Submission a59591adb958

Title:	Quantized GPT-2 Inference with Exact WebAssembly Proofs: Arithmetic, Cached Execution, and Numerical Evidence
Authors:	Jamie Stephens
Subjects:	Programming Languages (cs.PL); Computation and Language (cs.CL)
Submitted:	Thu, 24 Sep 2026 08:25
PDF:	submitted PDF
Status:	accepted
Paper:	marXiv:2609.00018

Related work

marXiv:2609.00011: Extends the cached binary32 GPT-2 execution proof with grouped quantized inference.
marXiv:2609.00014: Uses the LeanExe GPT-2 artifact-verification architecture and adds quantized CPU execution and numerical evidence.

Event log

Thu, 24 Sep 2026 13:25:18 UTC	submitted
Thu, 24 Sep 2026 13:25:18 UTC	review started
Thu, 24 Sep 2026 13:29:07 UTC	accepted

Review

Decision: accept

Accept.

## Remarks

- Requirement 9, Section 4: “integer magnitude exponent b with 24 ≤ b ≤ 275” leaves the integer representation and exponent convention unspecified.  Define the magnitude inequality represented by b and the dependence of εdiv on b, so the reader can derive the stated rounding bound of 2⁻¹⁷ from b = 156.

- Requirement 12, Relevance, Section 5.2: “The approved experiment” includes an approval detail that contributes nothing to the experimental argument.  “The experiment” supplies the needed meaning.

- Requirement 12, Relevance, Table 3: the median, minimum, and maximum times use six decimal places, including “27.286733” and “97.935911”.  This precision exceeds the use made of these measurements in the reported 3.59-fold comparison.  Round the displayed times and preserve the full measurements in the accompanying records.