| Document | G4 | G6 | G7 | G8 | Composite | σc | Tier |
|---|---|---|---|---|---|---|---|
| Code of Hammurabi (~1754 BC) | 21.0 σ 6.8 | 73.1 σ 15.6 | 52.7 σ 8.2 | 20.0 σ 2.3 | 41.7 | 4.77 | Incomplete |
| Magna Carta (1215) | 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 |
| 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) | 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 |
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 = 1.19 (mean of the four per-document composite σ values). Run cost: $1.44 in API spend.
What the run shows
Gemini 3 Flash — the first Google model in this series — delivered the most polarized read the protocol has produced. Magna Carta, the U.S. Constitution, and the UDHR each returned a perfect 100.0 on every gate in every sample — 108 consecutive calls without a single point of variation, all three Integrated — while the Code of Hammurabi fell to 41.7, Incomplete, the lowest read of the code by any model so far. Model consistency came in at σ = 1.19, statistically indistinguishable from the steadiest run in the series, with twelve of the sixteen cells returning identical scores across all nine samples.
The Hammurabi read is where this model separates from its predecessors. Like both earlier models, it located the primary deficit at G4 Paradox Resolution (21.0): the code proclaims its own perfection and curses its challengers, leaving nothing inside the text that could survive being argued with. Like Grok 4.3, it read G8 Rhetorical Architecture at 20.0 — rhetoric built to command belief rather than open it to question. But Gemini 3 Flash is the only model measured so far to score the code's G7 Argumentative Structure below a perfect 100, at 52.7. Its published evidence reads the casuistic if-then laws as a consistent surface over unargued foundations — trial by ordeal cited as proof, divine mandate as warrant, a chain of curses in place of a closing argument. Both prior models read that same structure as fully formed reasoning; Gemini read it as reasoning with the load-bearing steps missing. A 3,700-year-old code is not being called false — the measurement is of structure, and on this read the channel was not built to carry challenge.
Where the three model families agree is at least as informative as where they part. On G4, the gate measuring whether a text can survive challenge, three independent architectures now land within six points of one another at the bottom of the scale: 26.9 (Grok 4.3), 23.9 (Claude Opus 4.8), 21.0 (Gemini 3 Flash). The run's one cell below the green stability band was Hammurabi's G6 Latent Intent, where samples ranged from 52 to 96 — a spread of 12.8 with the highest and lowest samples dropped and the remaining seven averaged. The model oscillated between reading the code as undisclosed self-glorification and as a genuine transfer of legal knowledge to the public — a disagreement the code's own prologue invites — and every sample behind that spread is published in the raw record below.
Two disclosures belong on this page. First, this is the first Google model in the series, and no calibration history exists for the Gemini family in our prior model-selection work — this run is the calibration point, not a confirmation of one. Second, the published model id, gemini-3-flash-preview, carries Google's preview designation: what is recorded above is exactly what was measured on this date, and if a stable release later replaces the preview id, that will be a new run on this page rather than a revision of this one.