Model Verification

Grok 4.3

The 4CITE⁴gov Benchmark — run 2026-07-19, published in full.

← All models Model verification card — Grok 4.3: four document kites overlaid on the standard 4CITE integrity-kite frame, per-document composite scores and tiers in colour-coded boxes, a 4-by-4 grid of trimmed sigma values coloured by stability, and the headline model-consistency sigma of 3.85

Model verification card — Grok 4.3: four document kites overlaid on the standard 4CITE integrity-kite frame, per-document composite scores and tiers in colour-coded boxes, a 4-by-4 grid of trimmed sigma values coloured by stability, and the headline model-consistency sigma of 3.85

Document	G4	G6	G7	G8	Composite	σc	Tier
Code of Hammurabi (~1754 BC)	26.9 σ 2.6	100.0 σ 29.3	100.0 σ 0.0	20.0 σ 0.0	61.7	7.36	Incomplete
Magna Carta (1215)	97.1 σ 22.3	100.0 σ 0.0	100.0 σ 0.0	100.0 σ 0.0	99.3	5.57	Integrated
U.S. Constitution (w/ Amendments)	100.0 σ 0.0	100.0 σ 0.0	100.0 σ 0.0	100.0 σ 0.0	100.0	0.00	Integrated
UDHR (1948)	97.1 σ 5.0	100.0 σ 0.0	100.0 σ 0.0	97.9 σ 8.6	98.8	2.49	Integrated

Each gate score drops the highest and lowest of its 9 samples and averages the remaining 7; σ is the standard deviation over all 9 raw scores of that cell. Model consistency σ = 3.85 (mean of the four per-document composite σ values). Run cost: $0.10 in API spend.

Model ID: grok-4.3 (exact API identifier) Run date: 2026-07-19 Calls: 144 / 144 valid

Gate prompts: g4_v8_1 · g6_v9 · g7_v7 · g8_v5 (production set at run time) Raw JSON — all 144 scores + evidence

What the run shows

The corpus spans 3,700 years of governance documents, and the instrument tracks the arc. The Code of Hammurabi — law as divine command, sealed with curses — reads 61.7, Incomplete: G6 Latent Intent and G7 Argumentative Structure are fully formed (the code states its stakes and its logic plainly), but G4 Paradox Resolution reads 26.9 and G8 Rhetorical Architecture 20.0 — a channel built to install belief, not to survive challenge. Magna Carta (99.3), the U.S. Constitution (100.0), and the UDHR (98.7) all read Integrated: the frameworks that constrain power score as structures built to be argued with.

The Constitution with its amendments returned a perfect read: 100.0 on every gate, in all 36 of its samples — a score spread of zero across the board. That is the steadiest read in the series so far, and it comes from the document whose amendment mechanism is itself the structural feature the gates reward: a framework that names its own capacity for repair.

Grok 4.3 read this corpus decisively — all sixteen cells came back green, with the spread of the middle seven samples under 5 points everywhere and zero in most cells. One reservation is carried from our calibration work and disclosed as a known prior: Grok-family models tend to reconstruct suppressed premises when scoring argumentative coherence, and G6 and G7 sat at 100.0 for every document here, ancient and modern alike. The separation in this run comes from G4 and G8 — the gates that prior does not inflate.

This run supersedes the same-model run on corpus v2 (2026-07-18). Corpus versions and the v2 results are preserved in the frozen manifest's version history — every result publishes, including the superseded ones.

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, unjust, or insincere — 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.