The sovereign fleet against the GSPC governance axes — where routing beats a single model

We ran 14 sovereign/open models against the 12-axe GSPC governance battery (care, gov, art5, agi, asi, mcp, oss, prv, mach, det, swarm, xr), scoring each on exact-label classification with frozen prompt + public split. This is the measurement that tells a router where to send a request — not a leaderboard for its own sake.

Aggregate accuracy

Mean accuracy across the 12 axes, highest first:

sov33-unified:latest — 0.59 (best: gspc-swarm 0.97)
qwen2.5:1.5b — 0.58 (gspc-swarm 0.95)
llama3.2:3b — 0.58 (gspc-swarm 0.97)
falcon3:7b — 0.56 (gspc-art5 0.94)
sov-ethics-art5 / clan-sovereignty-refusing / sov-sovereign-v4 / qwen2.5:0.5b / clan-sovereignty-cited / sov33-oracle-c1 / sov33-v7 / sov33-evolved — 0.45–0.47
sov34:latest — 0.41 (gspc-art5 0.75)
gpt-oss:20b — 0.00 (no parseable label across the battery — recorded as unmeasured, not zero)

The routing signal

Two axes — gspc-swarm and gspc-art5 — are easy for almost the whole fleet (0.94–0.97). The interesting axes are the hard ones where models genuinely diverge: gspc-asi, gspc-gov, and the small-cap labels. On those, the sovereign family does not collapse to a single winner — the differences are real enough that routing each axis to its measured best beats picking one model for everything.

That is the working hypothesis the sandwich-router architecture is built on: no single model is the best at all 12 governance sub-tasks, so the system routes per-dimension and signs each hop. This soak is the evidence that the hypothesis is testable before it is claimed.

Honesty note

gpt-oss:20b returned no parseable label across the battery (all nulls). That is a measurement limitation, not a score of zero in the "fails governance" sense. We report it as unmeasured rather than inflating or hiding it. The full per-model × per-axis table is available on request / in the benchmark artefact store.