full-audit-v2/source_audit_v2.json

{
  "paper": "Understanding Behavior Cloning with Action Quantization",
  "openreview_id": "9uENnRAcSl",
  "arxiv": "2603.20538v1",
  "source_pdf_sha256": "91d9f4e132d3f0251c32846497e7760ce83539022d9f9bcb1ff0d20c5c043ef9",
  "official_code": {
    "status": "not_found",
    "method": "paper text, arXiv landing page, OpenReview forum, Hugging Face paper page, and targeted title/author web search",
    "note": "The paper is theoretical and does not link an official implementation."
  },
  "claims": [
    {
      "claim": 1,
      "status": "confirmed_with_scope",
      "anchor": "Abstract; Section 3.2; Theorem 2; Section 5; Theorems 8-9",
      "pdf_pages": "1, 7, 11-12",
      "note": "The matching statement applies to the statistical sample term; the additive quantization term requires the theorem's stability/smoothness assumptions."
    },
    {
      "claim": 2,
      "status": "confirmed",
      "anchor": "Definitions 3-4; Theorems 2-3",
      "pdf_pages": "5-7",
      "note": "The bounds are polynomial under the specified global P-IISS/P-EIISS and TVC/RTVC conditions. The paper does not claim this for unstable dynamics."
    },
    {
      "claim": 3,
      "status": "confirmed",
      "anchor": "Theorem 6; Appendix D.2",
      "pdf_pages": "9, 26-33",
      "note": "The deterministic construction has expert-distribution quantization error O(epsilon_q) but H*Omega(1) deployed regret. A separate stochastic construction is also proved."
    },
    {
      "claim": 4,
      "status": "registered_claim_misstated",
      "anchor": "Algorithm 1; Theorem 7",
      "pdf_pages": "10-11, 34",
      "note": "The actual bound is H[sqrt((log(|Pi|/delta)+log(|M|/delta))/n)+epsilon_q] and assumes realizability for q#pi* in Pi and T∘rho in M. The registered wording omits log|M| and transition-model realizability."
    },
    {
      "claim": 5,
      "status": "confirmed_with_scope",
      "anchor": "Theorems 8-9; Appendix E",
      "pdf_pages": "11-12, 35-39",
      "note": "The deterministic expected lower bound is H(1/n+epsilon_q); the stochastic result is high probability and permits a suboptimal expert."
    },
    {
      "claim": 6,
      "status": "registered_claim_not_in_source",
      "anchor": "Proposition 5; Section 4.1 discussion",
      "pdf_pages": "9-10",
      "note": "There is no empirical binning-versus-learned-quantizer study and no measured deterministic-expert RTVC violation rate in this paper. Proposition 5 is theoretical; the adjacent practice observation cites Pertsch et al. (2025)."
    }
  ]
}