Dataset Viewer
Auto-converted to Parquet Duplicate
gspc-gov
dict
gspc-care
dict
item_starved_not_trained
dict
{ "base_acc": 0.175, "tuned_acc": 0.775, "gain": 0.6, "n_eval": 40, "base_ci": [ 0.087, 0.32 ], "tuned_ci": [ 0.625, 0.877 ] }
{ "base_acc": 0.425, "tuned_acc": 0.875, "gain": 0.45, "n_eval": 40, "base_ci": [ 0.285, 0.578 ], "tuned_ci": [ 0.739, 0.945 ] }
{ "gspc-asi": 33, "gspc-mach": 33, "gspc-xr": 32, "gspc-prv": 32, "note": "<60 items — too few for honest train+40-eval split; need items authored first, NOT a fake run" }

GSPC Arena Results — Cross-Model Per-Axis Routing

Real measured results from the SOV3 arena: 8 models scored across 12 GSPC governance axes via the sovereign model relay, n=20/cell with Wilson confidence intervals.

Contents

  • clan_routing_table_FINAL.json — per-axis winner + CI (8/12 sovereign-family-won, falcon3 4/12)
  • arena_binom_12axes.json — 12 axes, all above chance (binomial-verified, correct per-axis baselines)
  • sov34_board.json — sovereign model full board
  • gsm8k_benchmark_agnostic_proof.json — proof the harness scores any benchmark (GSM8K, not just GSPC)
  • sovereign_family_map.json — the ~30 trained sovereign variants

Honest caveats

  • n=20/cell: routing winners have overlapping CIs; ~half the per-axis winners are statistical ties.
  • Scores are single-run temperature-0; indicative of specialization, not final rankings.
Downloads last month
43