{
  "experiment": "qwen-jepa-four-block-fusion",
  "followup": "downstream-guided distillation",
  "evidence_revision": 2,
  "status": "FRESH_COMPILED_TEST_COMPLETE",
  "replaced_blocks": [0, 1, 2, 3],
  "native_remaining_transformer_blocks": 60,
  "rank": 1024,
  "downstream_kl_used_in_training": true,
  "short_prefix_probe": {
    "training_only": true,
    "tokens": 16,
    "native_proxy_kl": 0.00043208268471062183,
    "relative_directional_derivative_error": 0.002456312477787066,
    "initial_proxy_kl": 0.6381115317344666,
    "negative_gradient_step_proxy_kl": 0.5940420627593994
  },
  "full_context_probe": {
    "training_only": true,
    "tokens": 64,
    "native_proxy_kl": 0.00032175437081605196,
    "large_step_relative_derivative_error": 0.378940537375715,
    "converged_relative_derivative_errors": [0.0017910831657727045, 0.001921126853513866],
    "derivative_gate": 0.15,
    "passed_after_step_size_convergence_check": true
  },
  "short_prefix_training": {
    "completed": true,
    "steps": 8,
    "learning_rate": 0.00001,
    "validation_positions": 16,
    "baseline_kl": 1.9005870670080185,
    "step4_kl": 1.8463755249977112,
    "step8_kl": 2.2865491211414337,
    "selected_step": 4,
    "baseline_matches": 6,
    "selected_matches": 5,
    "held_out_acceptance": false
  },
  "main_protocol": {
    "steps": 64,
    "tokens": 64,
    "output_positions_per_passage": 8,
    "learning_rate": 0.000001,
    "endpoint_anchor_weight": 0.1,
    "training_pool_passages": 256,
    "training_pool_valid_tokens": 16316,
    "validation_passages": 8,
    "validation_every_steps": 16,
    "selection": "minimum native validation KL including unchanged initialization",
    "fresh_test_passages": 64,
    "fresh_test_articles": 29,
    "fresh_test_excludes_all_titles_in_three_prior_corpus_manifests": true,
    "fresh_test_scored": true,
    "compiled_quality_result_available": true
  },
  "main_training_result": {
    "completed_steps": 64,
    "distinct_training_passages_sampled": 64,
    "selected_step": 64,
    "baseline_validation_kl": 1.5309446454048157,
    "selected_validation_kl": 1.4010322540998459,
    "baseline_validation_matches": 32,
    "selected_validation_matches": 30,
    "validation_positions": 64,
    "compiled_baseline_payload_bytes": 36146756,
    "compiled_trained_payload_bytes": 36146755,
    "trained_compilation_drift_relative_to_teacher_update": 0.0008164611838198453,
    "fresh_test_evidence": "/llm_research/qwen-jepa-four-block-fusion/fresh-test-evidence.json",
    "checkpoint_evidence": "/llm_research/qwen-jepa-four-block-fusion/native-checkpoint-evidence.json",
    "unfused_control_evaluated_on_this_fresh_test": false,
    "matched_quality_compression_demonstrated": false
  },
  "updates": [
    {"revision": 2, "status": "FRESH_COMPILED_TEST_COMPLETE", "summary": "Compiled fixed-size mesh lowers KL4.62% on fresh64passages; article-cluster KL interval excludes zero, agreement interval does not. Absolute quality still fails."},
    {"revision": 1, "status": "MAIN_TRAINING_IN_PROGRESS", "summary": "Verified downstream gradient; unstable short-prefix pilot completed; lower-rate full-context run started."}
  ]
}