BLEU-Probe Grounder Diagnostics

Plot-only diagnostics for the RepTok→GW→SONAR lightweight grounder loop, winning inspected config alpha=0.5 and beta max penalty=1. The fixed XPERF set is never used for probe training.
config: alpha=0.5, beta=1
direct XPERF BLEU: 10.854
max observed XPERF BLEU: 11.197
rounds: 9

Actual BLEU by Round

Actual BLEU by Round

Probe Learning Curves

Probe Learning Curves

Probe Held-Out RMSE by Round

Probe Held-Out RMSE by Round

Grounder Convergence

Grounder Convergence

Grounder Objective Components by Round

Grounder Objective Components by Round

Attention Weights by Step

Attention Weights by Step

Keep Gate by Step

Keep Gate by Step

Held-Out SONAR Loss by Step

Held-Out SONAR Loss by Step
machine-readable summary 
{
  "config": "alpha0.5_beta1",
  "direct_xperf_bleu": 10.853850617780271,
  "max_xperf_bleu": 11.196963305090026,
  "best_round_observed": 7,
  "rounds": 9,
  "source_cell": "reptok_sonar_bleu_probe_grounder_loop/alpha0.5_beta1",
  "analysis": "attention_best on fixed 200-image XPERF split",
  "plots": [
    "assets/bleu_by_round.jpg",
    "assets/probe_learning_curves.jpg",
    "assets/probe_rmse_by_round.jpg",
    "assets/grounder_convergence.jpg",
    "assets/grounder_components_by_round.jpg",
    "assets/attention_by_step.jpg",
    "assets/keep_gate_by_step.jpg",
    "assets/heldout_step_losses.jpg"
  ]
}