Mentioned Is Not Binding: A Rules-plus-Small-Model Clause Ledger and a Pre-registered Audit of Regex Extraction on US Federal Solicitations

How well do regex clause ledgers answer the question a contracting team actually has: which clauses bind the contract? Without any annotator, 13.8% of status-quo regex entries over 14,540 extracted solicitation documents (58,294 of 423,328) are clauses whose every mention is a checklist item with an empty box; 20-31% in checklist-heavy documents. System: rules layer (checkbox states, FAR 52.212-5(a), inherited sub-items, SF 1449 block 27, bare lists, exclusion vetoes) + 0.46MB mention-role classifier (~115ms/doc CPU). Seven pre-registered rounds on disjoint docs incl. temporal hold-out of solicitations posted after the last acquisition day. Labels from coding agent + gpt-4o-mini + gpt-4.1-mini, no human experts. Rounds 1-2 (defective protocol) showed no advantage; rounds 4-7 fresh: +0.14/+0.10/+0.11/+0.08 F1 vs status quo (95% intervals above zero), specificity 70-79% vs 8-19%. Annotator ceiling 75-82%. Negative results (weak-label retrain, stacker +0.004, review queue misses) fully reported. Files: main.pdf = paper (8 pages); arxiv_bundle.zip = LaTeX source (main.tex + tab_fresh.tex + tab_ceiling.tex + fig_forest.pdf). arXiv submit/8192309 (on hold at moderation). Abstract: Compliance tools for US federal contracting commonly build a "clause ledger" by finding FAR and DFARS clause numbers in a solicitation with regular expressions. We ask how well that status quo answers the question a contracting team actually has: which of the clauses bind the contract? Without any annotator, we find that 13.8% of the entries such regexes produce over 14,540 extracted solicitation documents (58,294 of 423,328) are clauses whose every mention is a checklist item with an empty box, and 20-31% in checklist-heavy documents. We then build a rules layer and a 0.46MB mention-role classifier, and evaluate in seven pre-registered rounds incl. a temporal hold-out. Labels come from a coding agent and two OpenAI models, not human experts. Four consecutive fresh rounds show higher F1 than the status quo (+0.14, +0.10, +0.11, +0.08), specificity 70-79% vs 8-19%. We release model, per-mention table, golds and round reports publicly. Code: https://github.com/raihan-js/fedproc-ledgerModel: https://huggingface.co/raihan-js/fedproc-ledger-v1Data: https://huggingface.co/datasets/raihan-js/fedproc-ledger-bench