v0.1.3 — Float placement, keywords and arXiv source

• Keep all figures and tables before References with float barriers, and move float declarations closer to their first discussion.
• Add the requested keywords immediately after the abstract. The rebuilt paper is seven pages.
• Include the regenerated, comment-stripped arXiv source ZIP. A fresh extraction compiles with pdflatex only; extracted text, external links and every rendered page match the release PDF.

No scientific numbers, claims, figure contents or results changed from v0.1.2. No new inference or cloud compute.

Validation: all 36 tests pass; pip-audit reports no known vulnerabilities in 87 dependencies; 56 executed source files verify. All six floats are on pages 2, 4 and 5; References starts on page 6. All 18 PDF links resolve, and the former repository URL redirects. See SHA256SUMS, arxiv-validation.json and v0.1.3_verification.json for checksums and evidence.

v0.1.2: method overview and audited readability edits

• Add a one-column vector method overview, with the Pythonic sensitivity shown as a separate dashed branch.
• Improve readability, add the author's affiliation/contact, and apply the editing audit's scope, descriptive-comparison and throughput fixes.
• Rebuild the six-page paper and reproducible archive. No new results or model inference; frozen inputs, scientific numbers and figure data are unchanged.

Validation: 36 tests pass; pip-audit is clean; executed sources and archive checksums verify. All 18 PDF links resolve, and the former repository URL redirects to grammar-vs-bits. See SHA256SUMS for artifact hashes and the included audit-resolution document for the wording changes.

v0.1.1: Grammar vs Bits, rename and reframing

• Rename the project to Grammar vs Bits and update repository links, citation/package metadata and archive names.
• Reframe the paper and README around the separate format-contract sensitivity, with its scope and inconclusive incremental recovery explicit.
• Add the model × precision × arm breakdown of all 578 sensitivity token limits. They remain counted as incorrect.

No new primary results or model inference. Primary numbers, frozen inputs and executed snapshots are unchanged. The rebuilt paper is six pages; 36 tests pass, pip-audit is clean, archive checksums verify, all 18 PDF links resolve, and the former repository URL redirects.

v0.1.0: conceptual replication and audit fixes

Conceptual replication with the primary 24,000-generation matrix, paired confidence intervals, failure-exclusion sensitivity, and a separate 4,800-generation Pythonic contract sensitivity. All recorded operational failures remain in the scores.

This release applies all High/Medium audit findings and the pointwise wording correction. It includes the six-page paper, exact executed worker snapshots, MIT code license, BFCL Apache-2.0 attribution, frozen inputs, raw outputs, pinned dependencies, and reproduction scripts. The original audit and its resolution are preserved.

The separate sensitivity substantially improves Llama parallel-call accuracy under whole-response constraints. Neither its incremental recovery nor the change from the JSON contract is resolved from zero. It changes the prompt, parser and grammar together and includes a disclosed worker interruption and missing-cell continuation.

Download the archive with release.json and validation.json. The archive embeds its per-file SHA-256 inventory; the standalone inventory is also attached. The PDF is available separately. All 36 tests pass and dependency auditing is clean. Reported-to-date project cost is $3.346, including failures and the additional sensitivity; this is not a settled invoice. All project apps and containers were stopped.