LisaMegaWatts/sshfighter-17-head-router-v1

SSH Fighter 17-Head Frozen-Expert Router v1

Seventeen character-specific heads route among nine immutable Agent Gym policies. The package contains every three-seed checkpoint, averaged inference weights, the grouped training tensor cache, and fresh-seed exact-engine evaluation rows.

Fresh exact-engine evaluation

condition	points rate	wins	losses	draws	switches/match
fixed-jumper	0.7974	12445	3161	0	0.00
static	0.8096	12635	2971	0	0.00
linear	0.7999	12483	3122	1	124.21
chebyshev	0.8013	12505	3101	0	108.91

• Matches: 62,424
• Engine snapshot: f478c1320f86f428c882e391c11e97ca4cf7b47b
• Technical failures: 0
• Evaluation seeds: [1608202601, 1608202619, 1608202647]
• Passive samples: 1,492,845 with all nine proposal codes, literal HP-kick oscillator state, disagreement, agreement-CUSUM, and momentum labels. These signals have no control authority.

Training scope

• Train/dev/retrospective-holdout samples: {'dev': 117504, 'holdout': 117504, 'train': 352512}
• Model seeds: [42, 123, 456]
• Experts remain frozen; only routing parameters are learned.
• The exported ensembles carry Python golden logits which the TypeScript evaluator must reproduce before executing a match.

Limitations

Training labels are retrospective contextual-bandit observations, not full counterfactual rewards for all experts. Fresh evaluation uses known synthetic opponent families and direct exact-engine inputs—not SSH transport or human adaptation. This is the pre-balance-intervention f478c13 stratum and must not be pooled with later mechanics fingerprints.