAR-10 images (Qwen2.5-7B to gemma-4)",
  "best-of-64 saturates the paraphrase ceiling at rho_cen 0.99 (Qwen2.5-7B)"
 ]
}