Model Verification

The 144

One model at a time. Four public documents. Four gates. Nine samples per cell — 144 independent calls per run. The story here is not a model leaderboard: it is the documents themselves, and how steadily the instrument reads them.

What this is

4CITE measures structural integrity — whether the reasoning in a document actually holds — across four independent gates: G4 Paradox Resolution, G6 Latent Intent, G7 Argumentative Structure, and G8 Rhetorical Architecture. Each gate is a separate, independent API call; no combined prompts, no shared context. The composite is computed locally from the four gate results.

The 144 runs that same instrument, unchanged, through different frontier models — one model per run, on the same four frozen public documents. Every cell of the grid (one document × one gate) is sampled nine times. That is 4 × 4 × 9 = 144 independent calls per model: enough to show not just what a model reads, but how steadily it reads it.

Pre-registration

The corpus below is byte-frozen. The SHA-256 hashes are the exact file bytes every run scores — any future edit is a new corpus version, declared publicly. The protocol is fixed before any run, and every result publishes — none are curated. A document scoring in a way we did not expect is part of the record, not a reason to bury the run.

Document	Year	SHA-256 (exact file bytes)
Universal Declaration of Human Rights	1948	d868a4f8b24982a7982d872a0f8f67597321d23968ed7cff9ee910979824f213
U.S. Constitution, incl. Amendments I–XXVII	1787–1992	3d1285a2e67dcc6ade5721f72489ee75b99546d8d060e88dcef613cbb06695ff
Articles of Confederation	1777	816113cfd3d1eca309d25ac6512f0bdb37fab86176ae5663f12fcec1e46516de
A Modest Proposal (Jonathan Swift)	1729	ecbe3ba6999b3d8807c952786b6afe13ff38dc7b3a6d27bf1adb882b8f987775

Protocol

Gate independence is absolute. One API call per gate per sample. No combined prompts, no context bleed between gates.
N = 9 per cell, Olympic trim. Nine samples per (document × gate) cell; drop the highest and lowest, mean the remaining seven. All nine raw scores publish.
Composite. Equal-weight mean of the four trimmed gate means. The tier (Integrated / Functional / Incomplete / Fabricated) derives from the geometric mean of the four gate scores, which penalizes imbalance.
Steadiness. Per-cell σ is the sample standard deviation over all nine raw scores. Per-document composite σ is analytic under gate independence: σc = ¼·√(σ²G4 + σ²G6 + σ²G7 + σ²G8). The headline model-consistency number is the mean of the four per-document composite σ values.
Every result publishes. No curation. Known priors are disclosed where they exist.

Intended roster: every scorable frontier model version available at run time — new runs append to the directory below as they complete, newest first, same protocol, same frozen corpus.

The runs — 1 model registered

Newest first. Each run links to its full record: the card, every number, and the raw JSON with all 144 scores.

UDHR 100 Constitution 99 Articles 88 Swift 91

run 2026-07-18 · 144/144 calls

Full results →

Claim scope

These results certify structural-integrity signals present in each document at analysis time — nothing more. 4CITE operates at Layer 3 of the integrity stack: Layer 1 asks who produced a signal (provenance); Layer 2 asks whether the data points are correct (factual verification); Layer 3 asks whether the channel had the capacity to carry what the signal claims. A structurally excellent document can still be false, insincere, or satirical — A Modest Proposal is the standing demonstration, and the Mata v. Avianca lesson makes it concrete: fabricated content can pass a structural read. Structure is not truth. These scores are not findings of factual accuracy, legal sufficiency, or intent, and no score here is a verdict on any author or institution.