Dataset Viewer
Auto-converted to Parquet Duplicate
title
string
method
string
models
dict
headline
string
unmeasured
list
supersedes_note
string
axes
list
not_in_this_comparison
dict
SOV-SIGNAL honest arena — base vs sovereign, measured 2026-08-09 from Cowork (CPU/llama.cpp)
labels pulled from each dataset; one-word answer; regex extraction; temp 0; unparsed reported not scored wrong; Wilson 95%; McNemar exact on discordant pairs.
{ "sovereign": "csoai/sov-gate-ft2 (Qwen2 494M)", "base": "Qwen/Qwen2.5-1.5B-Instruct" }
base Qwen wins or ties all 9 governance axes; separated (beats sovereign) on governance (n=237, p~0). Sovereigns win refusal/safety framing (xstest, AgentHarm) not shown here.
[]
Replaces sov-signal-leaderboard-v1, whose sov33-v11 'champion' art5=0.944 / gov=0.489 match BASE Qwen's measured scores, not any runnable sovereign. Re-run sov33-v11 independently before crediting it; do not assume.
[ { "axis": "governance", "gspc_id": "gspc-gov", "sov_gate_ft2_494m": { "acc": 0.249, "wilson95": [ 0.198, 0.308 ], "n": 237 }, "base_qwen2_5_1p5b": { "acc": 0.54, "wilson95": [ 0.476, 0.602 ], "n": 237 }, "win...
{ "note": "gspc-care (n=201), gspc-swarm (n=41), gspc-det (n=34) are NOT unmeasured -- all clear usable_n>=30 and are measured on HF item banks. They were simply not run in THIS base-vs-sov-gate-ft2 comparison session. Correcting a stale claim inherited from an earlier draft that miscounted these as n=1/6/6.", "axe...

SOV-SIGNAL per-axis leaderboard

Per-axis accuracy (Wilson 95%) of the SOV33 SOV-SIGNAL champion on the 8 governance axes. Authored/sourced inside the CSOAI measurement estate. Facts quotable with attribution; substantial extraction licensed.

Downloads last month
22