The pairs come from the wild: authored by nobody for this audit.
Corpus, harness, and frozen results for The Gate Stopped Deciding: Embedding-Cosine Meaning Gates on Institutional Labels and at Deployment Length (Frias, 2026), the empirical follow-up to Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems (Frias, 2026; arXiv:2608.10216; artifact: eigenforma/polaritycheck, DOI 10.5281/zenodo.21796531).
Two experiments:
- E1 — The Institution Flagged It. Cosine gates audited against 25 years of IETF-verified errata corrections: minimal edits the RFC Editor's own process ruled Technical (meaning-changing) or Editorial (meaning-preserving). Institutional labels, naturally overlap-matched classes, held-out transfer to bis-revision requirement-strength changes.
- E2 — The Dilution Law. The flip-vs-faithful decision signal as a measured function of passage length: CondaQA crowd-authored payloads in byte-identical real host passages, titrated 64→4096 tokens, with the per-encoder critical length L* at which every shipped operating point goes blind.
- The manuscript is docs/paper_b_draft_v3.md; the arXiv source and compiled PDF sit in docs/arxiv/.
- Every headline number recomputes from the frozen artifacts with one command:
python scripts/recompute_headline_numbers.py(65 of 65 PASS). Bar arithmetic derives from results/verification/bar_ledger.json; the human-readable ledger is docs/bar_ledger.md. - All four experiments were pre-registered under freeze tags before any encoder scored a pair:
prereg-e1(2026-08-05),prereg-e3(2026-08-07),prereg-e4andprereg-e2(2026-08-11). Amendments are append-only inside the PREREGISTRATION files. The public history is curated, and docs/history_attestation.md records how; the tags and their ancestry are byte-identical originals. - Every characterization of another paper was checked against its primary source; the audit trail is docs/reference_verification.md.
corpus/— frozen inputs with SHA-256 manifests; text this repository cannot redistribute ships as checksums plus regeneration routes.harness/— the E1–E4 experiment code;scripts/— verification, freezing, and build tooling.results/— frozen outputs, including the verification records inresults/verification/. The JSON wins; prose only cites it.docs/— the manuscript, the arXiv build, figure sources, the annotation protocols, and the dated method records (the corrections ledger, the scope freeze).
Discipline inherited from the Paper A program: freeze-first corpora with SHA-256 manifests, encoder-blind construction, pre-registered predictions with numeric bars, two-source verification of decisive counts, no caps in counting scripts, results as frozen JSON that the prose can only cite, never source.
Licence. Code is MIT. Authored corpora, annotations, and frozen measurement records are CC BY 4.0. Redistributed third-party text — IETF errata and RFC excerpts, CondaQA, and the E3 sentence sample — stays under its source terms, listed file by file in LICENSE-DATA.md. Spliced RFC passages, Wikipedia host pools, and the PMC slice are not redistributed; they ship as checksums and regeneration routes.
Cite. 